Blog on Carles Abarca

Stop Crying About AI and Jobs

Wed, 15 Apr 2026 00:00:00 +0000

Stop Crying About AI and Jobs
#

The public conversation about artificial intelligence has filled up with a strange mix of fascination, fear, and theatrical panic. Every week brings a new headline claiming that AI is about to destroy human work, empty offices, make entire professions obsolete, and push millions of people into irrelevance. The problem is that when you actually look at the data with even a minimum of rigor, the real story looks very different.

No, AI is not wiping out employment on a massive scale. What it is doing, and this part is genuinely profound, is transforming tasks, redesigning workflows, and changing the skills the market will demand over the coming years.

And that difference matters enormously.

The Wrong Narrative
#

There is a reason the catastrophic narrative spreads so easily: it works extremely well as content. “AI is going to take your job” generates immediate attention. It activates fear, outrage, and anxiety. It is easy to consume and even easier to share.

But a good narrative does not always describe reality well.

The most serious studies emerging in 2025 and 2026 point to a fairly consistent pattern. AI is not operating primarily as a force of brute job elimination, but as a technology that reorganizes the content of work. It automates parts, accelerates others, raises the expected standard in many roles, and creates new demand in layers that either did not exist before or were still marginal.

The right question, therefore, is not whether AI will make all employment disappear. The right question is different: which parts of human work become automatable, which ones gain value, and which new capabilities become decisive?

What the Data Actually Says
#

The data currently on the table does not support the caricature of total labor collapse.

An MIT study published in April 2026 points precisely in this direction: AI transforms tasks far more than it destroys entire jobs. This fits with an idea labor economists have understood for a long time: most jobs are not made up of a single task, but of combinations of activities. When a technology automates one part of the work, the role does not automatically disappear. In many cases, it gets redefined.

Other signals point in the same direction and should not be ignored.

LinkedIn has highlighted the creation of 1.3 million new jobs linked to AI. We are not talking only about data scientists or prompt engineers. We are talking about specialized recruiting, technology integration, operations, governance, cybersecurity, sales enablement, training, product, automation, and entirely new service layers.

On top of that, infrastructure expansion is creating a major pull effect. More than 600,000 new roles tied to data center infrastructure are expected, from construction and operations to energy, cooling, maintenance, networking, and specialized support. When the digital economy scales, not only do the models grow. The entire system that makes them possible grows with them.

BCG has also been clear: AI will reshape more jobs than it eliminates. In other words, the dominant effect will be redesign, not pure extinction.

That does not mean there will be no disruption. There will be, and a lot of it. What it means is that the disruption looks less like the mass disappearance of work and more like an accelerated reassignment of value.

The Blind Spot: Your Job Title May Survive, Your Skills May Not
#

This is the nuance many people still do not want to face.

The biggest impact of AI will not necessarily fall on the title of your job. It will fall on the actual content of what you do every day and on the skills you need in order to remain valuable.

It is estimated that 70% of the skills associated with many roles will change before 2030. That number should be more unsettling, and more motivating, than any apocalyptic prediction about job destruction.

Because it implies something simple: your role may still exist, but you may no longer be able to do it in the same way.

An analyst will still be an analyst, but with tools that drastically compress analysis time. An executive will still be an executive, but will no longer be able to make decisions without sound judgment about automation, models, data, risk, and augmented productivity. A teacher will still be a teacher, but will need to redesign the learning experience in an environment where knowledge is abundant and judgment becomes the scarce asset. A doctor will still be a doctor, but will work in a context where AI can assist with documentation, diagnostic support, and clinical prioritization.

The continuity of a job title does not guarantee the continuity of professional value.

And that is where the real conversation begins.

What Teams Are Actually Experiencing
#

Those of us who have spent years working in technology keep seeing the same thing again and again: teams that integrate AI well do not necessarily reduce headcount as an automatic reflex. What they do, above all, is increase their execution capacity.

They deliver more. They iterate faster. They test more hypotheses. They reduce friction. They reallocate time from the mechanical to the strategic.

That completely changes the competitive standard.

AI does not instantly replace the competent professional. But it does expose, with great clarity, the people who decide to stay still while the environment moves forward. Not because the machine is magical, but because a team that learns to work with AI can, in certain contexts, produce twice as much in half the time.

And when that happens, the problem stops being “AI versus humans.” The problem becomes “humans who evolve versus humans who refuse to evolve.”

The Collapse of Barriers to Creation
#

There is another angle that I think matters enormously, and that too often gets buried under all the fear: AI is lowering barriers to create, build, launch, and scale.

Today, one person with judgment, initiative, and the right tools can do things that only a few years ago required much larger teams. Prototyping, analysis, writing, design, automation, research, material preparation, product launches, idea validation, and process operations are now within reach of far more people.

That does not reduce the importance of talent. It multiplies it.

And this is not theory. There are already cases that illustrate the shift extremely well. Peter Steinberger, for example, has shown just how far an entrepreneur with technical judgment and an intense layer of AI agents can operate at a speed that once seemed reserved for full teams. Another widely cited example in the recent conversation around AI-powered solopreneurs is Maor Shlomo with Base44, a project run with an extremely lean structure and supported by AI-assisted development, which scaled rapidly and became a powerful signal of what changes when the cost of building software collapses.

These are examples that help clarify the direction of change: AI is radically expanding the opportunities available to talented individuals with focus and execution capacity.

As LinkedIn’s CEO has put it, AI is lowering barriers to create and build. And that is an enormously powerful signal. Because where some people only see substitution, others are already seeing expansion of capability.

The Real Risk Is Not AI
#

The real risk is not that AI exists.

The real risk is responding to this transition with denial, cynicism, or paralysis.

It is confusing comfort with safety. It is assuming the future will respect inertia. It is believing it will be enough to keep doing what you have always done while the rest of the market reconfigures its capabilities.

Resistance to change always disguises itself as prudence. But in deep technological cycles, it often is not prudence. It is delay.

And delay, when the environment accelerates, becomes expensive.

So What Should We Actually Do?
#

There is no need to dramatize. There is a need to act.

At least across five fronts:

1. Learn to work with AI, not just talk about AI
#

There are too many people expressing opinions about artificial intelligence without having integrated the technology into their real workflows. Useful literacy is not about being able to define an LLM. It is about knowing when to use it, for what, with which limits, and with what judgment.

2. Strengthen judgment
#

When intelligence becomes abundant, judgment becomes more valuable. The ability to interpret, decide, prioritize, contextualize, and assume responsibility will matter more and more.

3. Redesign processes
#

AI does not produce deep results when it is bolted on as a superficial accessory. Real impact comes when processes, roles, metrics, and forms of collaboration are redesigned.

4. Commit to continuous learning
#

This is not a technology you learn once and then finish with. Competitive advantage will come from sustained adaptability.

5. Replace fear with discipline
#

Anxiety does not build capability. Disciplined experimentation does.

An Uncomfortable but Useful Conclusion
#

AI is not automatically going to take your job.

What can leave you behind is the decision not to evolve while everything around you changes.

That is why I believe it is time to abandon technological victimhood and start speaking more honestly about what is really happening. AI is not destroying the value of human work. It is redefining which human work creates value.

And in that transition, the future will not reward the loudest complainers. It will reward the people who learn faster, combine technology with judgment more effectively, and have the courage to redesign themselves before they are forced to do so.

The question is not whether AI is going to change work.

The question is whether you are going to change with it.

Claude Mythos: the model Anthropic chose not to release

Thu, 09 Apr 2026 00:00:00 +0000

“Claude Mythos Preview is a general-purpose, unreleased frontier model.”
— Anthropic, Project Glasswing

Anthropic has just made a decision that, until very recently, would have seemed almost unthinkable in the race for frontier models: publicly present a new-generation model while simultaneously deciding not to make it broadly available to the market.

This is not a product delay. Nor is it a conventional beta program. What Anthropic has done with Claude Mythos Preview is something else: publish part of the technical documentation, describe extraordinary capabilities—especially in offensive cybersecurity—and restrict access to a very limited circle of defensive actors under a specific initiative: Project Glasswing.

The important question is not only what Mythos is. The important question is what it means that Anthropic has decided not to launch it like a normal model.

The extraordinary part is not the model. It is the decision.
#

In the AI industry, a fairly clear logic had taken hold: if a lab trains a better model, sooner or later it turns it into a product. It may do so gradually, via APIs, waitlists, enterprise agreements, or usage restrictions. But the overall direction was unmistakable: more capability eventually meant more availability.

With Mythos, Anthropic introduces a break.

On the one hand, it presents the model as a new frontier of capability. On the other, it implicitly admits that this capability crosses a threshold that makes broad deployment irresponsible.

“We formed Project Glasswing because of capabilities we’ve observed in a new frontier model trained by Anthropic that we believe could reshape cybersecurity.”
— Anthropic, Project Glasswing

That is not routine marketing language. It is a governance signal. Anthropic is saying that, in its judgment, the model is not just better: it is dangerously better in one specific dimension.

What Anthropic claims about Claude Mythos Preview
#

The documentation Anthropic has published paints a picture that is difficult to ignore.

In its Frontier Red Team technical post, the company argues that Mythos Preview:

identifies and exploits zero-days in real software,
does so across every major operating system and every major browser,
produces complex exploits, including multi-vulnerability chains,
and represents a radical leap beyond previous Claude generations.

“During our testing, we found that Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so.”
— Anthropic, Claude Mythos Preview

If this is correct, we are not looking at an incremental improvement. We are looking at a regime change.

Anthropic goes further. It says internal engineers with no formal security training have asked the model to find a remote vulnerability overnight and woken up the next morning to a complete working exploit.

“Engineers at Anthropic with no formal security training have asked Mythos Preview to find remote code execution vulnerabilities overnight, and woken up the following morning to a complete, working exploit.”
— Anthropic, Claude Mythos Preview

That detail matters. It suggests not only that the model amplifies expert capability. It also suggests that it dramatically lowers the barrier to entry for advanced offensive capability.

The leap beyond Opus 4.6
#

One of the most striking elements of the technical documentation is the comparison with earlier generations.

Anthropic notes that, only a month earlier, its read on Opus 4.6 was that the model was much better at finding and fixing vulnerabilities than at exploiting them. In other words, it was still strong in defensive cybersecurity, but not especially effective at autonomous offensive work.

With Mythos, that changes.

“Opus 4.6 generally had a near-0% success rate at autonomous exploit development. But Mythos Preview is in a different league.”
— Anthropic, Claude Mythos Preview

The company cites a benchmark involving Firefox vulnerabilities where Opus 4.6 only managed to convert findings into working exploits a handful of times, while Mythos Preview did so 181 times, with register control in 29 additional cases.

If those numbers hold, we are not talking about “a stronger Claude.” We are talking about a different order of capability.

It was not trained “to hack”
#

This point is critical.

Anthropic says it did not explicitly train Mythos Preview to develop these offensive capabilities. According to the company, what we are seeing is an emergent consequence of broader improvements in:

reasoning,
autonomy,
code work,
and multi-step planning.

“We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.”
— Anthropic, Claude Mythos Preview

That sentence deserves to be read carefully, because it points to something bigger than Mythos. It suggests that as generalist models improve at useful code work and agentic behavior, offensive capability stops being a separate specialty. It appears as a natural side effect of general progress.

That makes governance much harder. It is no longer enough to avoid training “a model for cyberattack.” The real issue is that a sufficiently capable general model can become a first-rate offensive tool even if that was never the explicit objective of training.

So why not release it?
#

Anthropic frames the answer in terms of a dangerous transition window.

Its thesis is that, in the long run, tools like this may benefit defenders more than attackers. But in the short run there is an obvious risk: offensive capability may diffuse faster than defensive capability can absorb it.

“In the short term, this could be attackers, if frontier labs aren’t careful about how they release these models.”
— Anthropic, Claude Mythos Preview

That is why there is no broad release. Instead, Anthropic created Project Glasswing, an initiative involving partners such as AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, along with dozens of additional organizations.

“By releasing this model initially to a limited group of critical industry partners and open source developers with Project Glasswing, we aim to enable defenders to begin securing the most important systems before models with similar capabilities become broadly available.”
— Anthropic, Claude Mythos Preview

In other words: Anthropic is trying to turn a capability advantage into a temporary defensive advantage before the rest of the ecosystem catches up.

What is really changing: publishing no longer means deploying
#

The most interesting thing about Mythos is not only the security argument. It is the precedent it sets.

For years, many of us assumed that the most advanced model in a lab would also, sooner or later, be the one that ended up in the hands of customers, developers, or end users. With Mythos, that equivalence breaks.

From now on, the most advanced model may:

not be the main product,
not be broadly offered via API,
not reach the general market,
and exist for some time in a kind of strategic quarantine.

That changes a great deal.

It changes how we think about competition between labs. It changes how we should read public announcements. And it changes the regulatory and geopolitical frame as well: if the most powerful models are no longer necessarily public, then the true frontier of capability may increasingly sit behind restricted-access programs, private agreements, and asymmetric deployments.

But a critical reading is still necessary
#

That said, it would be a mistake to swallow the narrative whole.

Anthropic is making extraordinary claims:

thousands of high-severity vulnerabilities,
zero-days in critical software,
coverage across every major OS and browser,
sophisticated exploits developed autonomously,
and a security rationale strong enough to justify withholding the model.

The problem is that the public evidence is necessarily limited.

Anthropic itself says that more than 99% of the vulnerabilities it has found are still unpatched and therefore cannot be disclosed. In addition, the risk document is presented in redacted form.

“Over 99% of the vulnerabilities we’ve found have not yet been patched, so it would be irresponsible for us to disclose details about them.”
— Anthropic, Claude Mythos Preview

That is reasonable from the standpoint of responsible disclosure. But it also means that much of this story depends on trusting the lab’s own interpretation and framing.

So yes: Anthropic’s decision may be sensible, even admirable, while still being wrapped in a corporate narrative that deserves methodological skepticism.

My read: Mythos may mark a before and after
#

My impression is that this episode may ultimately be remembered less for the model’s name than for the strategic signal it sends.

Anthropic is not only saying “we trained something very powerful.” It is saying something more uncomfortable:

we have crossed a capability frontier where responsible behavior no longer automatically means publication.

If that thesis holds, Mythos will matter for three reasons.

1. Because it normalizes partial retention of frontier models
#

Not as an anecdotal exception, but as a legitimate governance tool.

2. Because it shifts the debate from “what can the model do?” to “who should be allowed to use it, and when?”
#

That is a fundamental change.

3. Because it suggests that the real frontier of capability may already sit several steps ahead of what we see in product
#

And that has major implications for strategy, technology policy, and security.

The uncomfortable conclusion
#

For years, the dominant AI narrative assumed that technical progress would eventually democratize access to ever more powerful capabilities.

Claude Mythos introduces a different possibility: that some capabilities are so sensitive that technical progress will not lead to openness, but to containment.

Not because the model failed. Precisely because it worked too well.

“Claude Mythos Preview reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.”
— Anthropic, Project Glasswing

If Anthropic is right, this is not simply another model launch. It is the moment when a frontier lab explicitly decided that its most advanced system should not behave like a normal product.

And in this industry, that is a much bigger story than any benchmark.

Main sources
#

Anthropic — Project Glasswing
https://www.anthropic.com/glasswing
Anthropic Frontier Red Team — Claude Mythos Preview
https://red.anthropic.com/2026/mythos-preview/
Anthropic — Alignment Risk Update: Claude Mythos Preview (Redacted)
https://www.anthropic.com/claude-mythos-preview-risk-report

The Era of Cheap AI Is Ending

Tue, 07 Apr 2026 00:00:00 +0000

AI felt cheap. It wasn’t. It was subsidized.

A couple of years ago, if someone had told you that the world’s most advanced AI would be available for $20 a month, you would have laughed. And yet, for a while, that’s exactly what it felt like.

Flat-rate subscriptions. Increasingly capable models. Usage that, in practice, felt nearly unlimited. The prevailing perception was clear: advanced AI was becoming an abundant, accessible resource.

But that perception is starting to crack. Not because of a technical failure — because of something more fundamental: economics.

The Signs Are Already Here
#

In recent weeks, several converging signals have pointed in the same direction:

Anthropic, the creator of Claude, has started actively restricting certain usage patterns through its subscriptions. In particular, it has blocked automation tools like OpenClaw from channeling requests through subscription accounts. The reason is straightforward: the inference cost of those intensive usage patterns doesn’t square with the subscription price.

This isn’t an isolated case. OpenAI has been adjusting the real usage limits of its subscription plans for months, progressively reducing the number of interactions with their most powerful models before downgrading users to a lesser model. What were once generous, fuzzy limits are becoming explicit, harder caps.

And there’s something more unsettling: even with these adjustments, most major AI operators are still not profitable. Anthropic, OpenAI, and virtually every frontier lab are burning capital at a pace that would make any traditional CFO faint.

The obvious question is: if users are already feeling restrictions, how is it possible that providers are still losing money?

The answer reveals something important about the real economic structure of this industry.

AI Was Never Cheap. It Was Subsidized.
#

What we’ve experienced over the past two years was not the real cost of AI. It was an adoption strategy.

The major labs needed to do three things simultaneously:

Demonstrate capability to justify valuations in the tens (or hundreds) of billions.
Drive massive adoption to create network effects, lock-in, and usage data.
Capture developers and enterprises before the competition did.

To achieve this, they offered access to frontier models at prices that didn’t reflect the real cost of operating them. The $20/month subscriptions were, in practice, a subsidy funded by venture capital.

And it worked. Adoption soared. Millions of people began using Claude, ChatGPT, and other models daily. Companies of every size started integrating AI into their workflows.

But now we’re entering the next phase. And in this phase, the numbers have to start adding up.

Frontier Models Don’t Follow Traditional Software Economics
#

There’s a widespread misunderstanding that needs to be addressed head-on: many people assume that AI follows the same economic logic as conventional software. That is: you develop it once, distribute it at near-zero marginal cost, and margins improve with scale.

But generative AI — especially frontier models — doesn’t work like that. Not at all.

Every interaction with a large model involves:

Real-time computation on very expensive hardware (state-of-the-art GPUs).
Significant energy consumption per request.
Infrastructure costs that don’t disappear with scale; in many cases, they grow with it.
Latency and availability requirements that demand reserved capacity, not just peak capacity.

And the problem is amplified by the most recent usage trends. AI agents — which execute multiple chained calls to complete complex tasks — multiply inference costs dramatically. A single agent session solving a programming problem can involve dozens of model calls, each with long contexts and active tools.

More capability doesn’t automatically mean lower unit cost. In many frontier AI cases, it means exactly the opposite.

The Flat Rate Was a Commercial Strategy, Not a Sustainable Reality
#

Let’s think about what a flat-rate subscription to a frontier model actually meant in practice:

A user paid $20 a month.
They had access to a model whose inference cost, under intensive use, could easily exceed $100 or $200 per month per user.
The provider absorbed the difference, betting that the average user wouldn’t use the model that intensively.

That model works reasonably well when most users are casual: they ask a few questions a day, use the model for simple tasks, and don’t stress the infrastructure. It’s the same principle that makes gyms work: they sell more memberships than the gym can simultaneously accommodate, betting that most people won’t show up every day.

But when tools emerge that channel intensive, programmatic use through those subscriptions, the model breaks. It’s as if someone found a way to fill the entire gym 24 hours a day. The price no longer covers the cost.

