Instead of passively losing your job, take the initiative to start a company that develops in this direction.
A few days ago, OpenAI released 722 mathematical paper manuscripts at once on GitHub. These manuscripts are classified into 372 result “families” covering 17 mathematical fields, all generated by an unreleased, unnamed internal model whose performance is far superior to GPT-6 Astra.
Tweet from Mark Chen, Chief Research Officer of OpenAI
Certainly, the mathematics community still has quite a few disputes over the quality of these outputs and the way they are disclosed, but one point is almost impossible to deny: The speed at which AI generates research results has started to be calculated in units of “hundreds of papers”.
Almost at the same time, in the latest episode of *The Information Bottleneck* podcast, a deep learning veteran said a sentence that made many researchers feel a jolt: “After finishing this project and getting some pretty interesting results, I suddenly realized: Wow, my work will probably be replaced in about five years.”
Podcast link: https://www.the-information-bottleneck.com/p/yuandong-tian-on-recursive-self-improvement-600
The person who said this is Yuandong Tian. He worked at Meta FAIR for nearly a decade, conducted research on Go AI, reinforcement learning, and neural network interpretability, and is also one of the authors of the Coconut (Continuous Latent Space Chain of Thought) paper.
Now, he is the co-founder of Recursive Superintelligence. In May this year, the company exited stealth mode and announced that it had completed financing of more than 650 million US dollars at a valuation of 4.65 billion US dollars, led by GV and Greycroft, with participation from NVIDIA and AMD.
In this episode hosted by Ravid Shwartz-Ziv, Yuandong Tian explained why he chose to start a business: “Rather than passively losing my job, it is better to take the initiative to start a company that moves in this direction.” However, he also poured a little cold water on AI during the program. “If you let the model come up with ideas on its own, it will give you some very mediocre ideas that almost everyone knows, and they are not interesting directions at all.”
In the nearly one-hour conversation, Yuandong Tian talked about the AlexNet debate back in his CMU days, then moved to Go AI, from “action space” to latent space reasoning, and finally landed on recursive self-improvement, new architectures and open
The quarrel in Smith Hall: Is AlexNet really an accident?
Yuandong Tian came to the United States in 2008 to pursue his doctorate at the Robotics Institute of Carnegie Mellon University. At that time, machine learning was far from having the status it enjoys today. “At that time, many people actually did not think machine learning was a reliable path. People would say it was not very useful, because it would overfit, and no one could figure out what was happening inside it.”
His doctoral topic was the non-convex optimization problem in computer vision, such as restoring image distortion caused by water surface refraction. He used a hierarchical method for optimization, and proved that even if the problem is non-convex, as long as there is enough data, the optimality can still be proven to a certain extent. “The combination of data-driven and optimization is very interesting. I am still very proud of this work to this day.”
In 2012, AlexNet was born out of nowhere. Yuandong Tian sent an email to Alex Krizhevsky to get the code and played with it himself. He quickly realized that the hierarchical processing of convolutional networks was essentially the same as the hierarchical optimization idea in his paper, “It might be something I have been pursuing, just in a different form.”
But the atmosphere at CMU at that time was not friendly. “Everyone came to the office every day and fiercely debated in Smith Hall whether the result of AlexNet was a fluke or a real breakthrough. Most people said it was just a fluke, because it did not make sense.”
When he graduated in 2013, Yuandong Tian received offers from several institutions that wanted to do deep learning, but finally chose the Google autonomous driving team, partly for its reputation, and partly for a friend’s referral. However, during the one year and three months he spent there, he did not get to work on deep learning, because the team was more like a startup that needed to use mature solutions to solve practical problems instead of conducting scientific exploration. So he jumped to FAIR, calling it “one of the best choices I have ever made”.
Two people, one GPU: Yuandong Tian’s past with Go
At FAIR, the first project Yuandong Tian launched was Go AI. In 2015, his team developed DarkForest, which defeated top amateur Go players online relying on convolutional networks. A few months later, AlphaGo defeated professional Go players in the spring of 2016.
Facing this kind of “collision of ideas”, Yuandong Tian’s review concluded that AlphaGo had two things he did not have:
- Value network, while DarkForest only had a policy network at that time;
- He himself had almost no background in reinforcement learning before, “I was basically learning while doing, just like doing reinforcement learning on myself”. This experience led him to carry out a large number of reinforcement learning researches later.
When asked if AlphaGo was unexpected, his answer was very interesting: the method itself was not amazing, because convolutional networks can capture the board patterns that humans rely on when playing Go, “The really surprising thing is that it can actually beat professional players.”
