# How Far Can Artificial Intelligence Really Go? | Yann LeCun - World Governments Summit

*World Governments Summit* · https://www.youtube.com/watch?v=SbISeWn3FSk

## Summary

In this interview at the World Governments Summit, Yann LeCun, Turing Award Laureate and Meta's Chief AI Scientist, argues that open-source AI is essential for security, customization, and diversity of opinion, comparing it to the role of a free press. He contends that AI will become the infrastructure of the digital world and that infrastructure software historically becomes open source. He also pushes back on fears of AI taking over, saying we are still very far from human-level intelligence and that current LLMs lack persistent memory, reasoning, planning, and an understanding of the physical world.

## Key points

- Yann LeCun argues we need free and diverse AI platforms for the same reason we need a free and diverse press, so a single opinion doesn't dominate.
- He says infrastructure software historically becomes open source because it is more secure, more customizable, and enables ecosystems to form on top of it.
- He claims open source software like Linux is more secure because more eyeballs are on it, and Linux now runs most of the world's systems.
- He states we are still very far from matching human and animal intelligence with AI systems, and it will not happen in the next 2 years.
- He says current large language models lack persistent memory, reasoning, and planning, and cannot understand how the world works.
- He notes LLMs are trained on roughly 10 trillion tokens, which would take a single person 150,000 to 200,000 years to read 8 hours a day.
- He compares AI understanding of the physical world unfavorably to that of a cat, while acknowledging humans also run into glass doors.

## Notes

### Open-Source AI as Infrastructure
- Yann LeCun, Turing Award Laureate and VP & Chief AI Scientist at Meta, argues AI will soon mediate all our interactions with the digital world, as seen with AI-backed replacements for search engines.
- This requires a wide diversity of AI systems serving different languages, cultures, value systems, and interests — something that cannot be produced by a small number of companies on the West Coast of the US or by any single country.
- History shows infrastructure software ends up open source because it is more secure, more customizable, and enables ecosystems on top of it — as with the internet and cell phone software stacks.
- Since Llama 2 became available, people have been customizing open source models for vertical business applications and for particular languages and cultures.

### Why Open Source Is Safer
- "We need free and diverse AI platforms for the same reason that we need free and diverse press, so we don't have a single opinion coming to us from AI."
- Open source software is historically more secure because there are more eyeballs on it.
- Linux runs the world: most phones (everything not an iPhone), web servers, cloud services, car microcontrollers, and even the helicopter on Mars.
- The only reason to oppose open sourcing AI is a belief that AI is so powerful it will take over the world — which LeCun says is not happening.

### How Far AI Can Really Go
- We are still very far from matching human and animal intelligence and learning capacities; it is not going to happen in the next 2 years.
- Scientific breakthroughs are still needed for machines to understand how the world works, remember facts (persistent memory), reason, and plan — essential characteristics of intelligent behavior that current LLMs lack.
- LLMs are trained on roughly 10 trillion tokens (10 to the 13th power), about 2 × 10 to the 13th bytes, which would take one person 150,000 to 200,000 years to read at 8 hours a day.
- Even a cat has a much better understanding of the physical world than the most powerful AI systems — though LeCun concedes he has run into glass doors himself.

## Transcript

[00:00] How specifically does being open-source not lead us to our end? >> Dr. Yann LeCun, Turing Award Laureate, Vice President and Chief AI Scientist [music] of Meta. >> We need free and diverse AI platforms for the same reason that we need free and diverse press, so we don't have a single opinion coming to us from AI. We're still very far from matching human and animal [music] intelligence and learning capacities with AI systems. It's not just around the corner. It's not going to happen in the next 2 years. >> Can you talk to us, Yann, about open-source AI? Why is open-source the way forward from a security and innovation perspective? >> So, there's going to be a a future, not very far in the future,

[00:45] where all of our interaction with the digital world are going to be mediated by AI systems. We already see this with uh uh companies putting out basically replacement for search engines uh that are backed by AI systems. And so, in this in this future, if all of our digital diet is mediated by AI systems, um it's going to be there's going to be a need for a a a very wide diversity of AI systems that cater to different languages, different cultures, different value systems, different centers of interest. And it it can't be produced by a small number of companies on the West Coast of the US. Um um or any single country. Uh so, essentially AI tools, assistants, will become kind of a repository of all human knowledge.

