OpenAI master Karpathy's latest sharing: Why OpenAI is most interested in AI Agents

Andrej Karpathy, the co-founder of OpenAI, recently gave a brief speech at a developer event, talking about his and OpenAI’s internal views on AI Agents (artificial intelligence agents).

Andrej Karpathy compared the difficulties in developing AI Agents in the past with the new opportunities developed with new technology tools. He also joked that his work at Tesla was "distracted by autonomous driving." He believed that autonomous driving Driving and VR are both examples of bad AI Agents.

Regarding new opportunities, Andrej Karpathy believes that now is the time to return to neuroscience again and seek inspiration from it - just like what happened in the early days of deep learning.

On the other hand, Andrej Karpathy believes that ordinary people, entrepreneurs and geeks have an advantage over companies like OpenAI in building AI Agents. Everyone is currently competing on an equal footing, so he is looking forward to seeing progress in this area. Results.

He also revealed that if a paper proposes a different training method, OpenAI’s internal Slack will sneer at it, thinking that it is all leftover by them. They will discuss the latest AI Agents papers with great interest.

The following is the full text of this sharing:

Hello everyone.

I was invited to give some motivational words on the topic of AI Agents. I think AI Agents are very close to me in a way. Let me start with a story. This is a very early OpenAI story. At that time, there were only maybe a dozen people in OpenAI. Around 2016, at that time, The trend is actually RL Agents (reinforcement learning agents).

Everyone was very interested in building agents, but at the time it was mostly based on gaming, and the excitement was around gaming companies like Atari, and my project at OpenAI at the time was trying to put the focus on RL Agents Use a keyboard and mouse to play on your computer, not games.

I wanted to make them more useful and do a lot of work, and the project was called World of Bits.

A few colleagues and I finally published a paper. This paper is not amazing because it is essentially based on RL reinforcement learning methods. Our webpage is very simple, allowing users to easily book flights or order food.

All this is obviously not feasible because the technology is not ready yet and it would be unwise to do these things at that time.

It turns out that we should completely forget about AI Agents and do language models.

We’re back here five years later, and I was distracted a little bit by self-driving, but now AI Agents are cool again, and our toolbox is completely different, and the way we approach these problems is completely different. .

Actually, all of you have done research on AI agents, but maybe not with reinforcement learning methods. It’s crazy and I don’t think we could have foreseen this at the time. This is so much fun.

Let me take a moment to talk about why AI Agents are so popular. I think it's clear to a lot of people that AGI (Artificial General Intelligence) will take full advantage of the capabilities of AI Agents, not just one, but many. Maybe there will be organizations or civilizations of digital entities, which I think is very inspiring and even a little crazy.

However, I also want to pour some cold water on this. In my opinion, there is a class of problems that are easy to conceive and demonstrate in your mind, but very difficult to turn into a real product. Many things fall into this category, I think autonomous driving is an example.

It's easy to imagine autonomous driving and build a demo car to drive around the block, but it would take a decade to make it a product. By the same token, I think the same is true for VR, it will take ten years to make it work.

I think the same is true for AI Agents to some extent. While it's easy to get excited imagining it, I believe it would take a decade of involvement to actually make it work.

The other thing I wanted to say is, I think it's interesting now to go back to neuroscience and in some ways be inspired by it again, the early days of deep learning were inspired by neuroscience.

It's very interesting to think about the relationship between them, especially because I think a lot of people look at language models as part of the solution, but how do you build a complete digital entity that has all the cognitive capabilities of a human?

There is no doubt that we all agree that we need an underlying system to plan, think about and reflect on the activities we are doing, and this is where neuroscience plays an important role.

For example, the hippocampus is very important. What in AI Agents plays the role of the hippocampus to realize functions such as memory storage, mark retrieval, etc.?

We have a preliminary understanding of how to build the visual and auditory cortex, but there are still many unknown things in AI Agents.

For example, what does visual gaming look like in AI Agents? What does the thalamus, the seat of the subconscious, correspond to in AI Agents?

This is very interesting. I actually brought a neuroscience book with me today, Brain and Behavior by David Eagleman, which I found very interesting and enlightening.

Perhaps now we should look to neuroscience for some interesting inspiration and redesign individual neurons, as we did before.

Finally I would like to end with some words of encouragement. An interesting but not obvious thing is that the AI ​​Agents you (referring to the audience) built are actually at the forefront of contemporary AI Agents capabilities. All large LLM institutions such as OpenAI, DeFi, etc., I suspect they are not at the forefront.

You are at the forefront.

For example, OpenAI is very good at training Transformer large language models. If a paper proposes some different training method, then the discussion in our internal OpenAI Slack group will be something like, oh yeah, someone tried it for two and a half years, and it didn't work, and we have no idea about this method. I know the ins and outs very well.

But when the new AI Agents paper comes out, we are all very interested and think it is very cool, because our team did not spend five years on it, we don’t know anything more than you do, we are doing it Compete with all of you.

This is why I think you are at the forefront of AI Agents capabilities, which is very important to the development of AI Agents.

The above is the detailed content of OpenAI master Karpathy's latest sharing: Why OpenAI is most interested in AI Agents. For more information, please follow other related articles on the PHP Chinese website!