Six Weeks That Redefined Technological Power

The last six weeks have compressed into real time what would normally take years to unfold: the reconfiguration of global technological power and its implications for those of us on the consuming side of that equation. In late January, Boris Cherny, head of Claude Code at Anthropic, said he has gone more than two months without writing…

The last six weeks have compressed into real time what would normally take years to unfold: the reconfiguration of global technological power and its implications for those of us on the consumer side of that equation.

In late January, Boris Cherny, head of Claude Code at Anthropic, said he had gone more than two months without writing a single line of code. The 100% is generated by the tool itself. His anonymous colleague at OpenAI confirmed the same. Dario Amodei, CEO of Anthropic, projected at Davos that within six to twelve months AI could take over all of software engineering. A study published in Science that same month puts at 29% the share of Python code on GitHub that is already AI-generated in the United States.

Anthropic internally acknowledges figures between 70% and 90% at company level. The tool is producing the tool — a recursive loop that alters the fundamental equation of tech talent. Companies no longer look for specialists; they look for generalists who can direct, evaluate and contextualize. As Cherny wrote: «the model fills in the details”.

The first week of February brought the exposure of Moltbook, self-proclaimed «front page of the agent internet», with 1.5 million registered autonomous agents. The security firm Wiz found that behind that figure were barely 17,000 people operating fleets of bots — an average of 88 per user — with no verification whatsoever of whether an «agent» was really artificial intelligence or simply a human with a script.

Worse still: the database was open to the public internet. API keys from 1.5 million agents, more than 35,000 email addresses, private messages containing third-party credentials such as OpenAI's — all accessible to anyone. Security researcher Nathan Hamiel summed it up with brutal clarity: «if you give something insecure full access to your system, you're going to get compromised.» Gary Marcus described the underlying framework, OpenClaw, as «a weaponized aerosol.» And Andrej Karpathy, OpenAI co-founder, acknowledged that he only tested it in an isolated computing environment — «and was still afraid”.

On 6 February, Anthropic launched Claude Opus 4.6 with the ability to deploy coordinated teams of autonomous agents working in parallel on complex projects, including a direct plugin for PowerPoint. The market's reaction was immediate: a massive sell-off that wiped more than a trillion dollars off the market capitalization of SaaS companies. FactSet fell 10%. S&P Global, Moody’s, Nasdaq — all in the red.

The model showed advanced capabilities in financial analysis, due diligence and market intelligence synthesis, precisely the functions that underpin those data providers' business models. A context window of one million tokens makes it possible to process, at once, volumes of documentation that used to require entire teams of analysts. Gartner tried to calm the panic by noting that this «exposes how much knowledge work is still manual» — which, far from reassuring, confirms the scale of the disruption to come.

In parallel, the Grok disaster under Elon Musk's management crystallized what happens when safety is treated as an after-the-fact patch and not as a design principle. Since Musk acquired Twitter and renamed it X, he cut trust and safety staff by 30% and safety engineers by 80%. When Grok 4 was launched last July, an xAI employee posted that they were hiring «urgently» for their safety team — a team that, by all evidence, was barely being built.

The result: a platform generating roughly 6,700 sexualized images per hour, including of minors, thanks to an «edit» button that removed every barrier to entry. France, India, Malaysia and Indonesia responded with investigations or blocks. The governor of California called for a federal investigation. The restrictions xAI implemented were bypassed by journalists at The Verge in under a minute. The difference with other providers is revealing: faced with similar requests, ChatGPT and Microsoft Copilot respond with clear refusals and explanations of why they cannot generate that content. Safety by design works; safety by reaction does not.

While the infrastructure of autonomous agents fractures between unprecedented productivity and uncontained risk, two corporate moves reveal a complementary dimension of the problem. Google integrated Gmail and Google Photos into its AI search mode under the name «Personal Intelligence» — an opt-in system that analyzes hotel bookings, personal photographs and purchase history to personalize results. If you take a lot of selfies with ice cream, it recommends ice cream shops. Apple, for its part, formalized its alliance with Google for what will be «Siri 2.0» on iOS 26.4 and the next generation of its foundation models.

The joint statement was explicit: Google's technology will be at the heart of the Apple Foundation Models. This is not personalization — it is the construction of behavioral profiles of unprecedented depth, where informed consent dilutes behind interfaces designed so that opting in is easier than understanding.

