Jan LeCun is called one of the godfathers of artificial intelligence. In the late 80s, he and two other scientists — Jeffrey Hinton and Joshua Bengio developed the concept of convolutional neural networks. Today, they are the standard for image recognition and processing. They are used for object detection, medical diagnostics, face recognition, video analysis, and many other tasks.
Ian is Meta’s chief AI scientist, and his main goal is to create a new type of artificial intelligence that will be on par with human intelligence and will reason and understand the world in the same way as humans and animals do. He spoke at the AI for Good conference in Geneva about how to get there, whether artificial intelligence will be on par with human intelligence, and what we should really be afraid of. I recorded the most interesting moments from his speech.
Content
If you’re interested in creating superintelligence or making AI equal to human intelligence, then large language models (LLMs) are a dead end. This is not to say that LLMs are useless, but if we want to achieve more, we will have to invent new methods and new architectures.
Let’s try to ask four questions. How do we understand the physical world? How do we have a permanent memory? How do we reason? How do we plan?
These four entities are at the heart of intelligence. None of the LLMs that exist today are capable of planning, drawing conclusions at the right level, and generally do not have such fundamental capabilities as humans.
Someday we will learn to create systems that will surpass human intelligence in most areas, and I hope this will happen within the next decade. To do this, we need to solve two fundamentally important tasks.
The first is the type of response generation to a user’s request — inference. In the process of reasoning, LLM passes through a fixed number of layers of the neural network and finally produces an answer, also known as a token. The number of calculations that LLM can perform to generate an answer is — fixed. And this is the problem.
LLM spends the same number of iterations to find the answer to a simple question where it is enough to say «yes» or «no», and to a complex question where you need to think deeper. And it makes no sense. After all, a simple question can be answered immediately, but a more complex one triggers the process of reasoning.
During reasoning, people go through possible scenarios in their heads, and they don’t need to write anything down, except, for example, in math or programming. In most cases, we reason, plan, and make decisions using an abstract model of what is happening in our minds — what psychologists call second-order thinking.
Currently, artificial intelligence calculates the answer, but we want it to learn to think for a long time and search for it. This is a fundamentally different process.
The second important thing: AI needs to be «grounded» in our reality. It doesn’t have to be physical — it can be virtual, but there must be some base from which AI receives information with a higher bandwidth than just text. Humans receive a huge amount of information through sight, hearing, and touch. A four-year-old child already knows as much as an LLM trained on all available texts on the Internet. We won’t reach the human level of AI unless we let systems understand the real world.
We come to the main problem — the imperfection of modern architectures. Modern LLMs are trained to predict the next token, i.e., to pick up individual text fragments one by one, rather than model the surrounding reality as humans do.
A new architecture is needed for the system to understand the world. I advocate JEPA — Joint Embedded Predictive Architecture. This is a system that is fundamentally different from LLMs. In this new system, LLMs will still play their role, because they are great at turning our thoughts into coherent and understandable text.
But JEPA — is not LLM, it is a completely different entity.
It may seem strange to many people, but language — is a relatively simple thing because it is discrete: it has a limited number of words, and it is not that difficult to guess what the next one will be — which is what LLMs do. But the real world is much more complicated.
If you show an LLM a video of a conference room and ask them to predict what will happen next, they will get confused. It is impossible to guess every detail, like what each person in the room looks like or what color the carpet is The idea behind JEPA is not to predict tokens, but abstract representations. It operates at a higher level of abstraction, ignoring details that cannot be predicted. This results in a model that better understands what is happening.
When we figure out how to get to the level of AI close to human, it won’t be a quantum leap — we’ll go through levels of intelligence like rats, cats, dogs, etc. We will have time to make such systems safe. The AI of the future will be objective-driven, it will work to achieve a clearly defined goal, and with «constraints» that the system cannot violate. The current LLMs have a different essence of work, the goal is set by the user in the prompt — and this is a very weak specification, completely dependent on how and on what this LLM was trained.
In the new architecture, the goal is written from the beginning, and the system cannot do anything but solve the user’s task within the given constraints. The only thing left to do is to figure out how to formulate these goals and constraints correctly — a difficult but not impossible task. The probability that such a system will cause harm is about the same as the probability of the next plane you fly exploding — extremely small, because the design is done by professionals who have been improving safety systems for decades.
This does not mean that we need to regulate AI completely. AI used for medical diagnostics or driver assistance undergoes tests and inspections, receives approval from government agencies, just like aviation. There is no such regulation for chatbots, as the risks are low and no serious damage has occurred in the two years of their mass use.
My role at Meta is research and long-term strategy. My job is to inspire, to promote ideas like JEPA. It was FAIR, Meta’s research arm, that made Llama open source, as well as many other developments — more than a thousand projects in 11 years.
I believe that AI should be built on the basis of open source systems. Historically, open source platforms are safer, more reliable, and more flexible. This is how the Internet and mobile networks work. More eyes — faster bug fixes. This is especially important for AI: it is impossible to check the security of a system if you cannot modify and test it.
In my opinion, the biggest danger today is not the risk of AI doing something wrong, but that all our digital interactions will go through one or two «closed» AI agents, and you will get all the information from a couple of sources. Already, a lot of our digital consumption is the result of machine learning: not very smart, but they write, filter, and moderate a huge amount of content.
In the future, our entire digital diet will come from AI systems.
Therefore, we need a diversity of AI assistants with different languages, cultural values, and political views. This can be compared to the press: if we have only one monopolist, it is bad for society. And many countries will not accept the threat that all information will come from, for example, the United States or China. Open source is the only way to ensure such diversity.
Sometimes you hear suggestions that open-source — is bad, look, Chinese DeepSeek was trained on Llama. You might think: «what if we close down development in the West and slow down the spread of ideas so that competitors fall behind?».
But they will find out anyway, just with a delay. The problem is that we will move slower ourselves, we will slow ourselves down. It’s like shooting yourself in the foot. We need to be the best at taking new ideas and implementing them. In addition, we, in turn, learned from DeepSeek and other models, including our Chinese colleagues, who actively make their models publicly available.
Even if you completely isolate China intellectually, they still have excellent engineers and scientists, and they will quickly catch up, and perhaps even overtake the West. The DeepSeek example came as a shock to many in California.
In the future, I think the situation will be similar to the current operating system market: 2-3 dominant platforms, among which one or two will be open, plus a couple of closed and niche platforms for special tasks. I think so because I see a future in which LLMs and what will follow them will become the repository of all human knowledge and culture. To realize this, they will need access to data from all over the world, from different countries and regions.
But countries are not ready to share data for free — they need AI sovereignty, i.e. control over their own data and how artificial intelligence learns and works on it.
Therefore, the only way out is an international partnership in which future models will be trained in a distributed manner, sharing not data but model parameters. This will result in a common consensus model without each country losing sovereignty over its data. And it must be open source.
The main thing is that we need open, safe, diverse AI that enhances, not replaces, humans or creates virtual people.
Контент сайту призначений для осіб віком від 21 року. Переглядаючи матеріали, ви підтверджуєте свою відповідність віковим обмеженням.
Cуб'єкт у сфері онлайн-медіа; ідентифікатор медіа - R40-06029.