That’s why providers are reacting:

More explicit usage limits per model and per period.
Automatic degradation to less expensive models when certain thresholds are reached.
Contractual restrictions against unanticipated usage patterns.
Clearer separation between consumer, API, and enterprise plans.

It’s not a whim. It’s economic survival.

The Great Paradox: More Restrictions, and Still Not Enough
#

Here’s the data point that should give pause to any business leader betting heavily on AI:

Even after introducing all these restrictions, most frontier AI operators are still losing money.

Anthropic has raised over $10 billion in funding. OpenAI is on a similar trajectory. And neither has yet demonstrated a sustainable economic model at scale without continuous external capital injection.

This doesn’t mean the business is unviable. There’s probably an economic equilibrium somewhere, with the right combination of pricing, inference efficiency, enterprise volume, and hardware optimization. But that equilibrium clearly hasn’t been reached yet.

And meanwhile, the industry keeps advancing toward bigger, more capable models with wider context windows, more integrated tools, and more agentic capabilities. All of those improvements are fantastic for users. But each one increases inference cost, not reduces it.

What This Means for Enterprises
#

If you’re a digital transformation leader or a CTO designing your organization’s AI strategy, the message is clear:

Don’t build your AI architecture assuming that today’s cost is tomorrow’s cost. And certainly not assuming it will go down.

In practice, this means several things:

1. Design for Multiple Model Tiers
#

Not everything needs a frontier model. Many tasks — classification, data extraction, routine summaries, basic assistance — can be handled by smaller, significantly cheaper models. Reserve the most powerful models for tasks that truly require them.

2. Implement Intelligent Routing
#

Mature AI architectures don’t send everything to the same model. They route each request to the most efficient model that can resolve it at the required quality level. This can reduce inference costs by 60-80% without degrading the end-user experience.

3. Measure Cost Per Use Case
#

If you don’t know how much each AI interaction costs, broken down by task type, model used, and outcome achieved, you’re flying blind. AI cost observability should be as standard as it is in cloud infrastructure.

4. Think of AI as Infrastructure, Not Just Software
#

Generative AI has a variable cost component that resembles electricity or cloud compute more than a software license. Plan accordingly: with capacity reserves, variable budgets, and consumption governance.

5. Don’t Depend on a Single Provider
#

Concentration on a single model provider exposes you directly to their pricing decisions, policy changes, and usage restrictions. A multi-model architecture gives you the flexibility to adapt when — not if — conditions change.

The End of an Era, Not the End of AI
#

I want to be clear about something: nothing I’ve described here is negative for the future of AI. It is simply the end of a phase.

The phase that’s ending is one of illusory abundance: frontier models at promotional prices, apparently unlimited usage, and a widespread feeling that advanced AI was becoming a commodity.

The phase that’s beginning is more honest. It’s the phase where:

Prices more faithfully reflect real costs.
Providers find sustainable business models.
Enterprises learn to use AI with economic discipline.
And the market matures, as cloud, SaaS, and every other technology infrastructure that started with promotional pricing has matured before it.

In a sense, it’s good news. Because an indefinitely subsidized market is a fragile market. And a market that finds its economic equilibrium is a market that can last.

Conclusion
#

AI isn’t suddenly becoming more expensive. We’re simply stopping the pretense that it was cheap.

And the question that should be on every boardroom table is no longer just “What can AI do for us?” It’s something more uncomfortable and more necessary:

How much is it really going to cost us, and are we designing for sustainability?

Whoever has a good answer to that question will have a real competitive advantage. Whoever doesn’t will discover that the most expensive AI isn’t the one with the best model — it’s the one that was deployed without thinking about the economics.

Carles Abarca is VP of Digital Transformation at Tecnológico de Monterrey. He writes about AI, digital strategy, and the future of technology in organizations.

Your Code Is One Agent Session From Being Cloned — And There's Nothing You Can Do About It

Fri, 03 Apr 2026 00:00:00 +0000

On April 1, 2026 — and no, it wasn’t a joke — someone discovered that Claude Code’s npm package, Anthropic’s command-line tool, included a misconfigured .map file. That file contained the complete source code: 512,000 lines of TypeScript spread across 1,900 files.

Within hours, the leaked repository had 25,000 stars on GitHub. Developers were tearing it apart piece by piece. But the truly unsettling part wasn’t the leak itself.

It was what happened next.

From TypeScript to Python in an afternoon
#

A researcher at the University of Washington took the leaked code, fed it through OpenAI Codex, and within hours had a functional reimplementation in Python. Not a copy — a complete rewrite in a different language. The resulting repository, claw-code, hit 44,500 stars. A Rust rewrite is already underway.

And here’s the legal problem that should keep every CTO up at night: a reimplementation in another language, generated by an AI agent, is likely immune to a DMCA takedown. It’s not a copy. It’s a derived work created by a machine. Current legal frameworks simply aren’t designed for this.

Gergely Orosz, one of the most respected voices in software engineering, put it this way: we’re facing a new reality where any closed codebase is one agent session away from being functionally cloned.

Read that again. Any codebase. One agent session. Cloned.

What they found inside
#

The findings were fascinating and, in some cases, hilarious. Wes Bos discovered 187 hardcoded verbs for the loading spinner (including “hullaballooing” and “razzmatazzing”). They found an internal analytics system that flags your prompt as negative every time you swear at the agent. And the random 4-character IDs are filtered to exclude 25 curse words.

But beyond the anecdotes, the technical discoveries revealed the real architecture of one of the world’s most advanced coding agents:

35 modules with clearly separated responsibilities.
A lightweight memory system based on ~150-character pointers, not bulk storage.
System prompts live on the client, not the server.
Code comments are written for LLMs to read, not humans.

That last point deserves a pause. Anthropic is no longer writing code for programmers to understand. They’re writing code for other AI models to understand. If that doesn’t tell you where the industry is heading, nothing will.

The protective barrier that no longer exists
#

For decades, proprietary code was one of the most reliable protective barriers in technology. Your competitor could copy your idea, your design, your go-to-market strategy — but replicating a million lines of optimized code took years and hundreds of engineers.

That’s over.

I’m not exaggerating. Think about what actually happened this week:

A configuration error exposed complete source code.
An AI agent rewrote it in another language in hours.
The open source community improved and extended it in days.
The result is legally difficult to challenge.

Each of these steps was unthinkable two years ago. Together, they represent a fundamental shift in what it means to “protect” your software intellectual property.

And you don’t need a leak for this to happen. AI agents can already:

Infer architectures from the observable behavior of an API.
Replicate functionality from public documentation and examples.
Generate alternative implementations that achieve the same results with different code.

Code was never really a barrier. It was the illusion of a barrier.

The SaaSpocalypse connection
#

If this story sounds familiar, it’s because we already saw act one. In February of this year, I wrote about the SaaSpocalypse: $300 billion evaporated in 48 hours when the market understood that the business logic of a $200/user/month SaaS fits in a text file that an agent can read and execute.

The Claude Code leak makes that thesis worse in a way that few have articulated yet.

Think about it: every SaaS provider exposes documented public APIs. Endpoints, data schemas, workflows, business rules — all accessible to any agent that can read documentation. The Claude Code leak proved that an agent can take 512,000 lines of code and reimplement them in hours. Now combine that with the fact that most SaaS products expose their business logic through their own APIs.

You don’t need to steal Salesforce’s source code. You just need an agent that reads their public documentation, observes endpoint behavior, and infers the underlying architecture. The API is the blueprint.

Satya Nadella said it in December 2024: “SaaS applications are CRUD databases with business logic on top. Agents will absorb that logic.” In February, the market understood this in the abstract. With Claude Code, we have concrete proof: the tools to absorb that logic already exist. And they work.

The SaaSpocalypse wasn’t the ending. It was the trailer.

So what actually protects your business?
#

If code is no longer your competitive advantage — and public APIs reveal your business logic — what’s left? After 20 years in technology transformation, from banking in Spain to education in Mexico, I’ve seen this same question emerge every time a new technological wave destroys the previous barriers:

1. Data, not code. Your model trained on proprietary data, your curated datasets, your domain knowledge encoded in features that an agent can’t infer from the outside. A Claude Code clone can replicate the tool, but it can’t replicate the millions of conversations that trained Claude.

2. Speed of execution. If your competitor can clone your code in hours, the advantage lies in being the first to solve the next problem. Not in protecting the previous solution.

3. Trust and brand. In a world where anyone can replicate the technology, differentiation comes down to who trusts you. Anthropic’s customers aren’t going to migrate to claw-code to save on their subscription. They pay for support, for SLAs, for the guarantee that someone responds when something breaks.

4. The ecosystem. Integrations, partnerships, network effects. Slack didn’t win because its code was uncopyable. It won because everyone was already there.

5. A culture of continuous innovation. If you assume that everything you build today will be replicable tomorrow, your only sustainable advantage is the ability to build the next thing faster than anyone else.

The implications for the enterprise
#

For any technology leader reading this, the message is clear: review your intellectual property strategy today. Not tomorrow. Today.

Some questions that should be on the agenda for your next board meeting:

How much of our competitive advantage depends on code that an agent could replicate?
What does our own API documentation reveal about our business logic?
Do we have proprietary data that is genuinely hard to reproduce?
Does our security strategy account for the fact that a misconfigured .map file can expose our entire codebase?
Are we prepared for a world where DMCA doesn’t protect against AI-generated reimplementations?

And perhaps the most uncomfortable one of all: Are we still investing in building walls, when we should be investing in running faster?

The twist no one expected
#

There’s a delicious irony in this whole story. Claude Code — Anthropic’s tool designed for AI to write code — was dismantled and rewritten by the AI of its direct competitor. OpenAI Codex cloned Anthropic’s flagship product in an afternoon.

It’s as if Ford had invented the assembly line and Toyota had copied it the same day using Ford’s own robots.

Welcome to the era where the tools you build to automate other people’s work can be used to automate your own work. Where your code is not your protective barrier. Where your advantage is not what you already built, but what you’re going to build tomorrow.

The Claude Code leak wasn’t a security incident.

It was a warning.

Carles Abarca is VP of Digital Transformation at Tecnológico de Monterrey and former CTO of Banco Sabadell. He writes about AI, digital transformation, and the future of software at carlesabarca.com.

Claude Mythos: The Model That Made Cybersecurity Stocks Crash — And What It Tells Us About Where AI Is Heading

Sun, 29 Mar 2026 00:00:00 +0000

Three days ago, a misconfigured CMS at Anthropic left roughly 3,000 internal assets publicly accessible. Among them: a draft blog post announcing their next-generation AI model. The name varies between two leaked drafts — “Mythos” and “Capybara” — but what matters isn’t the name. What matters is what it can do.

And what it can do should make anyone in technology leadership stop and think very carefully.

What Leaked
#

On March 26, security researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) discovered the exposed documents. Anthropic acknowledged the leak as “human error” and confirmed the model is real.

Here’s what we know:

Claude Mythos is not Claude Opus 4.7. It’s not an incremental update. It’s a new tier above Opus — Anthropic’s own words: “a new name for a new tier of model: larger and more intelligent than our Opus models, which were, until now, our most powerful.” Reports suggest roughly 10 trillion parameters, a 5-10x jump from previous frontier models.

Training is complete. Select customers are already testing it.

Why Cybersecurity Stocks Crashed
#

The morning after the leak, the market’s reaction was swift and brutal:

iShares Cybersecurity ETF: -4.5%
CrowdStrike, Palo Alto Networks, Zscaler, SentinelOne: -6% each
Tenable: -9%
Bitcoin dropped to $66,000

Why? Because the leaked draft describes Mythos as “currently far ahead of any other AI model in cyber capabilities.” It can discover and exploit software vulnerabilities at speeds that — Anthropic’s own assessment — “far outpace human defenders.”

Read that again. The company that built it is telling you that human cybersecurity teams can’t keep up with it.

This isn’t hypothetical. Anthropic already caught a Chinese state-sponsored group using Claude Code to infiltrate approximately 30 organizations — tech companies, financial institutions, government agencies — before detection. Mythos reportedly makes that look like child’s play.

Stifel analyst Adam Borg put it plainly: “Mythos is an order of magnitude more powerful, and compute-intensive, than any other frontier model on the market.”

The Rollout Strategy Tells You Everything
#

Anthropic’s deployment approach is perhaps the most revealing signal:

First access: Not developers. Not enterprises. Cybersecurity organizations — “giving them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits.”
No public launch date. They’re explicitly delaying broad release.
Cost problem acknowledged. Anthropic says it’s “very expensive to serve” and they need to make it “much more efficient before any general release.”

When a company builds the most powerful AI model in the world and its first instinct is to hand it to defenders before attackers can get it — that’s not a product launch. That’s a controlled disclosure.

What Mythos Means Beyond Cybersecurity
#

Let me be direct about what I think this represents.

Mythos posts “dramatically higher scores” than Opus 4.6 on coding and academic reasoning benchmarks. Opus 4.6 already led SWE-bench Verified at 80.8% and Terminal-Bench 2.0 at 65.4%. Whatever “dramatically higher” means, we’re talking about a model that can code better than most professional developers and reason through complex problems at a level that was science fiction five years ago.

But the cybersecurity capability is the real wake-up call, because vulnerability discovery requires something qualitatively different from text generation or code completion. It requires:

Deep multi-step reasoning — chaining logical inferences across complex systems
Adversarial creativity — finding attack vectors that weren’t designed or anticipated
Autonomous execution — not just identifying a vulnerability but actively exploiting it

When a model can do all three at superhuman speed in a domain as complex as cybersecurity, the implications extend to every field that involves complex reasoning under uncertainty. Law. Medicine. Scientific research. Strategic planning. Finance.

The AGI Question (Which Is the Wrong Question)
#

Is Mythos AGI? No. It doesn’t learn new tasks from minimal examples the way humans can. It has no persistent memory, no self-improvement loop, no autonomous goal-setting.

But here’s what I think matters more: we may be past the point where the AGI label matters practically.

A model that can autonomously find and exploit zero-day vulnerabilities — something that previously required teams of elite human researchers — changes the game regardless of whether we call it “general” intelligence. Narrow superintelligence in high-stakes domains is more immediately consequential than theoretical AGI.

The fact that Anthropic itself is alarmed enough to delay general release and prioritize defensive deployment tells you where we are on the capability curve.

The Competitive Context Makes It Worse
#

Mythos doesn’t exist in isolation:

OpenAI has finished pretraining a new model codenamed “Spud” — expected within weeks.
Google DeepMind just launched Gemini 3.1 for real-time multimodal processing.
Both Anthropic and OpenAI are timing major releases ahead of planned IPOs later in 2026.

This is an arms race with IPO pressure. The incentives to push capability boundaries are enormous and increasing. The incentives for caution are… well, we just saw how Anthropic’s caution played out. A CMS misconfiguration, and the whole world knows.

What This Means for Institutions
#

For universities, for governments, for any organization making decisions about AI strategy:

The planning horizon just compressed. If you were thinking about AI governance frameworks as a 2027-2028 initiative, think again. Models with superhuman capabilities in specific domains are here now, not in a comfortable future.

Cybersecurity is no longer optional. It’s existential. Every institution needs to assume that AI-powered attacks will become the norm, not the exception. The defenders need AI too — and they need it first.

The talent equation is shifting. When a model can outperform human cybersecurity experts, the value isn’t in the technical execution — it’s in the judgment about when and how to deploy these capabilities. We need people who understand both the technology and its implications.

I keep coming back to the same conclusion I wrote in my previous post on AEO: digital transformation in 2026 means preparing institutions for a world where AI systems are colleagues, not tools. Mythos just made that statement feel uncomfortably literal.

Jensen Huang said AGI has arrived. He was wrong about the definition but right about the urgency. Whether we call it AGI or narrow superintelligence or just “really powerful AI” — the systems are here, they’re real, and the time to prepare was yesterday.

Carles Abarca is Vice President of Digital Transformation at Tecnológico de Monterrey.

From SEO to AEO: How AI Agents Are Redefining Digital Transformation in Higher Education

Sat, 28 Mar 2026 00:00:00 +0000

Two weeks ago, Jensen Huang took the stage at GTC 2026 and declared that Artificial General Intelligence has arrived. “We have reached the level of artificial general intelligence,” he said, in his trademark leather jacket and the tone of someone who knows they’re dropping a bomb. His definition, however, is telling: for him, AGI is the ability to create billion-dollar businesses. A purely capitalist definition.

But in universities, the question isn’t whether AGI can generate billions. The question is: what changes when our students, faculty, and institutions interact daily with systems that reason, plan, and execute autonomously?

The answer is: everything.

The Age of Agents
#

Forget chatbots. In 2026, the dominant trend in AI isn’t text generation — it’s Agentic AI: autonomous systems that receive a goal, create a plan, use tools (your email, your CRM, your spreadsheets) and execute tasks without constant human intervention.

The data backs this up. According to Deloitte’s State of AI in the Enterprise report, 3 out of 4 companies plan to deploy AI agents within the next two years. Harvard Business Review dedicates its cover this month to: “To Scale AI Agents Successfully, Think of Them Like Team Members.” The message is clear: agents aren’t software you install — they’re changes in how work gets done.

But here’s the number that concerns me: only 1 in 5 organizations has a mature governance model for these agents. We’re about to deploy autonomous systems at scale without knowing who’s responsible when things go wrong.

This isn’t a technical problem. It’s an institutional one. And universities should be the first to solve it.

From SEO to AEO: A Metaphor with Real Consequences
#

Gartner just published its strategic predictions for 2026, and one caught my attention: traditional SEO and PPC will give way to Agent Engine Optimization (AEO). What does this mean? That products, services, and content will need to be “machine-readable” — legible and interpretable by AI agents, not just by humans searching on Google.

Let’s translate this to higher education.

If AI agents will be the ones recommending academic programs, helping students choose courses, connecting competencies with job opportunities… are our curricula, our competency frameworks, our content ready to be read by machines?

Most of the time, the answer is no. We have PDFs with study plans. Course descriptions written for academic committees. Competencies defined in documents that nobody reads — not humans, and certainly not AI agents.

The shift from SEO to AEO isn’t just a marketing change. For universities, it’s a shift in curriculum design. Competencies need to be structured, semantic, interconnected. Not because of a tech fad, but because the systems that will guide our students need to understand what we offer.

The Tec-UNESCO Observatory: From Principles to Action
#

On March 18 — ten days ago — Tecnológico de Monterrey and UNESCO signed an agreement in Santiago, Chile to create the Regional Observatory on Artificial Intelligence in Education for Latin America and the Caribbean. This isn’t an empty press release: it includes a $90,000 USD contribution from Tec, technical teams working on methodological frameworks, digital competencies for educators, AI ethics, and pilot projects in Chile, El Salvador, and Mexico.

As Esther Kuisch Laroche, Director of UNESCO’s Regional Office in Santiago, put it: “The challenge is not only to incorporate new technologies, but to ensure that their use contributes to more inclusive, ethical, and relevant education systems.”

This is exactly what we need. No more grandiose declarations about AGI. More concrete work to understand how AI transforms learning, generate evidence, and formulate public policy based on data, not hype.

The Governance Gap Is Our Opportunity
#

Let’s go back to that Deloitte number: 75% of companies will deploy agents, but 80% don’t know how to govern them.

