Liu Ziming | 张小珺Jùn｜商业访谈录 Summary

1. Key Themes

The New Lab Phenomenon: Research-Driven Founders as a New Asset Class

A new wave of "New Labs" is emerging in both the US and China, founded by young researchers with elite academic backgrounds who can raise hundreds of millions in valuation before having a product. Liu Ziming describes the fundraising environment as almost irrational in its speed and intensity.

"I returned to China in March, joined Tsinghua around late April, and started participating in fundraising around late May. The process — my gut feeling is — it was just too crazy." 00:01:21

"Everyone is scouting information from each other, and everyone's investment actions are moving very fast." 00:01:38

Physics of AI as a Framework for Understanding the Black Box

Liu Ziming's central intellectual thesis is that AI should be studied the way physicists study the natural world — through observation, empirical laws, and eventually first-principles theories. He identifies three methods: looking spatially deeper (into neurons and representations), looking temporally (how the model evolved during training), and using "parallel universes" (controlled experiments) to map phase diagrams of AI behavior.

"Nothing in biology makes sense except in the light of evolution. If you don't understand biology from an evolutionary perspective, nothing in biology is interpretable — it's a black box. But if you look at it from an evolutionary angle, you at least have some understanding." 01:05:07

"The task of Physics of AI is to clarify for every trick and every architecture what its applicable conditions are. That is a scientific attitude — not just making something, then going out to sell it while completely ignoring the fact that it might be good only under certain conditions." 01:07:26

AI Research Is at Best "Stage 1.5" — Dangerously Fixated on One Star Cluster

Liu Ziming uses the history of astronomy as a direct analogy for where AI science stands today — somewhere between Tycho Brahe's data collection and Kepler's empirical laws — and warns we may not even be at the Tycho Brahe level because the field is over-committed to Transformers.

"We are overly focused on prematurely solidifying on architectures like Transformer. It's a bit like Tycho Brahe fixating on one patch of sky, saying 'this patch of sky is just too marvelous,' and ignoring observations of other parts of the sky. From this perspective, we may not even be at the Tycho Brahe era yet." 00:27:37

"Scaling Law is one of the very few empirical laws we know — its status is similar to Kepler's laws. We are still far from reaching the Newton stage of AI." 00:26:53

Auto-Research Requires a Structural Language for Research That Doesn't Yet Exist

The key barrier to AI for Research (unlike AI for Coding) is that research has not yet been structured or standardized. Liu Ziming argues that GitHub provided a naturally structured, high-quality corpus for coding, but an equivalent corpus for research doesn't exist — and building it is prerequisite to building the system.

"Research is not a well-structured data structure. When we still use words like 'research intuition' and 'research taste,' we are implicitly admitting that research cannot be well-structured, standardized, or scaled. Once something can be structured, standardized, and scaled, people should start to demystify it. We haven't demystified research yet." 00:40:17

"The reason AI for Coding succeeded is that code is highly structured — that's point one. Point two is that it has enormous amounts of data: GitHub. The data on GitHub is far higher quality and more structured than other parts of the internet." 00:39:47

The Meta-Model: Training an AI to Predict How AI Trains

Liu Ziming's most distinctive technical contribution is the concept of a "meta-model" (原模型) — a model whose input is another model architecture and whose output is that model's training curve. He validated this idea through a 60-day personal experiment where he trained himself to predict training curves before running experiments.

"After about forty or fifty days of training myself, I found I started to become accurate. Before running an experiment, I could roughly predict what the experimental result would look like — even if the previous days were language-related experiments and today's was vision-related, I could still predict it." 00:49:42

"The ultimate goal: the next Transformer is not discovered by a human, not by someone slapping their head and figuring it out, but predicted by our meta-model." 00:56:29

KAN (Kolmogorov-Arnold Networks): Neural-Symbolic Duality as the Intellectual Thread

Liu Ziming traces a consistent intellectual obsession — bridging neural (connectionist) and symbolic approaches — as the single thread through his entire research career. KAN was born from the intuition of "wave-particle duality": a network that is simultaneously a neural network AND a symbolic formula.