In 2018, the team reproduced AlphaZero, a solution that completely does not rely on human game records and plays against itself from scratch, and made some efficiency improvements to get OpenGo. In the official match with the Korea Baduk Association, this AI used only one V100 GPU, making four professional players fail to win a single game. “I think that was probably the first time someone used a single GPU to beat professional Go players.”
There was a small episode before the match. The team invited professional and semi-professional Go player friends to evaluate OpenGo’s strength, and the other side said that they could understand the opening moves, but after that they could not understand anything at all, and could only say “It looks very strong, but I don’t know if it is really strong”. So Yuandong Tian decided: then just go and play a match with real professional Go players.
The action space is more important than the algorithm
Talking about reinforcement learning in that era, Yuandong Tian believes that many methods are still underestimated to this day. For example, distributed reinforcement learning, which adjusts the model with the distribution of rewards instead of a single scalar; and different modeling methods such as Q-learning. “But now most of the time we use policy gradients, which is a very different era.”
He also values gradient-free optimization methods, and believes that large models give them new possibilities: “Large models now have a better high-level understanding of the system and the objective function to be optimized. This high-level understanding is much more useful than local gradient information.” In situations where the gradient fluctuates sharply and local optima are everywhere, the gradient may lead people to a dead end, while gradient-free methods naturally include exploration, “which is something gradient methods never do”.
Around 2018 and 2019, Yuandong Tian began to try to apply reinforcement learning to real-world problems, and the conclusion was quite unexpected: “What really matters is not the algorithm, but the design of the action space and state space.“
He took neural architecture search as an example. Humans know that network depth is very important, but AI does not. If AI starts to search for details such as the size of the convolution kernel of the first layer, all branches perform almost the same, the value function cannot give useful signals, and it can only go deeper layer by layer, leading to exponential explosion of the search space, “It doesn’t matter what algorithm you use at this time”. On the contrary, if the action space can be reorganized to make AI first “realize” that depth is very important, there will be a clear contrast between deep networks and shallow networks, and the direction can be seen after a few rollouts. Based on this, the team proposed Latent Space Monte Carlo Tree Search (LA-MCTS), whose core idea is to automatically find the action space that can maximize the discrimination of the search subspace.
Coconut: “I never use language when I think”
The co-host mentioned the Coconut paper that Yuandong Tian participated in, which allows the model to reason in the continuous latent space instead of the token space. Yuandong Tian said that this idea came from his self-observation: “I never use language when I think. I can clearly feel that I use another part of my brain when thinking, and then I have to translate my ideas into language to communicate smoothly with others.”
Starting from this intuition, the team did theoretical derivation and found that the latent representation can be a superposition state of multiple ideas. Using it for reasoning can bring significant advantages in problems such as graph traversal, “It may not be exponential, but it at least has certain advantages”.
Then why have cutting-edge models not adopted continuous chain of thought so far? Yuandong Tian gave two reasons.
- Data: A large number of human thinking processes are recorded in the form of language, while there is no such data for the latent space.
- The latent representation may be highly personalized, “Everyone more or less uses their own latent representation”, and the learning efficiency of large models is much lower than that of humans, so they can only use language as a substitute for the time being.
Ravid Shwartz-Ziv asked further whether this is similar to the regularization problem faced by self-supervised methods such as JEPA. Yuandong Tian agreed, and pushed the question further: Gradient descent may itself fall into a “local optimum at the meta level”. Some very good representations may not be learned by the current training paradigm at all, “We don’t know where the boundary is now”.
In his opinion, this is exactly the significance of interpretability research, “just like the difference between alchemy and chemistry”. However, the current models are so powerful that everyone is chasing results, and there are not enough people and time willing to devote themselves to this direction.
After co-writing the last paper with GPT-5
Talking about the transformation in the era of large models, Yuandong Tian believes that the analysis of neural network training dynamics that he has been doing continuously since 2018 is still connected in method when applied to large models, and there is no break. The real impact is at the empirical level: in the past, researchers put forward hypotheses, conducted experiments, and drew conclusions by themselves, and this process has no essential difference from that of one or two hundred years ago; but as the models get stronger and stronger, they can generate ideas on their own, “Researchers may be forced to think at the meta level instead of spending time putting forward specific hypotheses one by one”.
At the same time, the status of engineering capabilities has risen significantly. Many good works in the era of large models are “half research and half engineering”. Until the emergence of AI programming agents, the situation changed again: researchers with solid experience can let the agents quickly implement ideas and locate key problems, “Suddenly they have superpowers and can explore really deep things”.
The turning point was his last paper at Meta, for which he was also the only author. This paper studied the grokking phenomenon of neural networks, that is, the network first memorizes by rote, and suddenly learns to generalize at a certain moment of training.
During the whole process, he regarded GPT-5