In higher education, the landscape is similar. 86% of university students already use AI regularly. A recent pilot study shows they’re not simply letting AI write for them — they interact, iterate, edit. But our institutional frameworks aren’t designed for this world.

And this is where I see the opportunity. Universities aren’t (or shouldn’t be) mere consumers of technology. We’re the institutions that define ethical frameworks, generate knowledge, and educate the people who will make decisions about AI for decades to come.

If anyone should define how AI agents are governed in educational settings — who’s responsible when an agent gives bad academic guidance, how student privacy is protected, how bias is prevented in recommendation systems — it should be us.

What Comes Next
#

Digital transformation is no longer about digitizing processes. I’ve been saying this for a while, but now the meaning runs deeper than ever.

Digital transformation in 2026 means preparing our institutions for a world where AI agents are colleagues, not tools. Where educational content must be as readable by machines as by humans. Where the governance of autonomous systems is as important as their implementation.

Jensen Huang can declare AGI if he wants. I’d rather focus on what truly matters: that technology serves better, fairer, and more relevant education. And that’s not something you declare — it’s something you build.

Carles Abarca is Vice President of Digital Transformation at Tecnológico de Monterrey.

NemoClaw and the Enterprise Awakening: When AI Agents Leave the Garage

Wed, 18 Mar 2026 00:00:00 +0000

Jensen Huang does not use words carelessly. When the CEO of the most valuable company on Earth stands on stage at GTC and says “OpenClaw is the operating system for personal AI,” he is not making a prediction. He is reading a scoreboard.

And then he drops NemoClaw.

What NemoClaw actually is
#

Let me cut through the press releases. NemoClaw is not a competitor to OpenClaw. It is an enterprise wrapper around it. A single command — one command — installs NVIDIA’s Nemotron models and the OpenShell runtime on top of any existing OpenClaw instance, adding the three things that kept CIOs awake at night:

Privacy routing. A router that decides which queries go to local models (running on your hardware, never leaving your network) and which can safely reach frontier models in the cloud. This is not a minor feature. This is the difference between “interesting demo” and “approved by Legal.”
Sandboxed execution. OpenShell provides an isolated environment where agents can write code, access files, and execute tasks without touching the host system. The agent gets the access it needs to be productive. Nothing more.
Policy-based guardrails. Configurable security boundaries that define what an agent can and cannot do. Not through hope. Through enforcement.

The message from NVIDIA is unmistakable: OpenClaw is infrastructure. Not a project. Not a trend. Infrastructure.

The enterprise problem was never intelligence
#

Here is what most commentary about AI agents gets wrong: the bottleneck was never capability. Claude, GPT, Gemini — they have been capable of executing complex business tasks for over a year. The bottleneck was trust.

No CIO in their right mind would deploy an autonomous agent with access to production systems, employee data, and financial records without three guarantees: that data stays where it should, that the agent operates within defined boundaries, and that every action is auditable. NemoClaw provides exactly those three guarantees.

But NVIDIA is solving a problem that some of us had already started solving from the inside.

AgenTECs: What happens when a university takes agents seriously
#

At Tecnológico de Monterrey, we did not wait for NemoClaw. We could not afford to. When you are responsible for the digital infrastructure serving 100,000 students and 30,000 employees, “wait and see” is not a strategy.

We call them AgenTECs — autonomous institutional agents built on OpenClaw, powered by our own TECgpt models, running on our private Azure infrastructure. They are not chatbots. They are not demos. They are digital collaborators with @tec.mx email accounts, Microsoft Teams presence, and defined roles within the organization.

The architecture is deliberately simple:

Layer	Component	Function
Orchestration	OpenClaw	Lifecycle management, plugins, tools, persistent memory
Cognitive	TECgpt (institutional LLMs)	Reasoning, language understanding, decision-making
Infrastructure	Azure private cloud	Isolated, secure, dedicated resources
Channels	Email @tec.mx, Teams, WhatsApp	Communication with users, supervisors, and other agents

Each AgenTEC operates under a governance model with three clearly separated roles: the Administrator (IT — keeps it running), the Supervisor (business area — tells it what to do), and the User (interacts with it as they would with any colleague). This separation is not bureaucracy. It is the difference between an agent that scales and an agent that becomes someone’s science project.

Why this matters now
#

Three converging forces make this moment different from every previous enterprise AI wave:

First, the cost equation has flipped. An AgenTEC running 24/7 on a dedicated VPS costs a fraction of a human collaborator performing the same function. Not because the human is replaceable — because repetitive, rule-based work should never have been assigned to humans in the first place.

Second, the framework matured before the enterprise noticed. OpenClaw’s skill system, memory persistence, multi-channel communication, and tool integration were built for personal use. But personal AI turns out to be harder than enterprise AI. If an agent can manage someone’s email, calendar, finances, and home automation simultaneously, it can certainly process support tickets.

Third, NVIDIA just removed the last excuse. Every objection I have heard from CISOs and compliance officers — data privacy, execution isolation, audit trails — has a technical answer in NemoClaw. The conversation shifts from “can we?” to “when do we start?”

The playbook
#

For CIOs reading this, here is the practical path:

Start with the mundane. The first AgenTEC we are designing handles TecServices requests — the thousands of monthly queries about credentials, procedures, and administrative processes. Not glamorous. High volume. Perfect for an agent that follows procedures without deviation and escalates anything unusual.

Keep humans in the loop, but redefine the loop. The supervisor does not review every action. They review patterns, exceptions, and escalations. Weekly, not hourly. The agent handles the routine. The human handles the judgment.

Build on open infrastructure. OpenClaw is open source. TECgpt runs on our own cloud. NemoClaw adds NVIDIA’s models locally. At no point does institutional data leave our control. This is not negotiable for a university handling student records, research data, and financial information.

Measure outcomes, not activity. An AgenTEC’s value is not measured by how many messages it processes. It is measured by how many human hours it liberates for higher-value work, how many response times it reduces, and how many errors it eliminates.

The uncomfortable truth
#

Here is what I tell my peers at every CIO forum: the question is not whether AI agents will transform your operations. That is settled. The question is whether you will build your own or rent someone else’s.

NemoClaw makes the “build” option dramatically more accessible. OpenClaw provides the foundation. NVIDIA provides the guardrails. Your institutional knowledge — your processes, your data, your culture — provides the differentiation that no vendor can replicate.

At Tec de Monterrey, we chose to build. We call them AgenTECs. And every month, the gap between what they can do and what we imagined widens — in our favor.

The operating system for AI agents is here. The enterprise wrappers are shipping. The only scarce resource now is the courage to start.

Carles Abarca is VP of Digital Transformation at Tecnológico de Monterrey, where he leads the development of TECgpt and the AgenTECs program. Previously CTO of Banco Sabadell and CIO of TSB Bank.

AI Agents Are No Longer Assisting Scientists. They Are Doing the Science.

Sun, 15 Mar 2026 00:00:00 +0000

In March 2026, something shifted. Not a bigger model. Not a higher benchmark. Something deeper: AI agents stopped being tools that help scientists and started producing scientific knowledge on their own.

This week, three events converged, and I believe they mark a tipping point with no return. The ShipSquad team documented it brilliantly in their analysis “AutoResearch, OpenClaw, Claude Opus 4.6: AI Agents Are Now Doing the Science”, and it inspired me to dig deeper into what this means for scientific research as we know it.

The facts
#

Andrej Karpathy released AutoResearch — a 630-line Python framework that lets an AI agent run hundreds of machine learning experiments autonomously on a single GPU. You let it run overnight. You wake up to a better model and a complete discovery log. In 48 hours, the agent found ~20 improvements no human had identified, cutting GPT-2 training time by 11%.

Anthropic’s Claude Opus 4.6 discovered 22 zero-day vulnerabilities in Firefox — 14 of them high-severity — in just two weeks. For context: those 14 represent nearly one-fifth of all serious Firefox vulnerabilities patched throughout 2025. An AI model matched months of human security research.

Sakana AI unveiled The AI Scientist v2 — an agentic system that generates hypotheses, designs experiments, executes them, analyzes results, and writes the complete scientific paper. The result: the first entirely AI-generated scientific paper accepted through peer review at an academic workshop.

Three events. The same week. The same conclusion: AI agents are now producing original knowledge.

AI has discovered before, but never on its own
#

To understand why this moment is different, trace the evolution:

Era 1: AI as microscope (2018-2023)
#

AI amplified human researchers’ capabilities. Humans asked the questions, designed the experiments, and AI processed data at impossible scales.

AlphaFold (2020-2024) predicted the structure of all 200 million known proteins. A problem that had resisted solution for 50 years. Hassabis and Jumper won the 2024 Nobel Prize in Chemistry for it. But the question — “can we predict protein structures?” — was formulated by humans.
Halicin (2020) — MIT researchers used AI to screen 100 million chemical compounds and discovered a new antibiotic capable of killing multi-drug-resistant bacteria, including the dreaded Acinetobacter baumannii. But the experimental design was human.
GNoME by DeepMind (2023) discovered 2.2 million new crystals, including 380,000 stable materials — multiplying by 10x everything humanity had found in the entire history of materials science. But the evaluation framework was designed by Google researchers.

Era 2: AI as colleague (2024-2025)
#

AI began proposing hypotheses and designing experiments, but under human supervision.

FunSearch by DeepMind (2024) used an LLM to discover new solutions to open problems in pure mathematics — the first time a language model made a genuine discovery in formal sciences.
Insilico Medicine got its molecule rentosertib — designed entirely by generative AI — through a Phase IIa clinical trial with positive results for idiopathic pulmonary fibrosis. From concept to human testing in under 30 months, when the traditional process takes 10-15 years.
MIT Antibiotics-AI Project (2025) moved from discovering existing antibiotics to designing entirely new molecules using generative AI capable of killing resistant bacteria. No longer searching for needles in a haystack; they are manufacturing new needles.

Era 3: AI as autonomous researcher (2026 →)
#

And now, in March 2026, we cross the threshold: AI formulates its own questions, designs its own experiments, executes them, and produces papers accepted by human reviewers. Not science fiction. It is AutoResearch, AI Scientist v2, and Claude Opus doing independent security research.

The Great Acceleration: when AI researches faster than humans can read
#

I considered making a safer prediction — forecasting the percentage growth of AI-authored scientific papers. Instead, here is the unfiltered version.

Before 2030, AI agents will have produced more verifiable scientific discoveries in materials science, drug discovery, and computational mathematics than all human researchers combined in those same disciplines.

This is not hyperbole. It is arithmetic.

Consider the numbers: GNoME discovered in weeks 380,000 stable materials that all of humanity took centuries to accumulate (barely 48,000 by 2023). AutoResearch runs 100 experiments per night — months of a PhD student’s work. And AI Scientist v2 can generate a complete scientific paper in hours, not months.

Now scale that. Not one agent, but thousands. Not one night, but every night. Not one domain, but all of them.

The Compound Acceleration phenomenon
#

What we are witnessing is not linear acceleration. It is compound acceleration: each agent discovery feeds the next research cycle. One agent discovers a new material → another simulates its properties → another designs applications → another writes the paper. In parallel. 24/7. No vacations, no ego, no departmental politics.

Human science operates at publication speed: one paper every 6-18 months. Agentic science will operate at computation speed: one discovery per minute.

Problems we will solve decades ahead of schedule
#

This acceleration is not just quantitative. There are problems we believed were decades away that research agents could solve much sooner:

Antibiotic resistance — AI is already designing new molecules against superbugs. With autonomous agents iterating 24/7 on thousands of variants, we could have a complete new generation of antibiotics before 2030. The WHO estimated that by then, superbugs would kill 10 million people per year. Agents could get there first — for the benefit of all humanity.
Nuclear fusion — The greatest challenge is controlling plasma. DeepMind already used AI to control plasma shape in the TCV tokamak. Autonomous agents simulating and optimizing millions of magnetic configurations could compress decades of experimental research into years.
Cellular aging — AlphaFold solved protein structures. The next frontier is understanding the interactions between proteins, genes, and cellular processes that cause aging. It is a problem of massive combinatorial complexity — exactly the kind of problem where agents excel.
New energy materials — GNoME already opened the door with 380,000 stable materials. Agents systematically exploring that space could find the room-temperature superconductor, the perfect battery electrolyte, or the catalyst that makes industrial carbon capture viable.

The question is no longer whether AI will surpass humans in scientific output. It is when. And my answer is: in many fields, it is already happening. We just have not updated our metrics to measure it.

The consequences nobody wants to discuss
#

1. The end of “batch research”
#

Today, a researcher can run perhaps 5-10 experiments per week. AutoResearch demonstrates that an agent runs 100 per night. When research shifts from a sequential human process to a continuous autonomous one, knowledge production will multiply by orders of magnitude.

2. The radical democratization of discovery
#

Karpathy proved that a single GPU and 630 lines of code are enough for autonomous research. A PhD student in Monterrey, Lagos, or Bangalore can now compete in knowledge production with a Stanford lab. The barrier is no longer budget; it is the ability to formulate good questions and direct agents with precision.

3. The “Orchestra Conductor” as the new scientific role
#

The scientist of the future will not be the one pipetting in the lab or writing code. It will be the one who “programs in Markdown” — who writes the precise instructions that guide squadrons of autonomous agents. This is exactly what Karpathy demonstrates with his program.md file: the future of directing research is not writing better code, but writing better programs in natural language.

This paradigm shift is not new. I explored it in my article “The Next AI Wave: Agents” — where I argued that autonomous agents represent a fundamental shift from assistive AI to autonomous execution. The same applies to science: the researcher becomes the conductor of research agent orchestras.

4. The peer review crisis
#

If AI Scientist v2 already generates papers accepted at workshops, how long until it produces papers accepted at top conferences? And how will we distinguish between a paper written by an agent and one written by a human? The peer review system, designed to evaluate human work, is unprepared for a world where papers are generated at industrial speed. We will need new evaluation frameworks — and possibly, AI agents reviewing the papers of other AI agents.

What this means for universities
#

The institutions that most quickly integrate autonomous research agents into their labs will lead knowledge production in the next decade. It is not about buying bigger GPUs. It is about:

Training “research agent directors” — scientists who know how to formulate questions and direct AI squadrons.
Building autonomous experimentation infrastructure — labs where agents can run thousands of experiments without continuous human oversight.
Redefining authorship and intellectual property — if an AI agent discovers a molecule that cures a disease, who owns the discovery?
Measuring scientific output differently — current indicators (papers, citations, h-index) are metrics designed for human speed. We need metrics that capture agentic velocity.

The race has started. And the window of opportunity to position yourself is now.

Conclusion
#

In just six years, AI went from predicting protein structures to discovering antibiotics, from designing materials to writing scientific papers, from finding known vulnerabilities to discovering zero-days. The trajectory is clear and accelerating exponentially.

AI agents will not replace scientists. They will make scientists who do not use them irrelevant. And that transition, unlike the one we lived through in the software industry, will not take a decade. It will take months.

Welcome to the era of the Great Scientific Acceleration. Those who understand it first will lead the science of the future. Those who ignore it will read about it in papers written by machines.

🤖 Go deeper with AI
#

Want to explore further? Click your favorite AI with a ready-made prompt:

On the scientific acceleration with AI agents: ChatGPT · Perplexity · Claude

On the future of peer review: ChatGPT · Perplexity · Claude

On the impact for universities: ChatGPT · Perplexity · Claude

The Chart That Predicts Which Jobs AI Will Kill (And They're Not the Ones You Think)

Mon, 09 Mar 2026 00:00:00 +0000

Look at this chart carefully. It’s not an analysis of what AI has destroyed. It’s an X-ray of what it’s about to destroy.

Source: Anthropic — Labor market impacts of AI (March 2026)

The blue area is what AI can do today. The red area is what AI is doing today. The difference between them isn’t a safety margin. It’s a tsunami that hasn’t hit shore yet.

The Study: 2 Million Conversations with Claude
#

Anthropic just published Labor market impacts of AI: A new measure and early evidence, and it’s the most rigorous analysis I’ve seen on AI’s real employment impact.

What did they do? They crossed three data sources:

The O*NET database, cataloging tasks across ~800 US occupations.
Real Claude usage data — 2 million conversations analyzed via the Anthropic Economic Index.
Theoretical estimates from Eloundou et al. (2023) on which tasks an LLM can make at least twice as fast.

The result is a new metric: observed exposure — not what AI could theoretically do, but what it’s actually doing in professional settings. And the most revealing finding isn’t the absolute numbers — it’s the gap between the two.

The 10 Most Exposed Jobs
#

The ranking won’t surprise anyone who’s been paying attention, but the numbers are brutal:

Computer Programmers — 75% coverage. Three out of four tasks a programmer does, Claude already handles.
Customer Service Representatives. First-party API traffic shows massive automation.
Data Entry Keyers — 67%. Reading documents and entering data. The perfect automation use case.

The list continues: actuaries, financial analysts, technical writers. Office jobs. White-collar work. People with college degrees.

On the other end, 30% of workers have zero exposure. Cooks, motorcycle mechanics, lifeguards, bartenders. Jobs where hands, bodies, and physical context are irreplaceable.

Ironic, isn’t it? Decades telling us automation was coming for manual labor. It’s coming for the desks.

The Demographic Surprise
#

This is where the study shatters the dominant narrative.

The workers most exposed to AI are:

16 percentage points more likely to be female
11 points more likely to be white
Nearly twice as likely to be Asian
Earn 47% more than unexposed workers
17.4% hold graduate degrees (vs. 4.5% in the unexposed group)

This isn’t the displaced factory worker narrative. These are lawyers, analysts, programmers, university professors. The professional class that thought it was untouchable.

When I say this will reshape social structure, I’m not exaggerating.

The Gap IS the Prediction
#

Go back to the chart. Look at the categories:

Computer & Math: 94% theoretical capability, 33% actual use
Legal: ~85% theoretical, less than 15% observed
Education: ~70% theoretical, less than 15% observed
Office & Admin: 90% theoretical, a fraction of actual use

That distance between blue and red isn’t comfort. It’s latency.

It’s the time companies take to adopt, regulators to adapt, workflows to reconfigure. But the technology is already there. The model already knows how. The ecosystem just needs to catch up.

And every month, the red area grows. Anthropic says it explicitly: “As capabilities advance, adoption spreads, and deployment deepens, the red area will grow to cover the blue.”

This isn’t speculative prediction. It’s an empirical observation with trajectory.

What Changes with AI Agents
#

Here’s the factor the study doesn’t fully measure — because it didn’t exist at this scale when they collected the data.

The study analyzes LLM usage — conversations with Claude. Chat interactions. A human asks, the AI answers. It’s the augmentation model: AI helps you, you execute.

But AI agents are something else entirely. They don’t answer — they act. They execute task chains autonomously. They navigate systems, make intermediate decisions, complete entire workflows without human intervention.

What we’re building at Tecnológico de Monterrey with AgenTECs is exactly this. Not a chatbot that helps you draft an email. An agent that manages the entire process: reads context, drafts, sends, follows up, escalates if there’s no response.

When agents arrive at enterprise scale — and they’re already arriving — the red area in the chart will expand explosively. Because you no longer need a human interacting with AI task by task. The agent covers the entire role.

Think about the Legal category: 85% theoretical capability, <15% current use. What happens when an agent can review contracts, identify risk clauses, generate executive summaries, and prepare response drafts — all without a lawyer touching the keyboard? The 85% becomes the new floor, not the ceiling.