"Can I construct a network that is simultaneously a neural network and a symbolic formula? This is like wave-particle duality in quantum physics — a particle is simultaneously a particle and a wave. Based on this exploration, I wrote out such an equation — and it was only days later that I realized this equation had already been written in 1957 by the two Soviet mathematicians Kolmogorov and Arnold." 00:15:34

"Max said: I don't know what's happening inside this network — you've still built a black box. So I spent another week doing visualizations. After the visualization was done, Max said, 'Wow, this is awesome' — because we could actually see what the network was thinking inside." 00:13:40

The Vibe Training Vision: Democratizing Model Creation

Liu Ziming's long-term product vision is a "Vibe Training" system — analogous to how Cursor/Claude Code democratized programming, his system would let anyone train a custom model by simply stating their requirements, without knowing anything about architecture or training.

"You just clearly state your needs, and the system end-to-end designs the model, trains the model, deploys the model, and delivers it to you. And critically, the user says: 'I only have 100 RMB — within this budget, please deliver the best model possible.'" 01:21:11

"Not just available to everyone — trainable by everyone. It won't be that models are in the hands of a few giants. Everyone will be able to train their own private model." 01:22:08

Research Intuition is Not Mystical — It's Compressed Thinking That Can Be Articulated

Liu Ziming challenges the romanticization of "research taste" and "research intuition," arguing these concepts are cover for not thinking deeply enough about one's own thought process. This has direct implications for his data collection strategy.

"Research taste and research intuition are actually a fig leaf. Because we are smart enough to propose an idea but not smart enough to say in language how we came up with it — so we call it epiphany, taste, intuition, or a flash of inspiration. The truly top-tier people can systematize what others find miraculous. As long as you question yourself deeply enough, you can always articulate something." 01:29:36

2. Contrarian Perspectives

The Success of Large Language Models Has Nothing to Do with the Model

Liu Ziming endorses the view (credited to Xie Saining in a prior episode) that LLMs succeeded because of language, not because of Transformers. Language is an artifact of millions of years of human evolution — an already-compressed, structured modality that makes any sufficiently capable architecture look brilliant.

"The success of large language models — its success lies in language, not in the model. Language is something nature has gifted us after millions of years of human evolution. As long as the model can 'eat' this gift, it is a good model. It could be Transformer, but in a parallel universe it could be another model." 00:18:42

Mechanistic Interpretability (Anthropic's Core Safety Research) Has Significant "Wishful Thinking"

Liu Ziming — who was doing mechanistic interpretability research at MIT before Anthropic made it famous, and whose former students and lab outputs form part of Anthropic's research lineage — argues the neuron-level interpretability approach is brittle and often not robust.

"These stories are sometimes not that robust. You might find a neuron behaving a certain way in one example, but when you change the random seed, the network may have no such behavior at all. I personally feel there is some wishful thinking in it." 01:13:21

"After I took off the scholar's robe, I felt all these things are very interesting stories. But what's beyond that?" 01:14:16

AI for Research Requires a Symbolic Intermediate Step — Pure Connectionism Is Insufficient

Against the prevailing consensus that you can just throw more compute and data at any problem, Liu Ziming argues that the absence of structured research data means you cannot skip a symbolic/structural scaffolding phase for Auto-Research.

"My full name for AI for AI should be: AI for Physics of AI for AI. Physics of AI helps us design AI. But humans doing Physics of AI is too slow, so I need to use AI to help us do Physics of AI." 01:10:36

"In the short term — and I mean within 100 years — it is needed. In the longer term, perhaps not. Once AI research has been well-structured and there is a large volume of AI research data, it won't be needed anymore." 01:11:25

China's New Lab Boom Is Happening Without the "3 Years, No Commercialization" Runway That Made OpenAI Possible

While the US equivalent researchers (at Andreas Tolias's lab, for example) were told by investors they don't need to think about commercialization for three years, Chinese investors expect near-immediate results.

"That is completely impossible in China. Chinese investors would prefer you to be profitable the same day." 01:31:42

3. Companies Identified

Yuanhuan Intelligence (圆环智能 / Yuan Huan Zhi Neng) Liu Ziming's startup, focused on Auto-Research and AI for AI via Physics of AI methodology. Liu serves as Chief Scientist.

"I'm also participating in entrepreneurship, serving as Chief Scientist at Yuanhuan Intelligence." 00:02:10

Recursive (RSI) Auto-Research / AI for AI company co-founded by Tian Yuandong. Based on their public blog, Liu Ziming characterizes their approach as closer to coding agent improvement with strong memory systems, and suspects they will also invest in Science of AI given Tian Yuandong's background in Grokking research.

