The Next Breakthrough in AI Research: Build a Physics of AI First, Then Let It Discover Architectures Itself
Liu Ziming argues that real AFAI is not writing papers or rewriting training pipelines, but using the methods of physics to structure research first, then training a meta model to predict training curves, so that AI eventually discovers the next generation of architectures on its own.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
After ChatGPT, physics should be turned around to explain AI
Liu Ziming (刘子明) read physics as an undergraduate at Peking University, then went to MIT to work under Max Tegmark, where his early research was AI for physics. Once ChatGPT appeared, he decided to swap the two roles and turn to physics of AI: using the methodology of physics to understand and design AI. He puts both himself and Max in the category of ‘birds’, researchers who are good at carrying methods from one field into another, and that disposition is the foundation of his own particular definition of AFAI.
— Liu ZimingA worst-case impossibility proof does not stop a physicist
KAN is the work that made Liu's name, but Max was against it at first, because a 1989 paper had proved that in the worst case the Kolmogorov-Arnold theorem cannot be turned into an algorithm. Coming from physics, Liu took the view that the worst case is not worth worrying about: the universe we actually live in is not the worst possible universe. He built a prototype in two weeks and spent another week on visualisation, and only when Max saw the internal structure of the network did he react with admiration. It is a clean illustration of Liu's research style, which is to build the thing first and explain it afterwards.
— Liu ZimingLLMs succeeded because of language, not because of architecture
Liu agrees with Saining Xie (谢赛宁) that large language models are anti-bitter-lesson. What made LLMs work is language itself: language is the compressed product of a million years of human evolution, and a model only has to be able to ‘eat’ that inheritance in order to succeed, which is why the demands it places on architecture are low. Other modalities, vision above all, have not been compressed in the same way, so a world model needs an architecture designed to learn abstract representations. That is exactly why model design has become important again in the post-LLM era.
— Liu ZimingMost AFAI work runs neural to neural and never closes the loop
Liu sums up his own research as a single thread: grafting the neural and the symbolic worlds together. AI for science uses neural methods to capture symbolic scientific laws; science of AI runs the other way, using symbolic and scientific methods to understand and improve the AI black box. What most people call AFAI, he points out, mostly goes from neural to neural and never closes the loop back to the symbolic part. His AFAI is ‘AI for physics of AI’: using AI to accelerate the understanding of AI itself, and then using that understanding to design better models.
— Liu ZimingTop talent would rather go solo than be foot soldiers at big labs
Liu's view is that the reason a wave of neo labs has appeared in China and the US at the same time comes down to a change in what top talent chooses: between academia, being a foot soldier at a big company, and going out on their own, more and more people pick the last. At the same time the current paradigm, transformer plus scaling, leaves obvious low ground it cannot reach, such as AutoResearch and world models, which need a great deal of early-stage research rather than sheer engineering effort. Those directions are precisely where the new labs have room to survive.
— Liu ZimingThe hot money is all in world models and AutoResearch
Liu's read is that the most active directions for capital right now are world models and AutoResearch. A world model is essentially about giving a robot a ‘brain’, serving the robotics track; AutoResearch simply means training an AI researcher to do research. AFAI, for now, mostly gets filed by investors under AI for science. His own felt experience of raising money is that ‘it is insane’: investors sound each other out for information and move on deals extremely fast, and after a seed round you can keep adding rounds without limit.
— Liu ZimingRSI competes on diligence; AFAI competes on hit rate
Liu compares Yuandong Tian's (田渊栋) Recursive Superintelligence (RSI) with his own idea. As he sees it, the current RSI route looks more like a ‘diligent’ coding agent, where what counts is capacity to act and number of iterations; what he wants to build is a ‘smart’ AFAI, where what counts is the hit rate of the ideas proposed. His example: a coding agent might put forward a hundred ideas and have only one that works, whereas he wants a model that proposes three times and gets one that works. The two are orthogonal directions, and in the end they will have to be integrated.
— Liu ZimingTraining curves can be predicted in advance, and he is the evidence
Since the start of this year Liu has been treating himself as a model in training: every day he pulls a random dataset and a random model, predicts the training curve first, and then runs the experiment. For the first forty to fifty days his predictions were very poor, and then they suddenly became accurate. That ‘human experiment’ is why he believes the meta model exists, that is, a model that takes conditions such as a model architecture as input and outputs a training curve. With a meta model in hand, AI could rank ideas and predict experiments before running them, which would accelerate AI research enormously.
— Liu ZimingPeople who cannot train models should still be able to train models
The end product Liu envisages is a "training autopilot". The user does not need to know what a transformer is and does not need to understand training; they only have to state the requirement clearly, for example that the budget is only 100 yuan, and the system designs, trains, deploys and delivers the model end to end. His analogy: coding now lets people who cannot write programs write programs, and in the future a training autopilot will let people who cannot train models train models, incubating an "OpenAI" for each vertical field.
— Liu ZimingIn their own words · checked verbatim
Language is something we humans evolved over a million years; it amounts to a gift that nature gave us.
语言是我们人类演化了百万年，相当于是自然给我们的一种馈赠。
Liu Ziming18:40
What we want to do here is an AFAI that is a smarter FAI. Say I propose ideas three times, and one of them works.
我们这里想做的是AFAI是一个更聪明的FAI。 比如说我提三次想法，可是你们有一个就work了。
Liu Ziming47:41
But when I say stop, there is actually still a step in between: we first have to do physics of AI, and that is the thing that can help us design AI.
但我说stop就中间其实还有一步，我们是先要做physics of AI这个东西能够帮助我们去设计AI。
Liu Ziming1:10:46
Figures
| Days of self-training | 60 days | 49:49 |
| Days before his predictions became accurate | forty to fifty days | 50:56 |
| Share of AI papers that are well structured | about 1% | 59:07 |
| Projected chains of thought collectable per day | 200 | 1:01:12 |
| Projected volume of data collectable in two months | 10,000 items | 1:01:12 |
| ResNet's performance margin over the runner-up on ImageNet | 10% | 1:16:06 |
| Return to China | March 2026 | 44:35 |
| Start of fundraising | late May 2026 | 44:35 |
Glossary
- Physics of AI
- Studying AI with the ideas and methods of physics, in order to understand its behaviour and guide its design.
- Mechanistic Interpretability
- Understanding how a model works internally by taking apart the neurons and representations inside the network.
- Meta Model
- A model that takes a model as input and outputs a training curve, used to predict the results of AI experiments.
- Recursive Superintelligence (RSI)
- The route in which AI keeps improving itself, recursively raising its own intelligence.
How to listen
AI founders, investors following AGI roadmaps and mechanistic interpretability, and engineers who want to understand the research frontier.
The personal-background opening and the hobby Q&A at the end can be skipped; the middle stretch on the meta model route and his fundraising observations is the part to hear.