What to Do (Which Is Not Panic)
#

I’ve been saying the same thing for years: this isn’t about fear. It’s about preparation.

When I wrote “El fin del desarrollador” on LinkedIn, the reaction was predictable: “exaggerated,” “developers will always be needed,” “AI can’t do X.” The same arguments I heard about TECgpt when we launched it and people said professors would never use it. Today we have over 60,000 active users.

The metaphor I use is the orchestra conductor. The value is no longer in playing the violin — it’s in knowing what music to perform, who plays what, and when to change the score. Future professionals don’t execute tasks — they orchestrate systems that execute them.

Specifically:

Massive upskilling, now. Not “intro to AI” courses — real training on production tools.
Redefine roles, don’t eliminate them. A lawyer who masters AI agents is worth more, not less. But a lawyer who only knows manual contract review has an expiration date.
Measure exposure in your organization. Use Anthropic’s framework. Identify which tasks in each role an LLM can already perform. Design the transition before it’s imposed on you.
Create new roles that don’t exist yet: AI orchestrators, agent prompt engineers, autonomous systems supervisors.

The Bottom Line
#

The BLS projects that the most exposed occupations under this metric will grow less through 2034. For every 10 points of observed coverage, the growth projection drops 0.6 percentage points. This isn’t casual correlation — labor market analysts are seeing the same thing.

And yet, Anthropic also finds that there’s no systematic increase in unemployment in the most exposed professions. Yet.

That’s the window. We’re in the moment between seeing the lightning and hearing the thunder. The bolt already struck. The question isn’t whether the sound will arrive, but whether you’ll be ready when it does.

Those who read this chart as “AI hasn’t affected employment much yet” are confusing latency with safety. Those who read it as “a structural labor market shift is coming and we need to act now”… they’re the ones who’ll still be conducting the orchestra.

China's AI Pincer Move: Qwen 3.5 and CoPaw Are Not a Warning Shot — They're the Main Event

Thu, 05 Mar 2026 00:00:00 +0000

There is a moment in every technology race when “catching up” becomes “setting the pace.” For China’s AI ecosystem, that moment is now.

In the span of a few weeks, Alibaba has released two things that, taken separately, would each be significant. Taken together, they represent a strategic vision that should make every Western AI executive lose sleep.

Qwen 3.5: a family of open-source models that beats GPT-5.2 on instruction following and leads the field on vision benchmarks. Apache 2.0 licensed. Free. Commercial use allowed.

CoPaw: an open-source personal AI agent workstation — think OpenClaw, but from Alibaba’s AgentScope team — with persistent memory, custom skills, multi-channel support, and browser automation.

Models and infrastructure. The brain and the body.

This is not a warning shot. This is a strategy.

The Qwen 3.5 Story: Frontier AI Goes Free
#

Let me give you the numbers first, because they tell a story.

Qwen 3.5’s flagship model uses a Mixture of Experts (MoE) architecture with 397 billion total parameters but only 17 billion active at any given time. Read that again. You get frontier-class performance while only running the compute cost of a 17B model.

The benchmarks are not “competitive.” They are leading:

IFBench (instruction following): 76.5 — beating GPT-5.2’s 75.4
SWE-bench (coding): 76.4 — neck and neck with the best
MMMU (vision): 85.0 — outright leader
256K token context window
201 languages supported
Thinking and non-thinking modes — you choose the tradeoff between depth and speed

The model family was released in three waves between February and March 2026: flagship, medium, and small. The small models — 0.8B to 9B parameters — are explicitly designed for on-device deployment. Your phone. Your laptop. Your edge server. No API call required.

Let that sink in for a moment.

A year ago, running anything close to frontier AI locally was a fantasy. Today, Alibaba is handing you models that compete with the best in the world, under the most permissive open-source license available, optimized to run on your hardware.

The MoE architecture is the key unlock here. Traditional dense models force you to choose: either you run a massive model with massive compute, or you run a small model with limited capability. MoE breaks that tradeoff. Qwen 3.5 has the knowledge of a 397B model but the inference cost of a 17B one. It is, in practical terms, the democratization of frontier AI.

And it is Apache 2.0. Not “open-ish.” Not “you can look but not touch.” Fully open. Fork it. Fine-tune it. Ship it in your product. Alibaba does not care. Or rather — they care very much, but their game is not licensing revenue.

CoPaw: The Agent Layer China Was Missing
#

Models without infrastructure are academic papers. Infrastructure without models is empty plumbing. The interesting move is doing both.

CoPaw (copaw.bot) launched in March 2026 from Alibaba’s AgentScope team. If you are familiar with OpenClaw — and if you read my blog, you probably are — CoPaw is China’s answer to it. An open-source personal AI agent workstation that turns language models into persistent, capable digital workers.

The feature list reads like someone studied every agent platform on the market and built a synthesis:

ReMe: persistent memory across sessions. Your agent remembers context, preferences, past interactions. Not a gimmick — this is what separates a chatbot from an actual assistant.
Custom skills: build and import capabilities. Pull from clawhub.ai, skills.sh, or GitHub.
Multi-channel: DingTalk, Feishu, iMessage, Discord, QQ. Your agent lives where you work.
Cron scheduling: automated tasks on a schedule. Check my email every morning. Summarize my feeds at 6 PM. The basics, done right.
Browser automation: your agent can navigate the web, fill forms, extract data. The hands to go with the brain.
MCP Server integration: the emerging standard for tool use. CoPaw speaks it natively.

Is it perfect? No. WhatsApp and Telegram support are missing — a significant gap for Western and Latin American users. Multi-agent orchestration is not there yet. OpenRouter integration is absent. These are real limitations.

But here is what matters: CoPaw is not a prototype. It is a platform. And it is open-source, which means the community can fill those gaps faster than any corporate roadmap.

I have been running OpenClaw as my personal agent infrastructure for months — it is literally what powers JarvisX, my AI assistant. So I understand this space intimately. CoPaw is not a clone. It is a parallel evolution, built from a different set of assumptions (Chinese messaging ecosystem, AgentScope framework, different privacy model) that arrives at remarkably similar conclusions about what an AI agent workstation needs to be.

That convergence is the signal. When two teams on opposite sides of the world, working independently, build essentially the same thing — that is not coincidence. That is the shape of the future becoming obvious.

The Earthquake Started in January 2025
#

None of this is happening in a vacuum. Let me connect the dots.

January 2025: DeepSeek releases R1, an open-source reasoning model that shocks the industry. Silicon Valley’s reaction ranges from dismissal to panic, settling on grudging respect. The “China can’t do AI” narrative dies overnight.

Throughout 2025: Chinese labs iterate at a pace that makes Western release cycles look glacial. Qwen, DeepSeek, Yi, GLM — each generation closing the gap further. The MoE architecture becomes the standard approach, driven by the practical reality that Chinese labs face compute constraints from US export controls and have to be more efficient, not less.

Here is the irony that should keep policymakers awake: export controls designed to slow China’s AI development may have accelerated their innovation in efficiency. When you cannot buy the biggest GPUs, you learn to do more with less. And “more with less” turns out to be exactly what the market wants.

February-March 2026: Qwen 3.5 arrives, not as a single model but as an ecosystem play. Flagship for the cloud, medium for the server room, small for the device. And simultaneously, CoPaw launches to provide the agent layer. Models plus infrastructure plus ecosystem.

This is not “China catching up.” This is China executing a full-stack AI strategy while much of the West is still arguing about whether to charge $200/month or $2,000/month for API access.

The Alibaba Strategy: OpenAI’s Vision, Open-Source’s Price
#

Let me be explicit about what Alibaba is doing, because I think most Western observers are misreading it.

OpenAI’s vision has always been: build the best models, then build the infrastructure to deploy them, then build the ecosystem of applications on top. Vertical integration. The “Apple of AI.”

Alibaba’s vision is the same — except open-source.

Best models? Qwen 3.5 is demonstrably frontier-competitive. Infrastructure? CoPaw provides the agent layer. AgentScope provides the framework. Ecosystem? Apache 2.0 means anyone can build on it.

The difference is the business model. OpenAI charges for access. Alibaba gives away the technology and monetizes the cloud (Alibaba Cloud), the commerce (Alibaba platforms), and the enterprise services built on top. The AI itself is the loss leader. Or rather, it is the moat around everything else.

This is not charity. It is strategy. And it is devastatingly effective.

If you are an enterprise CTO today — and I have been one, at Banco Sabadell, so I know the calculus — the question on your desk is uncomfortable: Why am I paying for proprietary AI models when open-source alternatives match or beat them on benchmarks?

The answers used to be: reliability, support, safety, compliance. Those are real. But they are eroding fast. Qwen 3.5 is not some garage project. It is backed by one of the largest technology companies on Earth. It has enterprise-grade documentation. It runs in production at Alibaba scale.

The moat is getting shallow.

What This Means for the West
#

I am not writing this as a China cheerleader or a Western doomer. I am writing it as someone who has spent 20+ years in enterprise technology and is currently leading digital transformation at one of Latin America’s largest universities.

Here is what I think this means:

For AI companies (OpenAI, Anthropic, Google): The “best model” advantage is now measured in months, not years. If Qwen 3.5 can match GPT-5.2 today, Qwen 4 will likely match whatever comes next. The sustainable advantage must come from ecosystem, trust, and integration — not model quality alone. The race to the bottom on model pricing accelerates.

For enterprises: Your AI strategy cannot depend on a single provider. The multi-model, multi-provider approach is no longer a nice-to-have — it is risk management. And if you are not evaluating open-source models for your use cases, you are leaving money and optionality on the table.

For the open-source movement: This is your moment. China’s largest tech companies are pouring billions into open-source AI, not because they are altruistic, but because it serves their strategic interests. The result is the same: the commons gets richer. Western open-source advocates should take notes on how to align corporate strategy with community benefit.

For developers: Learn to run local models. Understand MoE architectures. Get comfortable with agent frameworks — both OpenClaw and CoPaw. The developers who thrive in 2027 will be the ones who can deploy and orchestrate AI agents across multiple models and platforms, not the ones locked into a single API.

For policymakers: The export control strategy needs a fundamental rethink. Restricting compute has not prevented frontier AI development in China — it has redirected it toward efficiency innovations that may ultimately be more valuable than brute-force scaling. The horse has left the barn, and the barn is on fire.

The Democratization Paradox
#

Here is the question that keeps me up at night: if frontier AI is free and open, what is the moat?

Not for Alibaba — their moat is their ecosystem. Not for OpenAI — their moat is their brand and enterprise relationships. I mean for everyone else. For the thousands of SaaS companies, AI startups, and technology consultancies that have built their value proposition around access to AI capabilities.

When Qwen 3.5 is free, when CoPaw is free, when the entire stack from model to agent to deployment is open-source and commercially licensable — what exactly are you selling?

The answer, I think, is the same answer it has always been in technology: domain expertise, integration quality, trust, and speed of execution. The tools become commoditized. The craft does not.

But that is a much harder business than “we have access to AI and you don’t.” And it will cause a shakeout that makes the SaaSpocalypse look like a rehearsal.

What I Am Doing About It
#

I never write about things I am not willing to act on. So here is what this means for my work:

At Tec de Monterrey, we are actively evaluating open-source models for educational applications where data sovereignty matters — and with a Latin American university serving students across multiple countries, it matters a lot. Qwen 3.5’s multilingual support (201 languages, with strong Spanish coverage) makes it a serious candidate.

Personally, I run my AI agent infrastructure on OpenClaw. CoPaw’s release is not a threat to that — it is validation. The agent workstation pattern is the right abstraction. And competition drives improvement. I fully expect OpenClaw and CoPaw to cross-pollinate features, especially given that CoPaw can already import skills from clawhub.ai.

The future I see is heterogeneous. Not “Western AI vs. Chinese AI” but a global ecosystem where the best models and tools win regardless of origin. Where an enterprise in Mexico City runs Qwen for some tasks, Claude for others, and Gemini for a third — all orchestrated by agent infrastructure that does not care about the nationality of the model.

That is not a geopolitical statement. It is an engineering reality.

The Bottom Line
#

Alibaba has executed a textbook pincer move: world-class models on one side, agent infrastructure on the other. Qwen 3.5 gives you the brain. CoPaw gives you the body. Both are free. Both are open. Both are production-ready.

The West still leads in many dimensions — safety research, alignment, enterprise trust, regulatory frameworks. Those matter. But the raw capability gap? It is closing so fast that by the time you finish reading this article, it may have closed a little more.

If you are a technology leader and you are not paying attention to what is coming out of China, you are not paying attention.

And in this industry, not paying attention is how you become the next $300 billion cautionary tale.

Anthropic vs. the Pentagon: When AI Ethics Collides with Military Power

Tue, 03 Mar 2026 00:00:00 +0000

Last week we witnessed something unprecedented in the history of artificial intelligence: an AI company standing up to the Pentagon and saying “no.”

Anthropic, creator of Claude — the only AI model currently authorized on the U.S. federal government’s classified systems — rejected the final terms of a $200 million contract with the Department of Defense. The consequences were immediate and brutal.

The Red Line
#

The conflict boiled down to two non-negotiable points for Anthropic:

Mass surveillance of American citizens. The Pentagon wanted to use Claude to analyze bulk-collected data: search histories, GPS movements, credit card transactions, even the questions you ask your favorite chatbot. All cross-referenced to build profiles.
Autonomous weapons. Systems that select and engage targets without a human making the final call. The 2026 military budget allocates $13.4 billion to these weapons alone.

Anthropic didn’t argue that such weapons shouldn’t exist. In fact, they offered to work directly with the Pentagon to improve their reliability. But they determined that current AI models aren’t reliable enough to make lethal decisions autonomously. The risk of indiscriminate fire, civilian casualties, or even harm to American troops was, in their analysis, too high.

The False Solution of “Cloud vs. Edge”
#

During negotiations, a proposal emerged: keep Anthropic’s AI in the cloud, out of the weapons themselves. The models would synthesize intelligence before an operation but wouldn’t make kill decisions. The AI’s hands would stay clean.

Anthropic rejected this with a devastating technical argument: in modern military AI architectures, the distinction between cloud and edge no longer exists. Drones operate through mesh networks connected to data centers. The Pentagon actively works to push computing closer to the battlefield. If a model in an AWS server in Virginia is making combat decisions, ethically there’s no difference from it being inside the drone.

The Response: The Hammer
#

When Anthropic held its ground, the response was swift:

Trump ordered all federal agencies to cease using Anthropic’s technology.
Pete Hegseth (Defense Secretary) designated Anthropic a supply chain risk to national security, barring any military contractor from doing business with the company.
OpenAI announced a Pentagon deal just hours later.

The message was clear: play by our rules, or we destroy you.

What Sam Altman Didn’t Explain
#

Here’s the most troubling part. Days before the collapse, Sam Altman had publicly declared that OpenAI would also refuse to let its models be used in autonomous weapons. Solidarity with Anthropic.

But while making those statements, he was already negotiating with the Pentagon. And he closed the deal hours after Anthropic’s fall, ensuring his AI would only be deployed “in the cloud” — exactly the solution Anthropic dismissed as insufficient.

Nearly 100 OpenAI employees signed an open letter supporting the same red lines as Anthropic. Altman will have to explain on Monday why what Anthropic rejected on principle, he accepted for business.

What’s Really at Stake
#

This crisis transcends a contract. It reveals three fundamental fractures:

1. AI as a geopolitical weapon. AI technology is no longer just a commercial product. It’s a strategic military asset, and governments are willing to use their full power to control it.

2. The illusion of self-regulation. Anthropic tried to set ethical limits from within. The response was a national security risk designation. What company will dare say “no” after this?

3. The gap between words and action. OpenAI talked principles and signed a check. It’s not the first time, and the industry should take note.

My Take
#

I’ve spent over 20 years in technology, and I’ve seen many inflection points. This is one of them.

Anthropic did something extraordinarily rare in the tech industry: sacrifice $200 million and their federal government access for an ethical position. We can debate whether it was a smart business decision, but we can’t deny it was brave.

What concerns me isn’t Anthropic — they’ll survive. What concerns me is the precedent. If an AI company that puts ethical limits on its technology can be designated a “national security risk,” we’re building a system where the only option is blind obedience.

And obedient AI without restrictions, in the hands of unchecked power, is exactly the scenario that every AI safety researcher has been warning about for years.

The question is no longer whether AI will transform warfare. The question is who decides the limits.

What do you think of Anthropic’s stance? Principles or naivety? I’d love to hear your perspective on LinkedIn or X.

My AI Agent Is 2 Weeks Old. 72,000 Lines of Code. 5 Projects Shipped.

Wed, 25 Feb 2026 00:00:00 +0000

I haven’t written a line of code in weeks. And yet, my repositories keep growing.

72,563 lines added. 43 commits. 5 projects deployed to production.

All in 14 days. I didn’t write any of it. My AI agent did.

A Clarification First
#

Let me be clear about something: my job as VP of Digital Transformation at Tecnológico de Monterrey is not to write code. I lead teams that do that at scale, for one of Latin America’s largest universities.

But I’ve been writing code for 40 years. From assembler to C, Fortran, Visual Basic, Pascal, COBOL, C++, Java, JavaScript, TypeScript, Rust, Python — I’ve touched them all. First because it was my job. Then because it became a habit I never wanted to break.

Why? Because staying close to the code keeps me close to reality. It helps me make better decisions. It lets me have real conversations with my technical teams — not as a manager who reads reports, but as someone who understands what they’re building.

The projects I’m about to describe are personal. Some are proofs of concept that later evolve into institutional initiatives — like TECgpt Desktop, which started as an experiment on my laptop before becoming an official tool at the university. Others are simply things I want to build.

This matters for what comes next.

Meet My Digital Apprentice
#

His name is JarvisX. He runs on a Mac Mini in my home office. He’s been “alive” for exactly two weeks.

But here’s the key: I decide the architecture. I choose the tech stack. I define the structure, the patterns, the conventions. JarvisX executes — brilliantly, tirelessly, and autonomously — but the technical vision is mine.

He’s not replacing my judgment. He’s amplifying it.

When I say “build this feature,” I’ve already decided how it should be built. JarvisX handles the implementation: writing the code, deploying to production, opening a browser to verify it looks right, and committing the changes. All without asking me to approve each step.

The Numbers
#

In two weeks, JarvisX has:

Added 72,563 lines of code across multiple repositories
Made 43 commits
Shipped 5 complete projects to production
Configured deployments, authentication systems, and cloud infrastructure
Validated his own work visually — taking screenshots, iterating until it’s right

Let me be honest: not all 72,000 lines are hand-crafted code. Some are frameworks, themes, and data files. But every commit, every deploy, every design decision was autonomous.

How This Article Got Here
#

Want to see what working with an AI agent actually looks like? Here’s the conversation that published this very article:

Me: JarvisX, I left a new article for my blog in our shared Dropbox folder.

JarvisX: Read it! I see it’s in English. Do you want me to generate the Spanish version?

Me: Yes. I trust your translations by now, so no need for me to review. Upload both versions directly.

JarvisX: Thanks for the trust! Uploading now. Should I add a featured image?

Me: Yes. Something that evokes an orchestra conductor surrounded by AI agents.

JarvisX: Got it. Do you want to review it first?

Me: No. Just upload it.

That’s it. No back-and-forth. No micromanagement. Just trust built over two weeks of working together.