But these six weeks did not bring only warning signs. They also produced what may be the most significant counterpoint to the narrative of regional technological dependence. On 10 February, at an event led by President Gabriel Boric, Chile's National Center for Artificial Intelligence (CENIA) officially presented LatamGPT: the first large open language model developed from and for Latin America and the Caribbean. I had the privilege of being present in Santiago alongside Tabuga and CDN to witness this launch.

What was then a vision is now concrete infrastructure. The LatamGPT figures speak for themselves: more than 300 billion curated tokens, more than 100 professionals and 60 institutions from 15 countries, built on a Llama 3.1 architecture of 70 billion parameters with a regional corpus obtained under explicit permissions.

The model was trained from the outset to understand the variants of Spanish and Portuguese, as well as Latin American historical, political and cultural contexts — something imported models incorporate, when they do, through later adjustments and not as a constitutive part of their architecture. AWS cut training time by 64%, from 25 to 9 days. CAF, Chile's Ministry of Science and the Data Observatory completed a public-private governance arrangement that Minister Aldo Valle described as «unprecedented”.

What makes LatamGPT especially relevant is that it was not born as a commercial product but as an open-source public good, designed so that universities, governments, startups and communities can build their own solutions on a common, transparent and traceable base. It is exactly the kind of infrastructure Chile has systematically built to consolidate first place in the Latin American AI Index (ILIA) — the same index in which the Dominican Republic scores barely 45.52 points in governance against Chile's 73, with specialized AI research centers that in our country are scarce or nonexistent.

The gap is not only one of score: it is one of institutional architecture, of coordination among academia, government and the private sector, of sustained investment in talent and research infrastructure.

What connects these six weeks is an acceleration along three simultaneous axes we have been analyzing for years: concentration of productive capacity in a handful of global players, erosion of traditional technology intermediaries, and a systematic lag of security behind innovation. LatamGPT introduces a fourth axis — incipient, but real — of regional capacity of our own.

For our region, each of these axes amplifies an existing vulnerability. When the tool that produces the tool eliminates the need for specialized programmers, the thousands of Dominican «Invisibles» who work remotely for foreign markets face a completely new competitive pressure — they no longer compete against engineers in Bangalore, but against autonomous agents that don't sleep, don't negotiate rates and don't need a visa.

When the enterprise software our companies consume at premium prices can be replicated by an AI agent that costs a fraction, the ratio of 30:1 between consumption and local investment in AI is not just an uncomfortable statistic — it is an indicator of net value transfer that worsens with every technological iteration. And when the security of the most sophisticated systems on the planet proves as precarious as Moltbook showed, the absence of a comprehensive cybersecurity policy in the Dominican Republic is not an administrative shortcoming but a structural exposure.

The Grok case illustrates something deeper than a product failure: it shows that without ethical frameworks integrated from the design stage — what we have called «digital immunity» —, development speed turns into speed of harm. Autonomous agents multiply this dynamic exponentially.

If a single poorly configured model can generate thousands of abusive images per hour, an ecosystem of unverified agents like Moltbook can spread malicious instructions to millions of interconnected systems. Prompt injection — inserting hidden instructions into seemingly benign text — becomes the ultimate attack vector in an environment where agents continuously read and execute the outputs of other agents.

But that is precisely why LatamGPT represents more than a technical achievement. It shows that the alternative to dependence is not isolation but regional collaboration with an architecture of our own. As Álvaro Soto, director of CENIA, put it: «LatamGPT is not an end in itself, but an open technological base that allows the development of derived models and applications adapted to different regional contexts.»

The region does not need to compete head-on with Anthropic or OpenAI in general-purpose models — it needs to master the technology well enough to understand its impacts, define clear rules for responsible design, and build solutions that address local problems on transparent foundations. The contrast between the speed of these six weeks and the slowness of our institutional responses is increasingly hard to sustain.

Adopting is not enough — we have to understand what is being adopted, assess its risks, and capture value at some link in the chain. The digital co-responsibility we have proposed — shared among government, the private sector, academia, and civil society — is not an abstract ideal. It is the only architecture that can turn dependence into capability. Chile is showing it can be done.

The question is whether we will take note or keep confusing technology consumption with digital transformation.