Shulou(Shulou.com)11/24 Report--

Karpathy believes that AI agents represent a future of AI, and its development needs to draw inspiration from neuroscience and encourage researchers in the field of AI agents to continue their efforts.

Recently, Andrej Karpathy, co-founder of OpenAI, former director of TeslaAI and now returning to OpenAI, shared his views on AI agents at a developer event.

Seven years ago, when the time was not ripe to study AI agents, he first talked about his early days at OpenAI (around 2016), when the industry trend was to study how to use reinforcement learning to improve AI agents.

Many projects are based on games like Atari to create AI players.

What he wanted to do at that time was a product with a wider range of applications.

But because of the limitations of the technology at that time, the effect was not good, so he and OpenAI changed direction and began to build a big language model.

Of course, I was distracted by autopilot during this period.

But now, five years later, the AI agent has become a very promising direction again.

Because there are entirely new technologies to study AI agents, the situation is completely different from that of 2016.

The simplest example is that no one is using reinforcement learning to study AI agents as they did in 2016.

The current research methods and directions were unimaginable at that time.

The AI agent represents a crazy future, although it may be a little far away because in the future, if AGI can appear, it will give full play to the power of the AI agent.

The future AI agent may not be a single individual, but there will be a lot of AI agent organizations, or even AI agent civilization.

This could be a very exciting, even crazy future.

But at the same time, everyone should stay awake and calm.

Because some technological trends may be easy to conceive and foresee, but it is difficult to get a product on the ground.

Many technologies belong to this type, such as autopilot.

The technical vision is easy to imagine, and a demonstration of a car driving on the block is easy to do, but it can take 10 years to make a product.

The same is true of VR.

AI agent may also belong to this category of technology, the application scenario is easy to imagine, the prospect is exciting, but it requires long-term technology development and accumulation.

AI agents need to draw inspiration from neuroscience, just as in the early development of deep learning, the development of AI agents may be inspired by neuroscience.

It's interesting to think about the relationship between AI agents and neuroscience.

Especially now many people use the large language model as part of the AI agent solution.

But how to build a complete digital entity with all the cognitive abilities of human beings?

Obviously, we all agree that some kind of potential system is needed to plan, think, and reflect on what we are doing.

This may be where neuroscience can play a role.

For example, the hippocampus is a very important part of the brain.

But what in the AI agent acts as the hippocampus for memory storage, tagging, retrieval, and so on?

We already know a lot about how to build the visual and auditory cortex, but there are a lot of things we don't know what it means in the AI agent.

For example, what is the thalamus, the location of the subconscious, equivalent to in AI Agents?

These are very interesting questions.

I brought a special book on neuroscience, which has David Eagleman's brain and behavior, which I find very interesting and enlightening.

As early AI studies did when designing neurons, drawing interesting inspiration from neuroscience may be the direction we should try again.

All of you here are at the forefront of the industry, but you may not know it, but the AI agent built by everyone here today is already at the forefront of AI agent capability.

Now all the institutions that are building large language models, such as OpenAI, I don't think they are at the forefront of this field.

At the forefront are all of you here.

For example, OpenAI is very good at training the Transformer language model.

If a paper proposes a different training method, then people in our Slack group within OpenAI will discuss and say:

"I tried this two and a half years ago, and it didn't work. "

We are very clear about the context of the method of training the model.

But when the AI agent paper comes out, all of us will be very interested and will think it's great.

Because our team has spent the last five years somewhere else.

We don't know more than you do in this field, and we stand on the same competitive level as you all.

That's why I think all of you here are at the forefront of AI agents, which is very important for the development of AI agents.

Reference:

Https://twitter.com/gptdaocn/status/1673781206121578498

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.