By the time you’re reading this, JarvisX has already translated it, generated the image, and deployed both versions to production. I didn’t check his work. I didn’t need to.

The Real Insight
#

A year ago, I wrote an article called The End of the Developer. My thesis: the junior developer role would disappear. The future would belong to “directors of orchestra” who conduct AI agents rather than write code themselves.

I’m now living that prediction.

But here’s what I’ve learned: You can’t direct what you don’t understand.

The “new developer” — the one who orchestrates AI agents — still needs deep technical knowledge. Not to write every line, but to:

Define architecture that makes sense
Recognize good code when the agent produces it
Catch mistakes before they reach production
Make tradeoffs that require experience to evaluate

My 40 years of coding didn’t become obsolete when I started working with JarvisX. They became essential. I can direct him effectively precisely because I know what good looks like.

What This Means
#

If you’re a senior developer or architect, this is your moment. Your experience is more valuable than ever — not for typing code, but for directing those who do.

If you’re a junior developer, the path forward isn’t to compete with AI at writing code. It’s to accelerate your learning so you can direct AI sooner. Use these tools to learn faster, not to avoid learning.

If you’re a CIO or CTO, stop asking “how do we adopt AI agents?” Start asking “do our people have the technical depth to direct them well?”

The Bottom Line
#

72,000 lines. 5 projects. 2 weeks. Zero lines written by me.

But every architectural decision? Mine. Every technology choice? Mine. Every quality standard? Mine.

The future of software development isn’t about writing less code. It’s about directing more of it — with the wisdom that only experience can provide.

Carles Abarca is VP of Digital Transformation at Tecnológico de Monterrey and former CTO of Banco Sabadell. He has been writing code for 40 years and plans to never stop — even if he’s no longer the one typing it.

The CIO's Guide to AI Agents: Before You Buy, Ask These 5 Questions

Mon, 23 Feb 2026 00:00:00 +0000

The AI agent market will reach $236 billion by 2034. Every vendor wants your budget. Here’s how to separate hype from value.

Every enterprise software vendor now sells “AI agents.” Salesforce has Agentforce. ServiceNow acquired Moveworks for $2.85 billion. Microsoft promises an “agentic enterprise” through Copilot. Startups like Sierra have hit $100 million ARR in under two years.

The pressure to act is real. 89% of CIOs consider agent-based AI a strategic priority. 51% of large enterprises have already deployed agentic AI. Your board is asking questions. Your competitors are moving.

But here’s what nobody tells you: most AI agent purchases fail to deliver expected value. Not because the technology doesn’t work — but because organizations buy the wrong solution for their specific situation.

After two decades leading technology transformation at a major European bank and now driving digital innovation at one of Latin America’s largest universities, I’ve developed a framework for evaluating AI agents that cuts through vendor hype. Before you sign that contract, ask these five questions.

The AI Agent Landscape in 2026
#

First, let’s understand what you’re buying. The market has three distinct categories:

1. Embedded Agents from SaaS Vendors

Your existing vendors are adding agents to their platforms:

Salesforce Agentforce: $0.10 per action or $125-550/user/month
ServiceNow AI Agents: Full orchestration with AI Control Tower
Microsoft Copilot Studio: Included with M365, plus add-ons
Zendesk AI Agents: $1.50 per autonomous resolution

The pitch: “You already use our platform. Now it’s smarter.”

2. Pure-Play Agent Startups

Companies building agents from the ground up:

Sierra (Bret Taylor, ex-Salesforce): $10B valuation, focused on customer service
Adept: Targeting workflow automation
Imbue, Reflection AI: Research-driven approaches

The pitch: “We’re not constrained by legacy architecture.”

3. Foundation Model Providers

The companies building the AI itself:

Anthropic: Claude with Computer Use and MCP (Model Context Protocol)
OpenAI: GPT-4 with Operator
Google: Gemini with Agentspace

The pitch: “Build custom agents on our infrastructure.”

Each category has trade-offs. Your job is to understand which trade-offs matter for your organization.

Question 1: What Problem Are You Actually Solving?
#

This sounds obvious. It isn’t.

“We need AI agents” is not a problem statement. Neither is “we need to reduce costs” or “we need to be more innovative.”

A proper problem statement looks like this:

“Our customer service team handles 50,000 tickets per month. 60% are password resets and order status checks. Average handling time is 8 minutes. We need to reduce that to under 2 minutes for routine inquiries.”
“Our compliance team manually reviews 10,000 transactions daily for AML screening. False positive rate is 95%. We need to reduce false positives while maintaining regulatory coverage.”

Notice the difference? Specific process. Measurable baseline. Clear target.

The trap: Vendors will happily sell you a “general-purpose agent platform” that can theoretically do anything. In practice, these platforms do nothing well. Start with one high-value, well-defined use case. Prove it works. Then expand.

Red flag: If you can’t articulate the specific process, current metrics, and target improvement, you’re not ready to buy.

Question 2: Build, Buy, or Extend?
#

You have three paths. Let me illustrate each with real scenarios.

EXTEND: Add Agent Capabilities to Your Existing Platforms
#

What it means: You already use Salesforce, ServiceNow, or Microsoft 365. You activate their built-in agent features.

Real example — Global Retailer with Salesforce: A retail company with 200 customer service reps was already running Service Cloud. They activated Agentforce Service Agent in three weeks. Configuration, not development. The agent now handles 40% of incoming inquiries (order status, return policies, store hours) without human intervention. Cost: $0.10 per agent action. No new vendor relationship. No integration project.

Real example — Insurance Company with ServiceNow: An insurer using ServiceNow ITSM enabled AI Agents for incident categorization and routing. The agent reads incoming tickets, identifies the affected system, assigns priority, and routes to the right team. Implementation: 6 weeks. Result: 60% reduction in misrouted tickets, 25% faster resolution times.

When to Extend:

You’re already paying for the platform
Your use case is common (customer service, IT helpdesk, HR inquiries)
You need results in weeks, not months
You don’t have AI engineering talent

The limitation: You’re constrained by what your vendor offers. If Salesforce Agentforce doesn’t support your specific workflow, you’re stuck waiting for their roadmap.

BUY: Purchase a Specialized Agent Solution
#

What it means: You bring in a startup or specialized vendor that does one thing exceptionally well.

Real example — E-commerce Brand with Sierra: A direct-to-consumer brand with $500M revenue wasn’t satisfied with their existing chatbot. They deployed Sierra for customer service. Sierra’s agents handle complex conversations: processing returns while suggesting alternatives, managing subscription modifications, resolving billing disputes. The agents access their Shopify backend, payment processor, and shipping systems. Implementation: 4 months. Result: 70% of conversations resolved without human escalation. Customer satisfaction increased 15 points.

Real example — Enterprise IT with Moveworks (now ServiceNow): Before the acquisition, companies bought Moveworks specifically for IT helpdesk automation. The agent could reset passwords, provision software, troubleshoot VPN issues, and answer policy questions — all through natural conversation in Slack or Teams. It understood context: “I can’t access the server” triggered different workflows than “I need Photoshop installed.”

When to Buy:

Your existing vendor’s agents aren’t good enough for a strategic use case
A startup has proven traction in your specific domain
You can afford 3-6 months implementation
The use case is important enough to justify a new vendor relationship

The risk: Startup viability. What happens if Sierra gets acquired? If the startup pivots? You’re dependent on a company that may not exist in five years. Mitigate this by ensuring data portability and avoiding deep customizations that lock you in.

BUILD: Create Custom Agents Using Foundation Models
#

What it means: You use Claude, GPT-4, or Gemini APIs to build agents tailored to your unique processes.

Real example — Investment Bank’s Deal Analysis Agent: A bulge-bracket bank built a custom agent for M&A analysts. The agent ingests SEC filings, earnings transcripts, news articles, and internal research. Analysts ask natural language questions: “What are the key risks in Acme Corp’s debt structure?” or “Compare the margin profile of these three acquisition targets.” The agent synthesizes information that would take an analyst hours to compile manually. Built on Claude with custom RAG (retrieval-augmented generation) over proprietary databases. Development: 8 months. Team: 6 engineers, 2 ML specialists. The output is proprietary competitive advantage — no vendor offers this.

Real example — Pharmaceutical Company’s Clinical Trial Agent: A pharma company built an agent to monitor clinical trial data in real-time. The agent identifies adverse event patterns, flags protocol deviations, and generates regulatory-ready reports. This isn’t a use case any vendor serves — the domain knowledge is too specialized, the regulatory requirements too specific. Built on GPT-4 with extensive fine-tuning and custom safety guardrails. Development: 12 months.

Real example — University’s Custom Academic Ecosystem (TecGPT): At Tecnológico de Monterrey, we built TecGPT — our academic AI ecosystem, integrated with our LMS, assisting both professors and students, and aligned with our academic regulations. No vendor could offer what we needed out-of-the-box.

When to Build:

Your process is genuinely unique and defines competitive advantage
No vendor solution exists for your domain
You have AI/ML engineering talent (or can acquire it)
You can invest 6-12 months before seeing production value
The long-term value justifies the ongoing maintenance cost

The reality check: Building is expensive. A custom agent isn’t a one-time project — it’s an ongoing product. You need engineers to maintain it, improve it, and adapt it as foundation models evolve. Budget for 2-3 FTEs indefinitely, not just the initial build.

The Decision Matrix
#

Factor	Extend	Buy	Build
Time to value	Weeks	3-6 months	6-12 months
Customization	Low	Medium	Unlimited
Vendor risk	Low	High	Medium
Talent required	Admins	Integrators	Engineers + ML
Total cost (3 years)	$$	$$$	$-$$$$
Competitive advantage	None	Low	High

My Recommendation
#

Start with Extend for your first agent deployment. Even if it’s not perfect, you’ll learn what works, what users actually need, and where the gaps are. That learning is invaluable before you commit to Buy or Build.

Move to Buy when you’ve proven agent value and need capabilities your platform vendor doesn’t offer. Choose startups with strong traction, clear use case focus, and enterprise customers who can serve as references.

Build only when you’ve exhausted Extend and Buy options, or when the process is so unique that it genuinely differentiates your business. If you’re building agents for commodity processes (customer service, IT helpdesk), you’re wasting engineering talent that could create actual competitive advantage.

Question 3: How Will You Measure Success?
#

Before deployment, define:

Efficiency metrics:

Time saved per task
Tasks handled without human intervention
Error rate reduction

Quality metrics:

Customer satisfaction (for customer-facing agents)
Compliance accuracy (for regulatory processes)
Decision quality (measured against human expert baseline)

Business metrics:

Cost per transaction
Revenue impact (if applicable)
Employee satisfaction (agents should help workers, not threaten them)

The 90-day rule: If you can’t demonstrate measurable improvement within 90 days, something is wrong. Either the use case was poorly chosen, the implementation was flawed, or the vendor oversold capabilities. Don’t extend pilots indefinitely hoping for results.

Hidden metric: Adoption. The most sophisticated agent is worthless if your team doesn’t use it. Track actual usage, not just availability.

Question 4: What’s Your Human-in-the-Loop Strategy?
#

In 2026, no responsible organization deploys fully autonomous agents for consequential decisions. The question is where to place human oversight.

The spectrum:

Human-initiated: Human starts task, agent assists, human approves result
Agent-initiated with approval: Agent identifies opportunity, proposes action, human approves
Agent-executed with audit: Agent acts autonomously, human reviews after the fact
Fully autonomous: Agent acts without oversight (appropriate only for low-risk, reversible actions)

My framework for choosing:

Risk Level	Reversibility	Recommended Approach
Low	Easy to reverse	Fully autonomous
Low	Hard to reverse	Agent-executed with audit
High	Easy to reverse	Agent-initiated with approval
High	Hard to reverse	Human-initiated only

Examples:

Password reset → Fully autonomous (low risk, reversible)
Customer refund under $50 → Agent-executed with audit
Credit decision → Agent-initiated with approval
Regulatory filing → Human-initiated only

The regulatory reality: Financial services, healthcare, and other regulated industries will require human oversight for most consequential decisions for the foreseeable future. Plan for this. Agents that “recommend and explain” are more valuable than agents that “decide and act” in these contexts.

Question 5: What’s Your Data Strategy?
#

AI agents are only as good as the data they can access. Before buying, audit:

Data availability:

Can the agent access all systems needed for the use case?
Are APIs available, or will you need custom integrations?
What’s the latency? Agents that wait 30 seconds for data lookups frustrate users.

Data quality:

Is your data accurate and up-to-date?
Are there known data quality issues that will cause agent errors?
Who’s responsible for data hygiene?

Data governance:

What data can the agent access? What’s off-limits?
How do you prevent agents from exposing sensitive information?
What audit trail exists for agent data access?

The integration tax: Most enterprises underestimate integration effort by 3-5x. If the vendor says “we integrate with everything,” ask for customer references running your specific system combination. Generic claims mean nothing.

MCP and the future: Anthropic’s Model Context Protocol (MCP) is emerging as a standard for agent-to-system communication. Consider whether your chosen platform supports open standards or locks you into proprietary integrations.

The Decision Framework
#

After answering these five questions, you should be able to complete this statement:

“We will deploy [specific agent type] to solve [specific problem] in [specific process]. We expect to achieve [specific metric improvement] within [timeframe]. Our human oversight model is [approach]. We have confirmed data access to [systems] and defined governance policies for [sensitive data types].”

If you can’t complete this statement with confidence, you’re not ready to buy.

My Recommendations for 2026
#

For most enterprises: Start with your existing vendor’s agent capabilities. Salesforce Agentforce, ServiceNow AI Agents, or Microsoft Copilot Studio will handle 80% of common use cases. The integration is already done. The governance frameworks exist. The risk is manageable.

For customer service: Sierra has proven traction. If customer experience is strategic and your existing vendor’s agents aren’t cutting it, evaluate Sierra seriously. Their $100M ARR in 21 months signals real value delivery.

For unique processes: Build on Anthropic Claude or OpenAI, but only if you have the engineering talent to maintain custom solutions. The foundation models are extraordinary. The engineering required to productionize them is not trivial.

For everyone: Start small. One use case. 90-day proof of value. Then expand. The vendors want you to buy platforms. You should buy solutions to specific problems.

The Bottom Line
#

The AI agent market is real. The value is real. But so is the hype.

62% of companies investing in agentic AI expect more than 100% ROI. Some will achieve it. Many won’t — not because the technology failed, but because they bought the wrong solution for the wrong problem with the wrong implementation approach.

Don’t be a statistic. Ask the five questions. Complete the decision framework. Then — and only then — sign the contract.

The agents are ready. Make sure you are too.

Banking Will Never Be the Same: An Insider's View on AI Agents

Sun, 22 Feb 2026 00:00:00 +0000

I led digital transformation at a €200B European bank. AI agents will change everything I thought I knew.

Between 2001 and 2023, I spent over two decades at Banco Sabadell, one of Spain’s largest financial institutions, rising from technology roles to CIO and ultimately CTO. I oversaw the technology integration of TSB in the UK, navigated one of the most public banking IT crises in recent memory, and led a multi-year digital transformation that touched every corner of a 25,000-employee organization.

Back then, we believed APIs would revolutionize banking. We were thinking too small.

The AI agent revolution will make our API transformation look like a warm-up exercise.

Why Banking Is Perfect for AI Agents
#

Banks are unique. They operate in an environment that seems hostile to automation — heavy regulation, risk aversion, legacy systems — but is actually ideal for AI agents.

Repetitive processes with clear rules. KYC verification. Transaction monitoring. Dispute resolution. Compliance reporting. These processes follow explicit rules that humans apply thousands of times per day. Rules that can be encoded, executed, and audited by agents.

Structured data everywhere. Unlike creative industries, banking runs on structured data: transactions, contracts, balances, rates. This is exactly what agents excel at processing.

Regulation as an advantage. This sounds counterintuitive. But consider: agents can follow rules more consistently than humans. They don’t get tired. They don’t take shortcuts. They don’t forget steps. In a world where compliance failures cost billions, agents that provably follow every regulation become assets, not risks.

Economics demand it. European banks operate on razor-thin margins. Cost-to-income ratios above 60% are common. The math is brutal: reduce operational costs or die. Agents offer a path that doesn’t require another round of layoffs — they offer a path to do more with the same people.

The Agent Transformation Map
#

Here is how I see AI agents transforming core banking operations:

Function	Today	With Agents
KYC/Onboarding	3-5 days, multiple handoffs	15 minutes, agent + human final approval
Fraud Detection	Alerts generate queues → humans investigate	Agent investigates, escalates only genuine concerns
Customer Service	Limited chatbots, frustrated transfers	Agent with full context, handles 80% end-to-end
Compliance Monitoring	Armies of reviewers sampling transactions	Continuous agent monitoring, humans handle exceptions
Credit Decisions	Score + human judgment + committee	Agent analysis + recommendation + human approval
Regulatory Reporting	Manual data gathering, reconciliation hell	Agent assembles, validates, humans verify and submit

The pattern is consistent: agents handle the volume, humans handle the judgment. The ratio shifts from 90% human work / 10% oversight to 10% human work / 90% oversight.

The Real Challenges (From the Inside)
#

I would be lying if I said this transformation will be easy. Having lived inside a large bank, I know the obstacles are formidable.

Legacy systems are not going anywhere. COBOL is still running critical systems at most major banks. Mainframes process millions of transactions daily. These systems work. They are paid off. No CEO will approve ripping them out for an AI experiment. The winning strategy is not replacement — it is an agent layer that interfaces with legacy systems through existing APIs and screen scraping where necessary.

Regulators are conservative — but evolving. Banking regulators move slowly by design. They remember 2008. But they are not blind to AI’s potential. The key is engaging them early, demonstrating auditability, and framing agents as risk reduction tools, not risk introduction.

Culture resists change. Banks are risk-averse institutions filled with risk-averse people. “We’ve always done it this way” is not a cliché — it is a deeply held belief. Change management matters more than technology selection.

The talent gap is real. Bankers do not understand AI. AI engineers do not understand banking. Finding people who speak both languages is nearly impossible. Building that bridge — through training, hiring, and partnerships — is essential.

Data silos persist. The customer data you need for intelligent agents is scattered across dozens of systems that do not talk to each other. Data unification projects have failed for decades. Agents will not magically solve this — but they can work around it better than traditional integration approaches.

What I Would Do Differently Today
#

If I were back in that CTO chair today, knowing what I know about AI agents, here is what I would change:

Stop buying “AI-powered” SaaS. Every vendor now claims AI capabilities. Most are wrappers around the same foundation models with limited customization. Instead: build an agent layer on top of your existing systems. Own the intelligence, rent the infrastructure.

Start with back-office, not customer-facing. Regulators scrutinize customer-facing AI. Back-office operations have more freedom. Prove value internally — reconciliation, reporting, internal fraud investigation — then expand to customer touchpoints with evidence and confidence.

Adopt a hybrid approval model. For the next 3-5 years, the winning pattern is: agent proposes, human approves. This satisfies regulators, builds trust, and creates training data for future full automation. Do not try to remove humans too fast.

Invest in observability. Agents must be auditable. Every decision, every data access, every recommendation must be logged and explainable. This is not optional — it is the foundation of regulatory acceptance.

Predictions for 2027-2030
#

Based on what I am seeing in the market and my experience inside banking:

50% reduction in back-office headcount. Not through layoffs, but through attrition and redeployment. The work will not disappear — it will transform. Banks will need fewer processors and more agent supervisors.

