The Technology Ladder: From CoT and Agent to Continuous Learning

Then we crossed another step. This year's step is Agent, because we found that with an Agent approach, it can handle even more tasks. Its range of capabilities expands, and its intelligence ceiling rises higher. Why is it a staircase? Because each subsequent step builds on the previous ones. Agent relies on CoT, and CoT relies on the step before that-the language model. So no step is wasted.

So, the development of AI and the trajectory of intelligence are traceable. This year's step is Agent, but the Agent step will also reach its limit. Even after it solves all the problems it can, it still can't replace your employees. It will have reached the upper bound of its capabilities.

Just like CoT: after CoT reached its limit, it had already surpassed the best humans at solving math olympiad problems and writing code. But it still stopped there; that technology did not reach AGI. So you see, the direction of AI intelligence is traceable.

After Agent, we think the problem to solve is continuous learning-how to make the model learn continuously, rather than requiring a massive training session. It should be able to engage in long-term continuous learning, like humans. This problem is related to task completion and others; they are the same issue. Standing at the Agent stage, we can see the next bottleneck: continuous learning. The next challenge to tackle is how to achieve continuous learning. This is visible and relatively clear.

This is the obstacle ahead; we must overcome it, and there is definitely a way, but it takes time. After continuous learning, we might arrive at a singularity. That singularity is when the model, after achieving continuous learning, can do everything humans can do.

It would then be able to develop its own versions, conduct its own research, and develop the next generation of AI models. So it reaches a singularity where it can iterate on itself. But this singularity is not really a singularity; it's a gradual process. It may be a long, gradual change, not a sudden leap. But out of habit, we tend to call it a singularity.

Because long ago, forecasters predicted a singularity here, but it's actually not a singularity; it's a continuous process. After this step, I think we'll get to embodied intelligence.

This is our speculation. We think the timeline should be: first solve learning to learn, then the intelligence singularity where it can self-iterate, and then embodied intelligence. After embodied intelligence, it enters the physical world and can do household chores and provide elderly care.

We think this is an ideal roadmap, but everyone has different opinions-there's no right or wrong. We just think this roadmap is the easiest. Because each step requires very little new work. With this roadmap, we don't have to work overtime. But if the roadmap is reversed, say we have to achieve embodied intelligence first, then it would be exhausting-a very hard job. We don't want that kind of roadmap; we want an easier one.

If we first solve continuous learning, then the self-iteration singularity, then embodied intelligence-the path is very easy. Because later on, you can use earlier technologies to help develop later ones. After the singularity, doing embodied intelligence doesn't require human effort anymore; the model can do it itself.

So this answers the question of what our long-term goal is. I told him, this is our long-term goal-AGI. Now back to reality. Last year's key reality was that everyone wanted to build a chatbot and compete for consumer-side traffic. This year's reality is that everyone wants to capture enterprise-side revenue and get involved, because if you don't, you're not even at the table, right?

But we don't see that as important. Internally, what we truly care about is the AGI roadmap I just described and how to break through the next technology. But strangely, the things you want most are the ones you don't get. The things you don't care about as much come easily.

There's a strategic advantage here: with our minds set on AGI, we work on AGI. So when we do applications, consumer side, or enterprise side, we don't need to put much thought into it-it takes very little effort. I think it's like standing on a high technological ground to work on lower-level technology; it's a dimensionality reduction attack.

At least on the consumer side last year, we saw that was the case. We didn't put much effort into the consumer side; at one point we even didn't want to maintain those users, but the users just wouldn't leave. Because they really wouldn't go, so they stayed. But slightly... and now this year, the enterprise-side revenue is showing optimistic growth. I think the numbers might be relatively good compared to peers, I estimate. But we didn't put much effort into this; we didn't even... do anything; it was just incidental. When deploying internet intelligence, the step toward AGI is something I must take - it's a stepping stone. On the road to AGI, I have to pass this step. So I offer these technologies to everyone via API, without doing anything extra. We are still doing AI; this is a byproduct. I just need a few people to maintain this API, no customer service, no sales, nothing - users come on their own.