OpenAI co-founder Karpathy shared: AI agents may take 10 years to really work

Organize: Twitter account "@GPTDAOCN"

Recently, Andrej Karpathy, the co-founder of OpenAI, gave a short 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 under new technological tools. He did not forget to ridicule his work at Tesla because he was "distracted by autopilot". He sees autonomous driving and VR as examples of bad AI agents.

For new opportunities, Andrej Karpathy believes that this is the moment 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 more advantages than companies like OpenAI in building AI Agents. Everyone is currently in a state of equal competition, so he is looking forward to seeing the results in this regard.

The following is the full text of this sharing:

Hello everyone. I was invited to speak some motivational words on the topic of AI Agents.

I think AI Agents is close to me in a way, let me start with a story, this is a very early OpenAI story, OpenAI maybe only a dozen people at that time, 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 game-based, and the excitement was around gaming companies like Atari, and my project at OpenAI was trying to focus RL Agents on keyboards and The mouse is used on the computer, not the game. I want to make them more useful and do a lot of work, the project is called World of Bits. A few colleagues and I ended up publishing a paper.

It's not an amazing paper, because it's actually based on RL reinforcement learning methods. Our web pages are very simple and allow people to book a flight or order some food and so on. None of this was obviously going to work because the technology wasn't ready and it wouldn't be wise to do this stuff at the time. Facts have proved that we should completely forget about AI Agents and do language models.

We came back here five years later, and I was a little distracted by autonomous driving, but now that AI Agents are cool again, our toolbox is completely different, and the way we approach these problems is completely different. In fact, all of you are working on AI Agents, but you may not be using any reinforcement learning methods. It's crazy, I don't think we would have seen this coming. It was just too much fun.

Let me take a moment to talk about why AI Agents are so popular. I think it's clear to many that AGI (artificial general intelligence) will take full advantage of the capabilities of AI Agents, not one, but many. Maybe there will be organizations or civilizations of digital entities, which I think is very inspiring, even a little crazy. However, I also want to throw some cold water on this. I think there's a whole class of problems that are easy to imagine, easy to build, easy to demonstrate, but really hard to make a product. A lot of things fall into this category, I think autonomous driving is an example.

Autonomous driving is easy to imagine, and easy to build a demo of a car driving around a block, but it took a decade to turn that into a product. In the same way, I think it's the same with VR, it's going to take a decade to make it work.

I think the same is true to some extent with AI Agents. It's easy to imagine scenarios for it, and it's very exciting, but I think if you're in it, you should invest a decade in it to actually work.

Another thing I want to say is that I think it's interesting to go back to neuroscience now and be inspired by it again in some ways, the early days of deep learning were inspired by neuroscience. It's very interesting to think about the relationship between them, especially 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 humans? Clearly, we all agree that we need some sort of underlying system to plan, think, and reflect on what we're doing, and this is where neuroscience comes into play.

For example, the hippocampus is very important. What in AI Agents plays the role of the hippocampus, which is used to store memory, mark retrieval, etc.? We have a general understanding of how to build the visual and auditory cortices, but there are many things we don't know what it means in AI Agents. Like what does a visual game look like in AI Agents? What is the equivalent of the thalamus, the location of the subconscious mind, in AI Agents? It's very interesting.

I actually brought a neuroscience book with me today, The Brain and Behavior by David Eagleman, which I found very interesting and enlightening. Taking some interesting inspiration from neuroscience, as we did in the early days when we engineered individual neurons, we should probably do it again today.

Finally I would like to end with some words of encouragement. An interesting but not obvious thing is that the AI Agents you build (referring to the live audience) are actually at the forefront of the capabilities of contemporary AI Agents. All the large LLM institutions such as OpenAI, DeFi, etc., I doubt they are 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 kind of different training method, the discussion in our Slack group inside OpenAI will be like, oh yeah, someone tried it for two and a half years and it didn't work, we're not interested in this method Very understanding of the ins and outs. But when the new AI Agents paper came out, we were very interested and thought it was very cool, because our team didn't spend five years on it, we don't know anything more than you, we are working with all of you People compete together. That's the reason to think you're at the forefront of what AI Agents can do.