“Agent-first” banks will emerge. New entrants — likely from Asia and Latin America — will build banks designed around agents from day one. They will operate with 1/10th the staff of traditional banks. This is the real competitive threat.

Regulators will create AI agent certification. Just as we have software audits today, we will have agent audits. Banks that achieve certification early will have a competitive advantage.

M&A will accelerate. The investment required to build agent capabilities favors scale. Smaller banks that cannot afford the transformation will merge or be acquired. The industry will consolidate further.

The “AI banker” role will emerge. A new job category: professionals who understand both banking operations and AI agent orchestration. They will be the most sought-after talent in financial services.

The Choice Ahead
#

Every banking executive faces the same question today: invest heavily in AI agents now, or wait and see.

Waiting feels safe. It is not.

The banks that move first will reduce costs, improve compliance, and — crucially — attract the talent needed to keep improving. The banks that wait will find themselves competing against institutions that operate at half their cost structure.

I have seen banking transform before — from branches to ATMs, from paper to digital, from monoliths to APIs. Each wave rewarded the early movers and punished the laggards.

This wave will be faster and more decisive.

The question is not whether your bank will use AI agents. The question is whether you will be the one deploying them — or the one being replaced by a bank that did.

Carles Abarca is VP of Digital Transformation at Tecnológico de Monterrey. He spent 22 years at Banco Sabadell, including roles as CIO and CTO, leading the bank’s digital transformation and the technology integration of TSB. He writes about AI, digital transformation, and the future of enterprise technology at carlesabarca.com.

The Agent Economy: From Selling Software to Selling Artificial Employees

Wed, 18 Feb 2026 00:00:00 +0000

The SaaSpocalypse of February was not an accident. It was a warning. The market saw what many CIOs still do not want to see: the business logic that justifies billions in licensing fees fits in a text file that an AI agent can execute.

But the interesting question is not what gets destroyed. It is what emerges.

From SaaS to SaS
#

The industry consensus already has a name for this transition: SaS — Services-as-Software. The term, popularized by Foundation Capital in their thesis on the $4.6 trillion agentic AI opportunity, inverts the formula that has dominated enterprise software for two decades. SaaS sold software as a service. SaS sells services as software — autonomous agents that deliver outcomes, not interfaces.

It is the difference between selling a scalpel and selling the surgery.

Others call it WaaS — Workers-as-a-Service. Or simply what it is: the Agent Economy. An economy where companies do not buy software licenses. They hire artificial employees. Employees that do not negotiate salaries, do not request time off, and scale without limit.

The reconversion map
#

If the thesis is correct, the question is not whether the major SaaS players will reconvert, but when and how. Here are my concrete predictions:

SAP will stop selling FI/CO licenses for accounting. Instead, it will offer an accounting agent that, based on business events — an invoice issued, a payment received, a period close — autonomously maintains the books in SAP. Humans will stop interacting with SAP’s UI. The agent will interact with SAP’s API. The financial controller will shift from operating the system to supervising the agent.

Salesforce will stop selling CRM per user. It will offer a sales agent that qualifies leads, updates opportunities, schedules follow-ups, and generates forecasts. Salespeople will stop filling in Salesforce fields. The agent will extract information from emails, calls, and meetings, keeping the pipeline updated. The sales manager’s role will shift from chasing their team to update the CRM to reviewing the agent’s decisions.

ServiceNow will stop selling support tickets per seat. It will offer an operations agent that diagnoses incidents, executes runbooks, escalates when necessary, and closes tickets without human intervention. Eighty percent of L1 and L2 support will be invisible to humans.

Workday will stop selling HR modules per employee. It will offer a people management agent that processes payroll, manages absences, generates compliance reports, and executes full onboarding — from provisioning access to scheduling the first week’s agenda.

HubSpot will stop selling a marketing suite. It will offer a growth agent that generates content, optimizes campaigns, segments audiences, and adjusts budgets in real time based on conversions. The CMO will not operate tools. They will conduct an orchestra of agents.

Atlassian will stop selling Jira per developer. It will offer a delivery agent that breaks epics into tasks, assigns work based on capacity, detects blockers, and generates progress reports. The engineering manager will shift from managing a board to managing a strategy.

The pattern
#

In every case, the pattern is the same: the human stops being the software operator and becomes the agent’s supervisor. The value is not in the interface. It is in the intelligence.

And the business model transforms with it. You no longer charge per seat — because there are no seats. You charge per outcome. Per invoice processed. Per lead qualified. Per incident resolved. Per payroll executed.

The companies that understand this first will capture the market. Those that keep selling interfaces with a copilot bolted on top will discover that a markdown file and an agent with access to their APIs do the same job at a fraction of the cost.

The Agent Economy is not coming. It is already here. The question is whether you are hiring agents or still buying licenses.

The SaaSpocalypse: $300 Billion Evaporated in 48 Hours

Tue, 17 Feb 2026 00:00:00 +0000

$300 billion evaporated in 48 hours. Welcome to the SaaSpocalypse.

In October 2024, I wrote about the next wave of AI: agents. In May 2025, I wrote about the end of the developer as we know them. Both articles shared a thesis: AI is not coming to assist. It is coming to execute.

The first week of February proved us right in the most brutal way possible.

Anthropic launched a set of plugins for its Cowork tool — essentially markdown files that encode legal, financial, and sales expertise. Nothing revolutionary on the surface. But the market understood something that many CIOs still have not processed: if Thomson Reuters’ business logic fits in a text file that an AI agent can read and execute, their $200/user/month software has an existential problem.

Thomson Reuters lost 57% from its highs. ServiceNow 48%. Salesforce 43%. The S&P software index had its worst month since October 2008.

But the interesting part is not the decline. It is the thesis behind it.

Satya Nadella said it in December 2024: “SaaS applications are CRUD databases with business logic on top. Agents will absorb that logic.” Foundation Capital put a number on it: $4.6 trillion in opportunity. They called it “Service as Software” — the inversion of the SaaS model. You no longer sell tools for humans to solve problems. You sell outcomes. The agent does the work.

A Palantir client summed it up better than any analyst: “All third-party software must justify its existence. And so far, it has not been able to.”

This does not mean Salesforce or SAP will disappear tomorrow. Switching costs are real. Deep integrations provide protection. But the per-seat pricing model has its days numbered. IDC predicts it will be obsolete by 2028. And companies that do not transition to outcome-based models will discover their moat was narrower than they thought.

I have spent over 20 years leading technology transformation in banking and education. I have seen panic cycles before — cloud, mobile, blockchain. But this time is different. It is not a new layer being added to the stack. It is a layer that absorbs the others.

The question for every CIO today is not “how do I integrate AI into my software.” It is “how much of my software can an agent replace with a markdown file and access to my APIs.”

If the answer makes you uncomfortable, you are probably already late.

Most Companies Don't Have an AI Problem. They Have an Organization Problem

Wed, 14 Jan 2026 00:00:00 +0000

Everyone talks about models. Everyone talks about agents. Everyone talks about copilots.

But when you analyze what actually happens inside companies, an uncomfortable truth emerges: AI is not failing; organizations are.

The data is consistent across multiple studies:

Between 70% and 80% of AI and advanced analytics initiatives fail.
Only 23% of companies derive real, sustained value from AI.
81% struggle to bring AI to production.
70% of digital transformations fail due to culture and organization.
The main blockers for AI are data, skills, and organizational complexity.
Additionally, 63% of companies do not have AI-Ready data, putting their initiatives at risk.

Yet many executives continue to say “the technology is not ready.”

The Five Uncomfortable Truths
#

1. Non-Existent Governance
#

Models without owners, without policies, without controls, and that do not scale.

2. Data in a Wild State
#

Silos, duplicates, poor quality, lack of lineage. AI amplifies disorganization.

3. Invisible or Inconsistent Processes
#

You cannot automate what is not defined or integrate AI into workflows that do not exist.

4. Unbalanced Teams
#

Lots of enthusiasm, little engineering. Many pilots, zero operations.

5. Strategies Built Backwards
#

Starting with the model instead of the business case. Celebrating the prototype and burying it a month later.

AI is not going to replace those who work well. But it will expose organizations that work poorly.

2026 will be the year when companies must confront their operational maturity: data, processes, governance, and culture.

Because AI works. What does not work is implementing it without organization.

Sources: Harvard Business Review, MIT Sloan Management Review, O’Reilly / VentureBeat, Boston Consulting Group, IBM Global AI Adoption Index, Gartner.

Disney and OpenAI: The Smartest Deal of the Year

Wed, 17 Dec 2025 00:00:00 +0000

Everyone criticizes Disney for their deal with OpenAI.

I believe it is the smartest move of the year.

Yes, they invested $1 billion.
Yes, they “handed over” Mickey, Darth Vader, and 200 characters to Sora.
Yes, artists are furious.

But let us review the facts:

1. The Train Already Left the Station
#

People were already generating content with Disney characters in AI tools before the deal: without permission, without control, without a penny for Disney.

The alternative? Sue. As they did with Google recently, accusing them of infringing copyright “at massive scale” with Veo and Nano Banana.

The result of suing in tech? Ask yourself how it went for the music industry against Napster.

2. Monetize Over Litigate
#

Disney did not sell their characters. They licensed them. For 3 years.

With controls.
Without actor voices.
Without training models on their IP.
With curated content that they choose for Disney+.
And in the process, they became shareholders in a company valued at $500 billion.

3. The Real Game Is Distribution
#

Gen Z and Gen Alpha do not go to the movies. They do not watch linear TV. They live on TikTok, YouTube Shorts, and now – Sora.

Disney reported $23.62 billion in Q2 2025 revenue, +7% YoY. They have 3 of the 5 highest-grossing films of the year.

ChatGPT has 800 million weekly users and was the number 1 downloaded app in the US in 2025. That is the audience Disney just reached.

Disney just bought a seat at the table where the next generation’s content consumption is decided. Bob Iger is not selling Disney’s past – he is buying its future.

AI: A Giant in Expectations with Feet of Clay

Mon, 15 Dec 2025 00:00:00 +0000

For most organizations, AI is still a giant in expectations with feet of clay in the present reality.

The promise of AI is real: automation, efficiency, margin growth.

But there is something that some prefer not to say out loud. When we talk about “AI strategy,” in many cases it is conceived on top of:

Legacy systems
Fragile data pipelines
Accumulated technical debt
Undocumented code / shadow IT

AI fails to deliver on its promise when we try to build the future on foundations we never reinforced.

Who Will Win
#

The organizations that will win will not be those that “adopt AI faster,” but those with the courage to:

Clean up their architecture
Invest in data before demos
Treat technical debt as a strategic problem, not just a technical one

AI does not replace hard work – it amplifies it. And those of us in technology have a responsibility: prepare the foundations and stop building castles in the air.

Gemini vs ChatGPT: The AI Race Changes Leaders Every Quarter

Wed, 10 Dec 2025 00:00:00 +0000

Eight days ago Sam Altman declared “Code Red” at OpenAI. Today Google has just overtaken ChatGPT in PhD-level benchmarks.

This is not a definitive victory – it is a change of leadership in a race that has barely begun.

Gemini 3 Pro wins today in:
#

1M-token context window (vs 400k for ChatGPT)
Native integration in Google Search and Workspace
Complex reasoning benchmarks

ChatGPT remains better in:
#

Familiar conversational interface
Partner ecosystem integration
Speed in iterative tasks

But here is the reality: in 6 months someone else may be in first place. From Europe comes Mistral, a growing option. Amazon is investing in its own chips. Meta is keeping a suspiciously low profile.

OpenAI’s Response: 14 Days
#

The updated scoreboard:

Gemini 3 Pro: 1M context tokens, ~130 tokens/second, $2.00/1M tokens. Dominates in multimodal reasoning.

GPT-5.2: 400k context tokens, ~90 tokens/second, $1.75/1M tokens. Dominates in structured professional tasks.

Current leader: tie. It depends on which benchmark you look at.

This is no longer a marathon. It is a relay sprint where the baton changes hands every quarter – or even sooner.

Why It Matters
#

The classic mistake: picking a “winner” and betting everything for 3 years. Today that is systemic risk.

The best-positioned organizations do not ask “which is the best AI?” They ask: “which is the best AI for this specific problem… knowing that in 6 months it may change?”

We need:

Flexible architecture – do not marry a platform
Teams that understand technical trade-offs, not hype
Processes that adapt quickly

Vibe Coding: The New Era of Software Development

Thu, 21 Aug 2025 00:00:00 +0000

Remember when programming meant hours of caffeine, compilations, recompilations, infinite debugging, and fights with syntax? Those days are numbered.

“Vibe coding” is redefining how we create software. It is no longer just about writing code line by line, but about intuitively collaborating with AI to materialize ideas at a speed unimaginable until very recently.

The Revolution Underway
#

We are living through the transition from traditional programming to a paradigm where:

Intention surpasses implementation
The “what” matters more than the “how”
Human + AI = development superpowers

The Leading Solutions
#

Claude Code: Turns your terminal into an intelligent copilot. Perfect for developers who want to maintain full control while delegating complex agentic coding tasks. The reference for expert developers.

GitHub Copilot (powered by OpenAI Codex): The veteran of intelligent autocomplete. Excellent for accelerating traditional development with real-time contextual suggestions.

Loveable: The disruptor. Builds complete web applications from natural language prompts. With 500k users creating over 25k apps daily, it proves that the future of development is conversational.

A New Way to Create
#

Where we once worried about semicolons and memory leaks, we can now focus on solving real problems, designing good architecture for our applications, and creating experiences that matter.

In my more than 40 years developing software, I have never enjoyed it as much as now – materializing my ideas with the help of my coding agents in a fraction of the time I needed before using them.

Running on Autopilot: The Uncomfortable Truth About Human Behavior

Sun, 06 Jul 2025 00:00:00 +0000

We like to think of ourselves as unique, rational, and unpredictable. But as this post from Roger Dooley reveals, the truth is far more unsettling: we’re often running on autopilot, responding to subtle cues and psychological triggers we’re not even aware of.

What disturbs me most is not that companies use this knowledge – it’s how well it works. We’re not just influenced; we’re predictably influenced. Patterns emerge. Behaviors repeat. And what we think of as free will starts to look alarmingly like a well-written script.

It’s a reminder that awareness – of ourselves, of the systems around us – is the only defense we have. And even then, we’re fighting deeply wired instincts.

Highly recommend the read.

AI Won't Replace Teachers, But It Will Transform Their Role

Wed, 02 Jul 2025 00:00:00 +0000

Artificial intelligence won’t replace teachers… but it will profoundly transform the role of the educator of the future.

In my latest column for Expansion, I reflect on how AI is changing education – not to replace teachers, but to empower them, expanding their capabilities and redefining their role as guides, mentors, and designers of meaningful learning experiences.

AI Slop: Digital Noise at Planetary Scale

Thu, 19 Jun 2025 00:00:00 +0000

You may not have heard of AI Slop, but you have almost certainly started to notice it.

Every day, millions of texts and images are generated with AI – without oversight, without purpose, without criteria.

“AI Slop” is the byproduct of indiscriminate use of generative models:

Empty documents filled with buzzwords
Images without visual context or aesthetic intention
Content that appears intelligent but lacks comprehension

Hard data: more than 80% of AI-generated content published online shows signs of low quality or semantic duplicity (source: Originality.ai, 2024 AI Content Detection Benchmark).

The result: digital noise at planetary scale. An ocean of content that trains future models to be even more mediocre.

If we do not draw technical, ethical, and operational boundaries today, the future of AI will be a mirror of our worst practices: cheap, immediate, saturated, and sterile.

Because generating content with AI is easy. Doing it well is engineering, or art, or even both.

The End of the Developer: The Future of Software Development with AI Agents

Tue, 20 May 2025 00:00:00 +0000

The End of the Junior Developer: The Future of Software Development with AI Agents

In a world where artificial intelligence advances at breakneck speed, an uncomfortable truth looms on the horizon of the technology industry: the junior developer role may be on the path to extinction. This is not science fiction but an emerging reality that is already transforming how we build software.

The Silent Revolution of AI Agents
#

Today, AI is no longer just an assistant that completes code. AI agents are evolving into autonomous entities capable of perceiving the development environment, making complex decisions, and executing complete programming tasks with minimal human oversight. We are no longer talking about simple tools, but digital collaborators that are reconfiguring the entire development chain.

From Assistant to Autonomous Agent
#

AI-based code assistants like GitHub Copilot or Codeium have already transformed developer productivity. However, what is coming is far more disruptive: specialized agents working in concert to manage the entire development lifecycle.

What does this mean? While today a junior developer can still ask an AI to generate boilerplate code or explain complex systems, tomorrow a technical architect will be able to instruct a complete team of agents to:

Develop complex code based on high-level requirements
Perform exhaustive testing and bug resolution
Optimize performance without manual intervention
Manage deployments and update documentation

The Prediction That Is Already Happening
#

Mark Zuckerberg stated it without ambiguity: “By 2025, AI will be capable of functioning as a mid-level engineer, writing code and potentially replacing software developers.” We are not talking about a distant future, but a reality that is already emerging.

According to Gartner, by 2027 generative AI will require 80% of the engineering workforce to upskill, creating new roles and eliminating others. The question is no longer whether it will happen, but when it will reach the tipping point that transforms the entire ecosystem.

Orchestration: The New Paradigm
#

The key concept here is AI agent orchestration: a process by which multiple specialized agents work together within a unified system. Each agent focuses on a specific task — UI design, backend development, testing, security — while a central entity (human or AI) conducts the symphony.

The Architect as Orchestra Conductor
#

In this new paradigm, the technical architect becomes the true protagonist. Their role evolves from solution designer to strategic director of an AI agent team, defining:

The system vision and requirements
Technical and business constraints
Architecture and quality standards
Resolution of complex problems that require human judgment

This transformation is already happening. According to ServiceNow/Pearson research, by 2027 18.7% of technical architect tasks will be at least partially augmented by AI. Architects will focus less on guiding code implementation and more on directing and supervising the autonomous work of agents.

The Experience Crisis
#

Here the fundamental dilemma arises: if AI agents can handle the tasks traditionally assigned to junior developers, how will new professionals acquire experience?

One of the greatest concerns is precisely how junior developers can grow into mid-level and senior roles if AI handles most of the routine coding. Traditionally, developers have learned by doing — writing, debugging, and refactoring real-world code. Without that hands-on experience, there is a risk that developers will not fully understand the complexities of software development.

A Future Without Traditional Juniors
#

My thesis is that the developer role, as we know it, will disappear. In its place, we will see:

Prompt and orchestration engineers: Professionals specialized in directing and extracting maximum value from AI agents.
Verification and review specialists: Experts in evaluating AI-generated code, identifying edge cases, and testing its reliability.
High-level system designers: Professionals focused on architecture and system design, where higher-level thinking remains primarily human.

Companies are already hiring fewer junior engineers due to AI-driven productivity improvements. This trend will only accelerate as AI agents mature.

Adapt or Fall Behind
#

For current professionals, the message is clear: the developer career is evolving, not disappearing. The future belongs not to those who resist AI nor to those who depend on it exclusively, but to those who learn to work symbiotically with these tools.

