acquires

A surprising idea is circulating among engineers who are dedicated to artificial intelligence (IA), and that sounds almost like a confession. It was raised by Dario Amodei, the CEO of Anthropic, in a essay published in April 2025 in which he acknowledged that not even his team fully understands how your AI models work. It is not an occurrence, although it may seem pure marketing. Deep down he is the CEO of one of the AI ​​companies most valuable in the world admitting in writing that his team doesn’t have the inner workings of their own product down pat. Most tellingly, Amodei is not alone. Geoffrey Hinton, award Nobel Prize in Physics in 2024 and one of the figures who has made deep learning possible as we know it, He said essentially the same thing. In other words: he only provided the learning algorithm; What happens next inside the machine is understood by no one, not even its creator. And Chris Olah, co-founder of Anthropic and head of its interpretability team, has been describing neural networks like a scaffolding on which the circuits grow on their own, without anyone having designed them directly. It is important that we keep in mind that we are dealing with three radically different profiles: a CEO with obvious commercial interests, an academic retired from industry, and a researcher who literally lives inside this problem. Even so, all three converge on the same idea, which also has a technical name: emergent capabilities, which are nothing more than abilities that the model did not have on a smaller scale and that appear suddenly when the size increases, without anyone having programmed them or seeing them coming. Scientific warning or sales pitch? This suspicion is reasonable and we are interested in stating it bluntly. Anthropic has built much of its brand identity around security and interpretability, versus competition, such as OpenAI or Google DeepMind, that prioritize rapid deployment. That its CEO publishes an alarmist essay about how little we understand about AI, when this company has been selling us for years the promise of being very careful with its models, fits perfectly with a commercial story: the more mysterious and powerful the creature seems, the more valuable the company that claims to be taming it is. Geoffrey Hinton left Google in 2023 precisely so he could speak without the ties of a corporate salary However, this reading begins to break down as soon as it is applied to Hinton. This scientist left Google in 2023 precisely to be able to speak without the ties of a corporate salary, he does not earn money from any generative AI company and his public discourse is systematically more pessimistic than any executive in the sector. You don’t have any products you need to sell or any funding rounds you need to justify. If this were only marketingit would be a marketing that does not benefit anyone specifically, least of all him. On the other hand, we should not overlook the work of Chris Olah, whose contribution does not consist of making statements: it consists of publishing scientific articles. The induction loops, the learning leaps that suddenly appear during training, the millions of monosemantic features extracted from Claude 3 Sonnet: This is all technical evidence accumulated over years, not a loose phrase in an interview. When Olah describes the inside of a model as arrays of billions of numbers from which we don’t fully understand how cognitive tasks emerge, he’s not selling anything. He is describing, with obvious frustration, the real state of a field that he co-founded. For this reason, the most plausible explanation is not the conspiratorial one, but the more uncomfortable of the two: it is probably true. The people who are building the most powerful AI models in the world appear to have no precise map of how and why those systems acquire capabilities for which no one explicitly designed them. This simply means that the convergence of an executive with incentives to exaggerate the alarm, an academic without any commercial incentives, and a researcher who has spent a decade documenting it empirically points less to an orchestrated campaign and more to an uncomfortable consensus within the field itself. The question that remains in the air, then, is not whether they are selling us smoke. It’s what happens when we continue to scale systems whose interior we still cannot fully explain, while the entire industry admits, almost in unison, that run faster than you understand. Image | TechCrunch | Arthur Petron (processed with ChatGPT) In Xataka | The two invisible settings that decide whether your AI is right or wrong In Xataka | “We are already at the AI ​​singularity”: Sam Altman, CEO of OpenAI