[01:39] And um it's going to be a a kind of infrastructure. And the history shows that infrastructure software always end ends up being open-source because it's more secure, it's more customizable. Uh it enables the creation of an ecosystem on top of it. And so the the entire infrastructure of our communication systems, the internet, um even our cell phones, uh software stack for the, you know, communication communicating with cell phone towers, it's all open source. For very good reasons. It's not by design, it's really the the market forces that push it. So there's a similar phenomenon which is definitely happening now in the last uh uh few months that Llama 2 has uh has been available and more open source uh

[02:29] uh and then it's been available that people are using those open source models and and customizing them for vertical applications in businesses, but also uh training them for, you know, a particular types of uh uh languages or or cultures. Uh and so it's a necessity. We we need it. We need free and diverse AI platforms for the same reason that we need free and diverse press. Um so we don't have a single opinion coming to us from from AI. Getting towards the the title, will AI lead us to our end? How specifically does being open source not lead us to our end? Is there something fundamental that makes it easier to secure? Yes. Uh so it's it's certainly been the case historically that uh open source

[03:16] software has been more more secure simply because there's more eyeballs on it, right? So the entire world today runs on Linux. Um maybe not your desktop, uh but probably more more than half of your phones run on Linux. Whatever is not an iPhone runs on Linux. Uh the web servers around the world run on Linux. All the cloud services run on Linux. Your car has several microcontrollers that probably run Linux. The uh helicopter on Mars runs Linux. I mean, Linux runs the world. And the reason is because it's open source, so you can port it to anything you want, and you can make it more secure. And and so, you know, there's kind of a similar story uh to be uh to be made for uh for for AI uh business models. More eyeballs, more

[04:04] ways to uh make it safe, um uh etc. Now, the only reason why you might think that it's not a good idea to open source is if you have this belief that somehow AI is so powerful that it is going to take over the world at some point. Okay. And uh and this is this is not happening. Um so, first of all, we're still very far despite all the wonderful things that AI can do today, uh we're still very far from matching human and animal intelligence and learning capacities with AI systems. We're very far. It's not just around the corner. It's not going to happen in the next 2 years. Um we still some need some scientific breakthroughs to get to the type of learning abilities that we observe in uh not just humans, but even your your cat.

[05:00] Um your cat has a much better understanding of the physical world than uh the most powerful AI systems that we have. >> I've seen my dog run into the glass door in my house. Y- Y- I I I I I debate that, but yeah. >> Sure, but you know, I've run into glass doors. Um so, the the uh H- Here is an interesting uh statistic. Um so, first of all, before we get to human-level AI, uh we're going to need some breakthroughs that allow machines to understand how the world works, which they can't do at the moment, uh can remember facts, which they can't really do at the moment. They don't really have persistent memory, at least the LLMs don't. Um they can reason, and they can plan. And those are characteristics that are

[05:45] essential to intelligent behavior. Current large language models cannot do any of those things, at least not to the level that uh we'd like them to to do. And the reason is is relatively simple. Um if you if you look at how much data is used to train those LLMs, it sounds staggering. It's uh typically 10 trillion tokens, right? A token is like a word, so it's 10 to the 13, right? A one with 13 zeros. Uh words or word equivalents. Each word or each of those tokens is stored in two bytes, so that's 2 * 10 to the 13 bytes. Seems staggering. If you evaluate how long it would take for humans to read this, it would take, you know, anywhere between 150,000 and 200,000 years for a single person to read this reading 8

[06:34] hours a day. Uh enormous. So, the conclusion is those systems are incredibly smart, right? Because they've they're capable of digesting this enormous amount of knowledge.