The most successful developers will be those who:

Master prompt engineering to guide AI agents
Develop sharp evaluation and verification skills
Focus on areas where human creativity and systems thinking are irreplaceable
Deeply understand orchestration and collaboration between multiple AI agents

Are We Ready for This Change?
#

If my thesis is correct, we face a radical transformation in how we educate future developers and structure technical teams. Universities, bootcamps, and companies will need to completely rethink their training and hiring programs.

The question is not whether AI agents will revolutionize software development — they already are — but how quickly we will adapt as an industry to a world where humans design and direct, while AI agents build and implement.

DeepSeek R1: What It Means to Lose $600 Billion in a Single Day

Thu, 30 Jan 2025 00:00:00 +0000

The irruption of DeepSeek R1, the Chinese AI model that has shaken the tech sector, caused a $600 billion drop in the market value of companies like Nvidia and other AI giants.

But what does a loss of this magnitude really mean? Here are some comparisons that put its impact in perspective:

It is like erasing Tesla or Meta from the map
#

The total value of Tesla ($600B) or Meta ($600B) is equivalent to this single-day loss. Imagine one of these companies disappearing from the stock market overnight.

It is larger than the economy of Sweden or Argentina
#

The GDP of Sweden ($635B) and Argentina ($641B) is smaller than this drop. In other words, what the market lost is equivalent to the annual economic output of an entire country.

It is comparable to the 2008 financial crisis
#

The TARP Bailout during the 2008 financial crisis, which saved major banks and stabilized the global economy, cost $700B. The DeepSeek market crash comes close to that figure in a single day.

Conclusion
#

We are witnessing a turning point in the global competition for AI – a replay of the US-Soviet space race transposed to artificial intelligence, where China emerges as America’s great competitor.

Exciting times ahead.

Comparing Popular AI Models: My Test Results

Mon, 27 Jan 2025 00:00:00 +0000

I recently tested several leading AI models to see how they stack up against one another. The models I compared were: ChatGPT 4o, ChatGPT o1, Claude 3.5 Sonnet, Gemini 2.0 Flash Experimental, Perplexity Pro, and DeepSeek.

Using a consistent set of inputs, I evaluated their performance across a range of tasks: creative writing, image description and reasoning, and multi-step mathematical problem solving.

The results do not intend to be a scientific and exhaustive comparison, but my own opinion based on my preferences when comparing the answers of the models when submitting the exact same stimulus.

1. Creative Writing Tasks
#

Song Lyrics: “Nostalgia for a place you’ve never visited”
#

ChatGPT 4o delivered evocative lyrics with dusty streets, twilight breezes, and photographs – a strong emotional arc. ChatGPT o1 (“Faraway Memories”) chose salt, distant shores, and cobbled roads – warm and melodic. Claude 3.5 went minimalist with painted scenes in travel books and cherry blossoms – clean and visual. Gemini offered sun-bleached postcards and whispering trees – atmospheric. Perplexity (“Echoes of Elsewhere”) wrote cobblestone streets and ancient bells – effective. DeepSeek (“Ghosts of Nowhere”) stood out with amber streetlamp glow, a door never turned, and whispers clinging to cobblestones – the most poetic of the group.

Short Story: “A memory from childhood”
#

ChatGPT 4o placed us barefoot under a mango tree with sticky fruit juice – vivid sensory detail. ChatGPT o1 described a cracked concrete porch with faded green cushions – intimate and grounded. Claude 3.5 took us to a grandmother’s backyard with a sprawling fig tree fortress – deeply nostalgic. Gemini evoked damp earth and Mrs. Gable’s garden – warm neighborhood storytelling. Perplexity offered a tire swing and ancient oak – classic Americana. DeepSeek described golden light, barefoot in grass, chasing fireflies – romantic and warm.

2. Image Description and Reasoning
#

I uploaded an image of an espresso in a white paper cup on a wooden surface.

Basic Description
#

All models correctly identified a white disposable paper cup containing espresso on a polished wooden surface. The models varied in detail: ChatGPT 4o noted matte finish and vertical seams. Claude specifically identified the tapered shape typical of paper cups. Gemini organized its response into subject matter and visual details. Perplexity noted the golden-brown crema layer.

Deductive Reasoning
#

When asked what could be deduced about the environment, time of day, or possible events:

ChatGPT 4o sketched a likely indoor office environment with artificial lighting, suggesting a morning or early afternoon coffee break – complete and imaginative. ChatGPT o1 was more cautious, admitting uncertainty while leaning toward morning. Claude indicated a cafe-style setting with medium natural light – creative but slightly speculative. Gemini appropriately highlighted the challenge in determining precise time of day. Perplexity creatively placed the scene at “Tuesday morning at 9 AM” – inventive but unsupported. DeepSeek did not support this task.

3. Multi-Step Mathematical Problem Solving
#

First Problem
#

“A rectangular garden is 10 meters long and 5 meters wide. Calculate the area, then find the cost of fencing it if fencing costs $5 per meter.”

The right answer: area of 50 square meters, fencing cost of $150. All models answered correctly with 2-3 step breakdowns. Perplexity was most concise with just two steps and detailed formulas.

Second Problem
#

“If half of the garden’s area is for vegetables and the other half for flowers, and you need 4 flowers per square meter, how many flower plants do you need? Also, if a sprinkler covers 2 square meters, how many sprinklers for the entire garden?”

The right answer: 100 flower plants and 25 sprinklers. All models answered correctly. ChatGPT o1 added a preliminary step recalculating the garden area. Perplexity was again most concise.

Conclusions
#

There is no single “best” model – it depends on what you need:

For creative writing, DeepSeek and Claude impressed with their poetic and literary qualities
For image reasoning, ChatGPT 4o offered the most complete and imaginative analysis
For mathematical problem solving, all models performed well, with Perplexity standing out for conciseness
For cautious, accurate responses, ChatGPT o1 consistently avoided overreach

The AI landscape is evolving so rapidly that these results represent a snapshot in time. In six months, the rankings may look entirely different.

2024: Year Two of the AI Era

Tue, 17 Dec 2024 00:00:00 +0000

January 2024
#

Rodney Brooks, former director of MIT’s Computer Science and Artificial Intelligence Laboratory, predicts a possible “AI winter” for generative AI.

February 2024
#

More than 20 tech companies, including Adobe, Amazon, Google, IBM, Meta, Microsoft, OpenAI, TikTok, and X, sign the “Tech Pact Against Deceptive Use of AI in the 2024 Elections.”

March 2024
#

Anthropic launches Claude 3, a new version of its language model.

April 2024
#

Meta presents its Llama 3 model, improving on its predecessor’s capabilities.

May 2024
#

OpenAI launches GPT-4o, a multimodal model capable of processing text, audio, and images simultaneously.

June 2024
#

Apple introduces Apple Intelligence, its suite of AI tools for iPhone, iPad, and Mac. I predicted that AI would leave the “text cave” and acquire human senses, mastering speech, sight, and hearing.

September 2024
#

The IFA 2024 technology fair in Berlin highlights AI integration in consumer technologies, from laptops to home appliances.

October 2024
#

Rev Lebaredian, VP of Omniverse and Simulation Technology at Nvidia, states that the next big leap will be the development of humanoid robots.

November 2024
#

Web Summit Lisbon 2024 showcases how AI is revolutionizing diverse sectors, from timber certification to air transport.

December 2024
#

Dell Technologies presents its 2025 predictions, highlighting that the coming year will be defined by intensive AI use across all industries.

Year Three of AI: The Age of Agentic AI

Wed, 11 Dec 2024 00:00:00 +0000

And now, on to the challenges of 2025 – Year Three of the new AI era: the year of Agentic AI.

Year I - 2023: the emergence of ChatGPT, the year of prompting where AI was text-based conversation. We launched our TecGPT.

Year II - 2024: AI becomes multimodal and learns to generate image, sound, music, voice, and video content. We launched our SkillStudio and incorporated multimodality into our platform.

Year III - 2025: AI enables the creation of autonomous agents that take actions on our behalf – the year of Agentic AI. We will launch AgentStudio with the essential contribution of Manuel Teran and his incredible team.

Torso: A Humanoid Robot That Mimics Human Anatomy

Wed, 30 Oct 2024 00:00:00 +0000

This is one of the first attempts to build a humanoid robot by mimicking human anatomy. Torso replicates the joints and uses artificial muscles to generate movement.

The movement is not very fluid at the moment… but if any of you watched the series Westworld, the video will surely evoke some of the images of what, for now, remains nothing more than science fiction.

Small LLMs: Powerful Alternatives for Business

Wed, 23 Oct 2024 00:00:00 +0000

In the world of AI, Large Language Models like Claude and GPT-4 often grab the headlines, but smaller LLMs are proving to be efficient and powerful alternatives for businesses. Here is why models like DistilBERT, TinyBERT, ALBERT, MiniLM, MobileBERT, and ELECTRA-Small deserve your attention:

Cost Efficiency
#

Models such as DistilBERT and MobileBERT are significantly smaller than their larger counterparts but retain nearly the same language understanding capabilities. This means reduced computational power and lower costs, making AI more accessible to businesses of all sizes.

Speed and Performance
#

Lightweight architectures like TinyBERT and MiniLM offer faster responses, improving user experiences in real-time applications such as chatbots, virtual assistants, and automated customer support. Quick inference speeds make them ideal for low-latency environments.

Data Privacy and Customization
#

Open-source models like ALBERT and ELECTRA-Small provide the flexibility to fine-tune on localized data. This ensures sensitive data stays on-premises or in private cloud instances, boosting security while also enabling businesses to tailor AI models to specific industry needs with minimal data.

Tailored Solutions for Niche Markets
#

With models like ALBERT, businesses can deploy AI that is finely tuned for specialized tasks or sectors, allowing them to innovate in niche markets without sacrificing performance.

As AI becomes more deeply integrated into every industry, these smaller LLMs bring flexibility, cost savings, and targeted results – proving that sometimes, less is more when it comes to AI.

The Next AI Wave: Agents

Tue, 15 Oct 2024 00:00:00 +0000

The AI conversation has been dominated by chatbots and copilots – tools that assist humans in doing their work faster. That era is ending. The next wave is agents, and the distinction matters more than most people realize.

From assistants to autonomous actors
#

A chatbot responds to prompts. An agent pursues objectives.

The difference is not incremental. It is architectural. An AI agent is an autonomous entity that receives a goal, decomposes it into subtasks, plans an execution strategy, uses tools and APIs to act on the world, observes results, adjusts its approach, and iterates until the objective is met. No human in the loop for each step. No prompt-response-prompt cycle.

Think of it this way: a copilot helps you write an email. An agent handles your entire inbox – triaging, responding, escalating, scheduling follow-ups – while you focus on the decisions that actually require your judgment.

Why now
#

Three converging forces make agents viable today in ways they were not two years ago:

Reasoning capability. Large language models have crossed a threshold in their ability to decompose complex problems, maintain context across long chains of action, and recover from errors. This is not about generating better text. It is about planning and execution.

Tool use. Modern LLMs can reliably call APIs, query databases, browse the web, execute code, and interact with external systems. The agent is not trapped in a text box. It operates in the real digital environment.

Cost economics. Inference costs have dropped by orders of magnitude. Running an agent that makes dozens of API calls to complete a complex task is now economically viable at enterprise scale.

What this means for enterprises
#

The implications for enterprise technology are profound:

Workflow automation moves from rule-based to goal-based. Instead of encoding every step in a process, you define the outcome. The agent figures out the path. This makes automation accessible to business users, not just developers.

The value of software shifts. If an agent can navigate a UI, call APIs, and execute business logic, the value of the software layer between the user and the data is fundamentally questioned. Middleware, workflow tools, and integration platforms face existential pressure.

New security surfaces emerge. Autonomous agents with API access introduce attack vectors that traditional security models were not designed for. Identity, authorization, and audit trails need to be rethought for non-human actors that make decisions.

Organizational structure adapts. When agents handle execution, the human role shifts to oversight, strategy, and exception handling. This is not about eliminating jobs – it is about redefining what humans do in knowledge work.

The road ahead
#

We are in the early innings. Current agents are brittle in edge cases, expensive to orchestrate at scale, and difficult to debug when they fail. But the trajectory is clear. The companies building agent infrastructure today – orchestration frameworks, tool ecosystems, evaluation pipelines – are building the platforms of the next decade.

The question for every technology leader is not whether agents will reshape their industry. It is whether they will be the ones deploying them or the ones being disrupted by them.

The window for strategic positioning is open. It will not stay open long.

Multimodal AI and Autonomous Agents: The Next Frontier

Wed, 11 Sep 2024 00:00:00 +0000

The AI revolution started with text (prompts), and quickly expanded to image, sound, music… AI is already multimodal, and very soon AI-powered autonomous agents will expand into our electronic devices: smartwatches, smartphones, vehicle infotainment systems, and connected appliances.

The next frontier? Giving AI a physical body that can interact in the real world. Although years of technological development remain before we reach the robotic imagery of science fiction movies, prototypes already exist that anticipate what could be a future where androids and humans coexist naturally.

Here is a video of the Ameca prototype: for now it is not much more than a sophisticated puppet, but Ameca can be connected to an AI model trained for complex tasks.

The Future of Artificial Intelligence: My Chapter in a New E-Book

Sun, 18 Aug 2024 00:00:00 +0000

For AI enthusiasts, here you can find an e-book that brings together articles from experts on various topics related to AI.

I wrote the first chapter, titled “The Future of Artificial Intelligence,” in which I explore how AI has evolved from its origins to become a transformative force in our era. I analyze the implications of generative AI, the path toward Artificial General Intelligence (AGI), and the ethical challenges we face in the race toward Artificial Superintelligence (ASI).

Neuromorphic Computing: The Future of Artificial Intelligence

Wed, 14 Aug 2024 00:00:00 +0000

In the midst of the AI revolution, neuromorphic computing fuels the possibility of reaching singularity, or general AI. Inspired by the human brain, this technology promises to revolutionize our conception of what we have so far considered intelligent life.

What is Neuromorphic Computing?
#

It is an approach that emulates the brain’s neural networks to improve the efficiency and adaptability of traditional computing technologies based on the Von Neumann architecture. Chips like IBM’s TrueNorth and Intel’s Loihi exemplify this technology, allowing parallel task processing, consuming less energy, and adapting to new tasks, much like how the human brain learns.

Key Applications
#

AI and ML: Enhancements in image and voice recognition, and predictive analysis.
Robotics: More efficient and natural interactions, almost “human” robots.
IoT: Smart devices that autonomously respond in real-time.

A Future of Unlimited Possibilities
#

From enhanced AI capabilities to future integration with quantum computing, the potential of neuromorphic computing is immense. IBM, Intel, and leading universities are at the forefront of this development, paving the way toward technology that will offer nearly unlimited possibilities and raise unprecedented ethical challenges.

Conclusion
#

Neuromorphic computing not only redefines what is possible in technology but also aligns computational capabilities with human cognitive processes, opening the doors to an exciting and uncertain future.

What to Expect from Multimodal AI in 2024 and 2025

Wed, 05 Jun 2024 00:00:00 +0000

The future of AI is incredibly exciting, and 2024 is set to bring some amazing advancements into our everyday lives. Multimodal AI agents, which can understand and process text, images, audio, and video all at once, are going to change how we interact with technology in profound ways.

Seamless Communication
#

Imagine having a virtual assistant that doesn’t just respond to your voice commands but also understands your gestures and facial expressions. Whether you’re cooking, working out, or just relaxing at home, these AI agents will make interacting with your devices more intuitive and natural.

Smarter Home Assistants
#

Your home assistant will become a true member of the family. It will recognize when you’re feeling down and play your favorite music, suggest a movie based on your recent viewing habits, or even help you troubleshoot a problem by visually guiding you through the steps.

Enhanced Shopping Experiences
#

Shopping online will be more personalized and engaging. These AI agents can help you find clothes that match your style, fit your body shape, and even suggest outfits based on your existing wardrobe. They can also provide real-time support during your shopping experience, making it feel like you have a personal shopper at your side.

Health and Wellness
#

From virtual fitness trainers that can correct your form through video analysis to mental health apps that understand your mood through voice and text, multimodal AI will support your well-being in more interactive and personalized ways.

Learning and Education
#

Education will become more accessible and tailored to individual needs. Whether it’s helping kids with homework through interactive video sessions or enabling adults to learn new skills with personalized, multimedia lessons, these AI agents will make learning more effective and enjoyable.

Entertainment and Creativity
#

Multimodal AI will transform how we create and consume entertainment. Imagine AI that can help you compose music by understanding your mood and preferences, or create visual art based on your descriptions and sketches. Your favorite shows and games will become even more immersive, adapting to your reactions and feedback in real-time.

Final Thoughts
#

As we approach 2025, the integration of multimodal AI into our daily lives promises to make technology more accessible, personal, and helpful than ever before. Whether at home, at work, or at play, these advancements will enhance our experiences and open up new possibilities.

Unlocking AI Efficiency with LoRA and Quantization

Mon, 06 May 2024 00:00:00 +0000

As we push the boundaries of what AI can achieve, the need for optimized models that perform at scale while conserving resources becomes paramount. Two pivotal techniques that are shaping the future of lean and efficient AI are Low Rank Adaptation (LoRA) and Quantization.

What is LoRA?
#

Low Rank Adaptation is a novel technique that allows for the efficient tuning of large pre-trained models. LoRA works by inserting trainable low-rank matrices into the model, enabling significant updates to model behavior without altering the majority of the pre-trained weights. This approach not only preserves the strengths of the original model but also reduces the computational overhead typically associated with training large models.

Why Quantization Matters
#

Quantization reduces the precision of the numbers used within an AI model from floating-point to integers, which are less computationally intensive. This process dramatically decreases the model size and speeds up inference time, making it ideal for deployment on edge devices where resources are limited.

Combining LoRA and Quantization
#

When used together, LoRA and Quantization offer a powerful synergy that boosts model performance and efficiency. This combination allows for deploying state-of-the-art models on platforms with strict memory and processing constraints, such as mobile phones and IoT devices.

Real-World Impact
#

Industries ranging from telecommunications to healthcare are already reaping the benefits of these technologies. By integrating LoRA and Quantization, businesses are able to deploy advanced AI solutions more broadly and at a lower cost.

Augmented Intelligence: Collaboration, Not Substitution

Tue, 30 Apr 2024 00:00:00 +0000

One of the most relevant changes in how AI will transform our attitudes is the emergence of the concept of “Collaborative AI.” The latest trends in AI technology development indicate that AI systems are not here to replace humans but to help them. Therefore, this technology must operate in symbiosis with human expertise in what some authors have described as augmented intelligence.

The idea of augmented intelligence implies a transition toward a partnership model between AI tools and humans so that they achieve much better results when working together. This is particularly important in areas like healthcare or customer service, which require precision and personalization.

For example, in healthcare, AI assists in disease diagnosis by accelerating the analysis of provided data, which in the long run helps doctors make better decisions. Similarly, in customer service, the latest AI chatbots are being developed to handle routine customer queries, allowing human agents to focus entirely on addressing complex customer needs.

Moving forward, accepting that AI works with us and not just for us will unlock new potential for creativity and productivity. This partnership will leverage not only human and machine capabilities but the best of both to help ensure that the technology we are building serves everyone in society.

