AI Research's Next Breakthrough: Build a Physics for AI First, Then Let It Discover Architectures on Its Own
Liu Ziming argues that real AFAI is not writing papers or tweaking training pipelines, but using physics methods to structure research first, then training a meta model to predict training curves, and ultimately letting AI discover the next generation of architectures itself.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
After ChatGPT, physics should turn around and explain AI
Liu Ziming studied physics as an undergraduate at Peking University, then went to MIT to work under Max Tegmark, starting out in AI for physics. After ChatGPT appeared, he decided to swap the roles of AI and physics, turning to physics of AI — using physics methodology to understand and design AI. He counts both himself and Max as "bird people," good at migrating methods across fields, which laid the foundation for his distinctive definition of AFAI.
— Liu ZimingA worst-case impossibility proof does not stop a physicist
KAN is Liu Ziming's breakout work, but Max opposed it at first, because a 1989 paper proved that the Kolmogorov-Arnold theorem cannot be turned into an algorithm in the worst case. Coming from physics, Liu Ziming held that you need not care about the worst case — the actual universe is not the worst-case universe. He built a prototype in two weeks, then spent one more week on visualization, and only after Max saw the network's internal structure did he exclaim and approve. This reflects his research style of "build the thing first, explain it later."
— Liu ZimingLLM success comes from language, not architecture
Liu Ziming agrees with Xie Saining's view: large language models are anti-bitter-lesson. The key to LLM success lies in language itself — language is a compression produced over millions of years of human evolution, and a model only needs to be able to "eat" this gift to succeed, so it places little demand on architecture. By contrast, modalities such as vision have not been well compressed, and world models need architectures designed to learn abstract representations — which is exactly why model design becomes important again in the post-LLM era.
— Liu ZimingMost AFAI only goes neuron-to-neuron, with no closed loop
Liu Ziming sums up his own research as one main thread: grafting the neural and symbolic worlds together. AI for science uses neural methods to capture symbolically expressed scientific laws; science of AI turns around and uses symbolic and scientific methods to understand and improve the AI black box. He points out that what people usually call AFAI is mostly neuron-to-neuron, with no closed loop back to the symbolic part; his AFAI is "AI for physics of AI" — using AI to accelerate understanding of AI itself, then guiding the design of better models.
— Liu ZimingTop talent would rather go solo than be a big-company grunt
Liu Ziming believes the root cause of the many neo labs appearing simultaneously in China and the US is that top talent's choices have changed: between academia, being a grunt at a big company, and going solo, more and more people choose the last. At the same time, the current paradigm (transformer + scaling) has obvious low-hanging basins it cannot reach, such as AutoResearch and world models, which need a lot of early-stage research rather than pure engineering grind — and these directions happen to be the living space for new labs.
— Liu ZimingHot money is all in world models and AutoResearch
Liu Ziming judges that the directions where capital is most active right now are world models and AutoResearch. A world model is essentially giving a robot a "brain," serving the robotics track; AutoResearch is training an AI researcher to do research. AFAI, meanwhile, is currently mostly classified by investors under AI for science. His firsthand feel from fundraising is that it is "insanely crazy" — investors probe each other for information and move extremely fast on projects, and after a seed round there can be unlimited follow-on rounds.
— Liu ZimingRSI competes on diligence, AFAI competes on hit rate
Liu Ziming compares Tian Yuandong's Recursive Superintelligence (RSI) with his own ideas. He thinks RSI's current route is more like a "diligent" coding agent, emphasizing action and iteration count; what he wants to build is a "smart" AFAI, emphasizing the hit rate of the ideas it proposes. He gives an example: a coding agent may propose 100 ideas and only 1 works, whereas he wants a model where 3 proposals yield 1 that works. The two are orthogonal directions that will ultimately need to be integrated.
— Liu ZimingTraining curves can be predicted in advance, and he is his own evidence
Since the start of this year, Liu Ziming has treated himself as a model being trained: each day he randomly pulls a dataset and a model, predicts the training curve first, then runs the experiment. For the first forty or fifty days his predictive ability was poor, then it suddenly became accurate. This "human experiment" convinced him that a meta model exists — a model that takes conditions such as model architecture as input and outputs the training curve. With a meta model, AI could rank ideas and predict experiments in advance, greatly accelerating AI research.
— Liu ZimingPeople who cannot train models should still be able to train models
Liu Ziming envisions the final product as a "training autopilot." Users need not know about transformers or understand training; they just have to state their needs clearly (for example, a budget of only 100 yuan), and the system can design, train, deploy and deliver a model end to end. He draws an analogy: coding now lets people who cannot write programs write programs, and in the future training autopilot will let people who cannot train models train models, and incubate "OpenAIs" in every vertical domain.
— Liu ZimingIn their own words · checked verbatim
Language is something we humans evolved over a million years; it is, so to speak, a gift 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 AFAI that is a smarter FAI. For example, I propose three ideas, and one of yours 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 a step in between: we first need to build this thing called physics of AI 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 until predictive ability became accurate | forty or fifty days | 50:56 |
| Share of well-structured AI papers | about 1% | 59:07 |
| Estimated chains of thought collectable per day | 200 | 1:01:12 |
| Estimated data volume collectable in two months | 10,000 | 1:01:12 |
| ResNet's performance advantage over second place on ImageNet | 10% | 1:16:06 |
| Time of return to China | March 2026 | 44:35 |
| Time fundraising started | end of May 2026 | 44:35 |
Glossary
- Physics of AI
- Studying AI with the ideas and methods of physics, to understand its behavior and guide its design.
- Mechanistic Interpretability
- Understanding a model's internal working mechanism by dissecting the neurons and representations inside the network.
- Meta Model
- A model that takes a model as input and outputs the training curve, used to predict AI experiment results.
- Recursive Superintelligence (RSI)
- The route of letting AI continuously self-improve, recursively raising its own intelligence.
How to listen
AI founders, investors watching AGI routes and mechanistic interpretability, and engineers who want to understand the frontier of AI research.
The personal history at the start and the hobby Q&A at the end can be skipped; focus on the middle section on the meta model route and funding observations.