“If you block an AI because you think it is dangerous, you are in medieval obscurantism”

The phrase “this AI model is so dangerous that it would be a mistake to make it public” is something more old than the first public versions of ChatGPT. This narrative of “dangerous” AI is more marketing than reality, but it is still the discourse that companies like Anthropic and OpenAI push every time they launch a new model.

Because it is still a tool (that is drinking up the world’s resources, but a tool) and Yann LeCun has a message for those companies.

If they really believe it is dangerous, they live in medieval obscurantism.

Fear narrative = nonsense

On April 10, Anthropic released Claude Mythos Preview. The company claimed that it was the best AI model ever created and it was something only they could say because few, very few had access to said model. Analysis and tests from companies like Mozilla indicated that yes, it was really amazing in environments like cybersecurity. Not revolutionary, but a very capable tool.

And of course, if it fell into the wrong hands, that tool could be used to exploit systems and find vulnerabilities (something that those companies that had access exploited). It was “too powerful,” according to Anthropic, a narrative of fear that the United States Government itself was responsible for enhancing by prohibiting non-U.S. citizens, including Anthropic workers, from they could use the commercial model: Fable 5.

According to the godfather of AI, all of this is nonsense.

Yann LeCun He is one of the most famous names on the artificial intelligence scene. It was in 2013 when the former Facebook put this academic from deep learning and machine learning at the head of the company’s AI ship. In the 1980s, LeCun developed an algorithm to train artificial neural networks, and since then he has built a career in both the educational and private sectors.

At Vivatech, LeCun participated in a discussion with Steven Levy, a Wired journalist, and they discussed several issues related to the present and future of AI. On stage, LeCun made it clear what his position is on the current language models (the LLMs) and what he thinks about this narrative of fear.

Levy and Yann were talking about the Chinese model more focused on open source when the ‘guru’ commented that “many people, particularly those close to the United States government, believe that AI cannot be open source because that scenario is already occupied by Chinese companies. They are good and cheap models, some free to use, and in the middle of that we have companies like Anthropic.”

LeCun Hasn’t had any problems in the past when referring to rivals in this field and, in this case, argued that “Anthropic, and a few others, are a lobby that, essentially, wants AI not to be open source because they believe it is an inherently dangerous technology.” Then, Yann began to list changes that have really had a weight in the history of humanity, such as the advances that gave way to new ages. But… AI?

“So, What is AI today? At the very least, it is a way to disseminate knowledge. At the moment, it is not a way to generate new knowledge and the only way they work is through repositories of human knowledge and they allow people to have easy access to those repositories. So, it’s basically that, one more way to make people smarter because they can access human knowledge.”

For LeCun, these LLMs are a vitaminized Wikipedia, but if he has not been shy in the past talking about rivals and players in the field of AI, he did not do so at the Paris fair either. “If you are going to block a tool thinking that it is very dangerous, you are in medieval obscurantism,” said the academic. Levy added, pointing out that basically “that thing of discovering a tool and then taking it out of our hands It is more typical behavior of the Catholic Church calling Galileo a “heretic” for his theories“

Because, for LeCun, it all comes down to a different word than “danger”: “control.” “This is how they control the message and what people can and can’t do with their tools. If you buy a pen, you don’t want the company that sells it to you telling you what you can write with it. There’s a limit to what companies can tell you about using their tools.”

In the end, LeCun here is not taking into account that these companies with proprietary models (anyone with a proprietary service, really) can dictate what you do with the tools you license and when they can turn off the tap. We are seeing it even with video games that companies remove your SSD when they consider that you should no longer have it, but that is another topic.

Yann finished his argument by pointing out that sharing knowledge is culture, that this is good and that, although there is a negative counterpart (sharing culture can also lead to radicalization), what we should not do is control what people can do with AI tools. “There is a great arrogance and superiority complex in the idea that only a few are able to control AI and rub it on the masses who do not have access.”

If LLMs are not the answer, what is?

In his speech, LeCun made it clear that, for him, LLMs are not the grail (as several of the CEOs of AI companies who passed through the Vivatech panels did sell). These language models are a Wikipedia on steroids and a vitaminized predictive “keyboard.” There are two things that Yann acknowledges that LLMs do very well: mathematics and code, but the revolution will come with real agents and physical AI.

For LeCun, what we are not getting with LLMs They are two things. “The first is that, if you want to build an agentic system like the ones everyone is talking about, you can’t make it reliable without it having the ability to anticipate the outcome and consequences of its actions. Most of us have the ability. Maybe some politicians don’t.”

“But we are certainly capable of anticipating the outcome of our actions to plan a sequence of actions to accomplish a series of tasks. That is what a world model is.” LeCun comments that we must work on models that predict the future state of the world with the formula t + 1 in which “1” can be 10 milliseconds, one second, one minute or ten years. “That’s a model of the world.”

Therein lies the complication, since companies want to achieve this through language models that, as he has already commented, are nothing more than predictive keyboards that have become more advanced and “can never predict exactly what word will follow a sequence of words.” AND, if they can’t do that, they can’t predict real world data.

Take video creation as an example. He argues that he can record a panoramic view of the stage and audience, but if he stops at a certain point and asks an AI to reconstruct what’s left, it will fail. He calls it “a mathematically intractable problem because we don’t know how to represent distributions over an infinite number of plausible outcomes” and says the idea is that AI should find an abstract representation of observations and make predictions in that abstract representation space, but “not try to reconstruct all the details that are not actually predictable.”

Precisely, LeCun has been working on that since he left Meta. He confessed that he has spent 15 years working on the idea of ​​self-supervised learning through video prediction, but that it failed during the first decade. Then it occurred to him to stop pursuing those generative models that try to predict at the pixel level to move on to that abstract work and that is where he is seeing progress.

The problem is that nothing is clear right now about this fact and what’s more, in a panel that followed, CEOs of AI companies criticized Yann’s position, pointing out that it was less about saying what is not right and more about providing solutions.

But well, beyond the “sting” that there may be, and more on a personal basis, it was a pleasure to listen to Yann LeCun and Steven Levy talking, and throwing darts, about this field that has so much left ahead of it.

Images | Xataka

In Xataka | The future of geopolitics is played in AI models: Claude Fable 5’s veto indicates that Europe is offside