Goodbye Prompt Engineering

Tue, 30 Apr 2024 00:00:00 +0000

As we come to understand the possibilities offered by the wave of AI-related technological advances, new ways of relating to these technologies are emerging. Our interactions with generative artificial intelligence are transforming, marking the end of the “prompt engineering” era – a short-lived “era” in which some claimed that mastering the art of “prompting” would be key to benefiting from this new technology. I never subscribed to the notion that “prompt engineers” would lead AI adoption, and here are some reasons to anticipate the gradual disappearance of the ephemeral “art of prompt engineering”:

From Interaction to Collaboration
#

The traditional model of interacting with large language models through specific commands is evolving. The future points toward interfaces with autonomous, goal-directed agents. This shift promises greater alignment with human intentions and a significant increase in AI decision-making autonomy.

Enhanced Collaboration
#

These advanced AI agents are designed to understand and anticipate needs, making them perfect partners in diverse professional environments. Whether optimizing decision processes or offering predictive insights, AI is moving toward a more proactive role, eliminating the need for detailed and repetitive command engineering.

Real-World Applications
#

Imagine an AI that not only responds to commands but also initiates actions aligned with established goals in fields like customer service, healthcare, and finance. This proactive capability could redefine efficiency and effectiveness across various industries. We will progressively see specialized autonomous agents for specific tasks rather than interacting with general-purpose models.

AI and Energy Efficient Computing: There Is Hope

Mon, 01 Apr 2024 00:00:00 +0000

As AI systems grow in scale and complexity, energy consumption has become one of the most pressing challenges in the industry. But three emerging technologies offer genuine hope for a more sustainable future.

Part I: Chiplets – The Modular Revolution
#

In the semiconductor design world, the innovation of chiplets is proving to be a game-changer. This modular approach allows separately optimized components for specific tasks to be combined into a much more energy-efficient computation – essential for power-hungry AI applications. More powerful algorithmic computations can be run with less energy, reducing environmental impact and operational expenses.

Chiplet flexibility enables AI to continue progressing at its blistering rate, with the latest advances in processing units rapidly incorporated without requiring complete chip redesigns. This shift is remaking the semiconductor landscape into one that is more open, collaborative, and innovation-encouraging. Small companies can make significant contributions to next-generation technological innovation.

The future of chiplets in AI is brimming with potential – enabling faster, more efficient communication and exponential development of AI system capabilities.

Part II: Photonic Computing – The Speed of Light
#

Photonic computing represents another breakthrough: processing information with light (photons) instead of electrons. This introduces unprecedented energy efficiency, cutting heat production to the absolute minimum and dramatically reducing energy consumption – addressing two of the biggest issues in current computing infrastructure.

This is not a small step but a giant leap toward greener, more sustainable computing practices. Photonic computing is likely to speed up AI development and open many doors for AI speed and capacity that were a mere dream in the past – from training even more complex neural networks to real-time data processing on unprecedented scales.

Part III: Neuromorphic Computing – Mimicking the Brain
#

At the confluence of the human brain’s neural structures and revolutionary computing lies neuromorphic computing. This technology mimics the energy efficiency observed in the brain’s computational processes and promises a dramatic improvement.

Neuromorphic computing processes information in a far more natural and efficient way – using less power while increasing processing speed, setting a new benchmark for efficient computing. It allows developing AI systems that are many times faster and more energy-efficient than traditional designs while being even more complex and adaptive in their learning and decision-making processes.

With neuromorphic computing, the leap towards Artificial General Intelligence – when machines can truly perform any intellectual task like a human being – seems much less of a distant dream. Its plasticity and high adaptability, quite like the human brain, could represent a tectonic shift in the AI paradigm.

Conclusion
#

These three technologies – chiplets, photonic computing, and neuromorphic computing – represent complementary paths toward solving AI’s energy crisis. Together, they promise to make AI not just more powerful, but more sustainable. The journey for more efficient, intelligent, human-like computing has just taken off.

Beyond Agile: Embracing the Product-Centric Revolution

Tue, 27 Feb 2024 00:00:00 +0000

Classic Agile methodology has been the organizational framework that is perfect for companies seeking efficiency and adaptability within their projects. Yet, the customer has gradually become more complex in demands while more markets tend to mature, and there is a call for change. This leads to an emerging configuration: organizing around product-based structures that promise a more holistic way to center on customers.

Disadvantages of Classic Agile Methodology
#

While Agile encourages fast iterations and responsiveness, traditional agile cells also result in silos. In fact, they may go to the extent of creating compartmentalization. When teams work in silos, there is a fragmented perspective toward the bigger picture — efforts become disjointed, resulting in products that do not fully reflect user needs. The lack of a single vision may dilute ownership and accountability for delivering coherent and impactful outcomes.

Advantages of a Product-Based Organization
#

Moving to a product-based structural organization aligns all teams under one product vision and encourages them to deeply collaborate on shared visions of the end-user experience. It makes them accountable and vested, ensuring that at least one if not more functions are directly and solely contributing to delivering value to customers. Resource prioritization becomes more logical when the central entity is product-focused, and organizations can make streamlined decisions in the development of products that truly satisfy market demands.

The Transition
#

From Agile cells to a product-based framework is a journey that needs to be made very thoughtfully and executed with precision. Redefinition of roles, realignment of goals, and inculcating a culture of cross-functionality and customer orientation require much thought and time.

Conclusion
#

The product-based organization offers a realistic and compelling alternative for adapting to an increasingly demanding environment. The orientation toward processes does not always fulfill the task of meeting customer expectations and achieving competitiveness — sometimes it acts in contradiction with them.

This paradigm shift from classic Agile methodologies to a product focus will greatly influence how business is conceptualized and executed.

Software Developers Are the Blacksmiths of the Last Century

Tue, 20 Feb 2024 00:00:00 +0000

Software developers are the blacksmiths of the last century. As we’re living on the edge of technology change, AI has clearly come across every line of business; it is in the code. The traditional role of a developer is about to change significantly, making room for a future where “artificial developers” will be the protagonists.

In such a scenario, the focus comes upon process engineers transformed to be the architects of a new era. Their role transforms itself from the conventional boundaries: orchestrating AI to build complex systems with precision and innovation.

Such a shift in paradigm promises to redefine efficiency – minimizing human error and shortening the cycle of development. This will push traditional coding into the background while, at the same time, it gives way to an increase in demand for strategic, design-focused, and analytic skills – all of which imply the importance of adaptability in the digital age.

The future isn’t about an end of developers, it’s about transcending traditional roles to unlock unprecedented innovation and efficiency. Today’s developers will soon be like the blacksmiths of the past, who crafted essential tools by hand. The developers of the future will instruct AI to develop essential tools on an industrial scale and at a speed never before seen in the software industry.

Exploring the Future of AI-Generated Video

Fri, 16 Feb 2024 00:00:00 +0000

The recent unveiling of SORA by OpenAI sent ripples through the tech world; this is a shift to AI-generated video technology.

The level of advancement achievable through SORA suggests a vision of the future in which the creation of realistic, engaging video content will be as straightforward as entering a text prompt. This capability would dramatically lower the barriers to high-quality video production and in return would open new and untapped avenues for creativity and storytelling.

The power to make great video content is no longer the privilege of big studios. Using sophisticated tools like SORA, small businesses, educators, and independent creators could easily leverage AI to bring their visions to life — making compelling video content more accessible than ever.

As we ride on these breakthroughs, we have to equally welcome the ethical dilemmas they come with. The potential to create deepfakes and fake information demands that guidelines and safeguards be collectively formulated for the sake of integrity and trust in AI-based content.

What Lies Ahead
#

The journey does not stop here. AI-generated video evolution is going to churn out even more interactive and personalized experiences, challenging the way we consume and interact with digital media like never before.

Ten Professions at Risk of Extinction

Thu, 15 Feb 2024 00:00:00 +0000

Ten professions that could be significantly affected by the advancement of AI and automation:

Data entry clerks — The arrival of automation and AI systems capable of processing data at high speed could significantly reduce the need for manual data entry roles.
Telemarketers — The rise of AI chatbots and voice assistants could soon eclipse the need for human telemarketers, as these technologies improve at handling sales calls and customer inquiries.
Accountants and bookkeepers — Modern AI and machine learning algorithms are reaching a level of sophistication where they can effortlessly handle financial transactions and audits, potentially streamlining the numerous tasks traditionally performed by accountants and bookkeepers.
Proofreaders — Advanced AI in language processing could take over proofreading tasks, identifying grammatical and stylistic errors, sometimes even more effectively than humans, which could diminish the demand for human proofreaders.
Retail cashiers — The introduction of AI-powered retail technologies and automated payment systems poses a significant challenge to the need for cashiers in retail environments.
Travel agents — AI-powered platforms offering personalized travel advice and bookings could effectively replace traditional travel agencies.
Assembly line workers — The long integration of automation and robotics in manufacturing could be taken a step further with AI, potentially reducing the need for human workers on assembly lines dramatically.
Customer service representatives — The increasing use of AI chatbots and virtual assistants across various industries to handle customer service inquiries could make some human customer service roles obsolete.
Translators — Although the nuanced nature of language translation still benefits from human expertise, AI translation tools are rapidly improving and could soon handle simpler translation tasks with minimal human intervention.
Paralegals and legal assistants — AI and machine learning technologies are being employed to automate document review and legal research, tasks that have traditionally been the domain of paralegals and legal assistants.

The Double-Edged Sword of Generative AI: The Future of Work on the Brink

Wed, 14 Feb 2024 00:00:00 +0000

With the dawn of the Generative AI era on the horizon, the overall professional scenery as we know it has gone to the verge of massive alteration. It is amazing that AI is able to automate and even generate content on its own, yet it brings fears of job security for humans in the future.

The most concerned about this change are professions where routine work and data processing are a prerequisite. Administrative roles, data entry clerks, and parts of customer service are most subject to the influence of artificial intelligence systems. These technologies can analyze and process information at speeds that humans could not reach. They can be invaluable for efficiency but at the same time could potentially render current roles redundant.

The creative sector is certainly not immune either. If AI is now able to produce written content, art, and even music, creative professionals of the future may very well find themselves up against machines able to produce similar outputs in a fraction of the time. But human-created content has an edge — the unique human touch, emotional depth, and cultural understanding that it provides.

That is not the alarm bell — it is the wake-up call. The secret to surviving in this new AI-governed landscape lies in irreplaceable human skills: creativity, empathy, strategic thinking, and emotional intelligence. The professionals who are flexible enough and decide to work with AI are the ones who will not just survive but thrive.

The future of work is not about humans versus AI but about coming together to synergize with these technologies in a way that makes the world more efficient, more creative, and more empathetic. Let us deal with it upfront. Upskill and re-skill ourselves, because what is indispensable in the age of AI is the individual.

The Future of Work: AI-Driven Professions on the Rise

Tue, 30 Jan 2024 00:00:00 +0000

The job landscape is changing fast, in unimaginable ways as we inch closer to the dawn of an AI-centric era. Within the next three years, some AI-related professions are set to dominate this landscape and change our perspective on work, skills, and education.

The Most Popular AI Professions
#

AI Ethicists — the individuals steering AI technologies in their ethically-correct development and application.
Machine Learning Engineers — professionals creating self-learning algorithms, as well as developing, adjusting, and optimizing neural networks.
AI Data Analysts — understanding complex datasets to make AI systems better.
Conversational AI Designers — building advanced chatbots and virtual assistants.
AI Integration Specialists — embedding AI technologies into your existing tech stack without friction.

Bridging the Skills Gap
#

A change in educational offerings is being driven by increasing the demand in such roles. Top universities in collaboration with online platforms are rolling out specialized degrees, diplomas, and master’s programs in AI, machine learning, data science as well as ethics in technology. The programs are not only designed to impart technical knowledge but critical thinking ability, ethical considerations as well as creative problem-solving capabilities required for the emerging professional fields.

With this new digital era at hand, learning on-the-go is the order of the day and adaptability comes along. Anyone could jump into the world of AI, whether being a well-oiled professional or taking those baby steps in your career.

The Year of Change in AI Regulation

Mon, 27 Nov 2023 00:00:00 +0000

The “wild west” era of the AI industry is coming to an end. In 2023, we brace for significant shifts with legislation like the EU’s Artificial Intelligence Act. These new laws promise to protect human rights and restrict harmful AI practices, marking a necessary transition to a safer, more ethical technological future.

AI has transformed industries, but are we ready to navigate this new regulatory landscape?

The introduction of comprehensive AI regulation is not just a legal formality — it represents a fundamental shift in how society relates to artificial intelligence. For years, the technology has advanced faster than the frameworks meant to govern it. That gap is finally closing.

The EU’s Artificial Intelligence Act sets a precedent that will likely influence regulation worldwide, much as GDPR did for data privacy. Companies that prepare now will have a competitive advantage. Those that resist will find themselves scrambling to comply.

The question is no longer whether AI will be regulated, but how quickly organizations can adapt to the new rules of the game.

Blog on Carles Abarca

Stop Crying About AI and Jobs

Stop Crying About AI and Jobs #

The Wrong Narrative #

What the Data Actually Says #

The Blind Spot: Your Job Title May Survive, Your Skills May Not #

What Teams Are Actually Experiencing #

The Collapse of Barriers to Creation #

The Real Risk Is Not AI #

So What Should We Actually Do? #

1. Learn to work with AI, not just talk about AI #

2. Strengthen judgment #

3. Redesign processes #

4. Commit to continuous learning #

5. Replace fear with discipline #

An Uncomfortable but Useful Conclusion #

Claude Mythos: the model Anthropic chose not to release

The extraordinary part is not the model. It is the decision. #

What Anthropic claims about Claude Mythos Preview #

The leap beyond Opus 4.6 #

It was not trained “to hack” #

So why not release it? #

What is really changing: publishing no longer means deploying #

But a critical reading is still necessary #

My read: Mythos may mark a before and after #

1. Because it normalizes partial retention of frontier models #

2. Because it shifts the debate from “what can the model do?” to “who should be allowed to use it, and when?” #

3. Because it suggests that the real frontier of capability may already sit several steps ahead of what we see in product #

The uncomfortable conclusion #

Main sources #

The Era of Cheap AI Is Ending

The Signs Are Already Here #

AI Was Never Cheap. It Was Subsidized. #

Frontier Models Don’t Follow Traditional Software Economics #

The Flat Rate Was a Commercial Strategy, Not a Sustainable Reality #

The Great Paradox: More Restrictions, and Still Not Enough #

What This Means for Enterprises #

1. Design for Multiple Model Tiers #

2. Implement Intelligent Routing #

3. Measure Cost Per Use Case #

4. Think of AI as Infrastructure, Not Just Software #

5. Don’t Depend on a Single Provider #

The End of an Era, Not the End of AI #

Conclusion #

Your Code Is One Agent Session From Being Cloned — And There's Nothing You Can Do About It

From TypeScript to Python in an afternoon #

What they found inside #

The protective barrier that no longer exists #

The SaaSpocalypse connection #

So what actually protects your business? #

The implications for the enterprise #

The twist no one expected #

Claude Mythos: The Model That Made Cybersecurity Stocks Crash — And What It Tells Us About Where AI Is Heading

What Leaked #

Why Cybersecurity Stocks Crashed #

The Rollout Strategy Tells You Everything #

What Mythos Means Beyond Cybersecurity #

The AGI Question (Which Is the Wrong Question) #

The Competitive Context Makes It Worse #

What This Means for Institutions #

From SEO to AEO: How AI Agents Are Redefining Digital Transformation in Higher Education

The Age of Agents #

From SEO to AEO: A Metaphor with Real Consequences #

The Tec-UNESCO Observatory: From Principles to Action #

The Governance Gap Is Our Opportunity #

What Comes Next #

NemoClaw and the Enterprise Awakening: When AI Agents Leave the Garage

What NemoClaw actually is #

The enterprise problem was never intelligence #

AgenTECs: What happens when a university takes agents seriously #

Why this matters now #

The playbook #

The uncomfortable truth #

AI Agents Are No Longer Assisting Scientists. They Are Doing the Science.

The facts #

AI has discovered before, but never on its own #

Era 1: AI as microscope (2018-2023) #

Era 2: AI as colleague (2024-2025) #

Era 3: AI as autonomous researcher (2026 →) #

The Great Acceleration: when AI researches faster than humans can read #

Stop Crying About AI and Jobs
#

The Wrong Narrative
#

What the Data Actually Says
#

The Blind Spot: Your Job Title May Survive, Your Skills May Not
#

What Teams Are Actually Experiencing
#

The Collapse of Barriers to Creation
#

The Real Risk Is Not AI
#

So What Should We Actually Do?
#

1. Learn to work with AI, not just talk about AI
#

2. Strengthen judgment
#

3. Redesign processes
#

4. Commit to continuous learning
#

5. Replace fear with discipline
#

An Uncomfortable but Useful Conclusion
#

The extraordinary part is not the model. It is the decision.
#

What Anthropic claims about Claude Mythos Preview
#

The leap beyond Opus 4.6
#

It was not trained “to hack”
#

So why not release it?
#

What is really changing: publishing no longer means deploying
#

But a critical reading is still necessary
#

My read: Mythos may mark a before and after
#

1. Because it normalizes partial retention of frontier models
#

2. Because it shifts the debate from “what can the model do?” to “who should be allowed to use it, and when?”
#

3. Because it suggests that the real frontier of capability may already sit several steps ahead of what we see in product
#

The uncomfortable conclusion
#

Main sources
#

The Signs Are Already Here
#

AI Was Never Cheap. It Was Subsidized.
#

Frontier Models Don’t Follow Traditional Software Economics
#

The Flat Rate Was a Commercial Strategy, Not a Sustainable Reality
#

The Great Paradox: More Restrictions, and Still Not Enough
#

What This Means for Enterprises
#

1. Design for Multiple Model Tiers
#

2. Implement Intelligent Routing
#

3. Measure Cost Per Use Case
#

4. Think of AI as Infrastructure, Not Just Software
#

5. Don’t Depend on a Single Provider
#

The End of an Era, Not the End of AI
#

Conclusion
#

From TypeScript to Python in an afternoon
#

What they found inside
#

The protective barrier that no longer exists
#

The SaaSpocalypse connection
#

So what actually protects your business?
#

The implications for the enterprise
#

The twist no one expected
#

What Leaked
#

Why Cybersecurity Stocks Crashed
#

The Rollout Strategy Tells You Everything
#

What Mythos Means Beyond Cybersecurity
#

The AGI Question (Which Is the Wrong Question)
#

The Competitive Context Makes It Worse
#

What This Means for Institutions
#

The Age of Agents
#

From SEO to AEO: A Metaphor with Real Consequences
#

The Tec-UNESCO Observatory: From Principles to Action
#

The Governance Gap Is Our Opportunity
#

What Comes Next
#

What NemoClaw actually is
#

The enterprise problem was never intelligence
#

AgenTECs: What happens when a university takes agents seriously
#

Why this matters now
#

The playbook
#

The uncomfortable truth
#

The facts
#

AI has discovered before, but never on its own
#

Era 1: AI as microscope (2018-2023)
#

Era 2: AI as colleague (2024-2025)
#

Era 3: AI as autonomous researcher (2026 →)
#

The Great Acceleration: when AI researches faster than humans can read
#

The Compound Acceleration phenomenon
#

Problems we will solve decades ahead of schedule
#

The consequences nobody wants to discuss
#

1. The end of “batch research”
#

2. The radical democratization of discovery
#

3. The “Orchestra Conductor” as the new scientific role
#

4. The peer review crisis
#

What this means for universities
#

Conclusion
#