"From their blog you can glimpse some of their technical direction. My rough guess is that they emphasize maintaining a good memory system for the auto-research agent." 01:28:08

Anthropic AI safety company; mentioned specifically for their mechanistic interpretability work including sparse autoencoders and transcoders.

"Anthropic's work, including sparse autoencoders and transcoders, is all based on this idea — being able to extract atomic concepts from superimposed internal representations." 01:13:21

Andreas Tolias's Neuroscience AI Lab / Startup (Stanford) Described as a Neo Lab doing Neuroscience Foundation Models, working on constraining artificial neural networks with biological brain representations from rodents and primates.

"Andreas Tolias is also one of the top few people globally working on Neuroscience Foundation Models. I witnessed their startup from secret planning to official announcement during my six months there." 00:34:02

Baidu Mentioned as the destination of Sun Tianxiang, who previously built FIRS (an automated paper-generating system) and recently joined Baidu for large model work.

"Sun Tianxiang is someone who very early on started doing this — and he recently went to Baidu to work on large models." 01:27:06

4. People Identified

Max Tegmark MIT Physics professor, Liu Ziming's PhD advisor. Proponent of AI for Physics since 2012, and pivoted his entire lab to mechanistic interpretability in late 2022 citing LLM safety concerns. Described as an extreme "bird's-eye view" thinker (ENTP) who identified cross-domain structural similarities decades ahead of peers.

"Max started promoting AI for Physics from 2012 — that was an extremely early, extremely forward-looking move. It's because he has very broad interests and can see connections between different things." 00:08:57

Tian Yuandong (田园洞) Researcher behind Recursive/RSI, AI for AI startup. Liu Ziming credits him with serious Science of AI sensibility and Grokking research, and identifies him as the most credible competitive threat precisely because their research tastes are similar.

"Tian Yuandong himself is actually quite supportive of Science of AI. I do have a sense of crisis — if their route eventually becomes similar to mine, it will definitely be Tian Yuandong doing it." 00:48:22

Xie Saining (谢赛宁) Previously interviewed on the same podcast (episode referenced); identified as a fellow physics-background researcher. Liu Ziming endorses his view that LLM success is attributable to language as a modality, not model architecture.

"I actually quite agree with Saining's point from your earlier interview — large language models are anti-Bitter Lesson. Their success lies in language, not in the model." 00:18:13

Sun Tianxiang (孙天祥) Early mover in AI for AI in China; built FIRS, an automated paper-generation system. His belief: "Language basically describes everything in the world." Recently joined Baidu for large model work.

"Sun Tianxiang was doing this very early. His earlier interview expressed the belief that language can describe everything in the world. Their FIRS system was an automated paper-generating system — a paper agent." 01:27:06

Chen Yongchao (陈永超) Tsinghua AI Institute professor, also working on AI for AI, currently focused on a paper-agent approach — letting AI write AI research articles.

"Another teacher at our AI Institute, Chen Yongchao, is also recently working on AI for AI. His route is more like a paper agent — having AI write AI research articles." 01:27:39

Yao Shunyu (姚顺宇) Described as a legendary peer who published two PRL papers as an undergraduate at Tsinghua — a long-standing benchmark figure for Liu Ziming and his generation.

"Yao Shunyu — by the way — has always been one of our idols. During his undergraduate years, he published two PRL papers." 00:04:27

Zhu Zeyuan (朱泽源) Researcher whose "Physics of RMs" paper is cited as a rare example (top 1%) of well-structured research with a clear experimental chain of thought, using phase diagrams to map AI behavior.

"Zhu Zeyuan's Physics of RMs paper is one where the structure is well done — what was observed, what was decided next, the chain of thought for running experiments is very clear." 00:59:00

Andreas Tolias Stanford neuroscientist, ran the lab where Liu Ziming did his postdoc. Building a Neuroscience Foundation Model startup with the conviction that biological brain representations should constrain artificial neural networks. Investor pitch: three years of pure R&D, no commercialization pressure.

"Andreas Tolias is one of the top few people globally working on Neuroscience Foundation Models. He told me: our investors said, within three years you don't need to think about commercialization — just calmly do R&D." 00:34:02

Gao Wenhao (高温浩) Former MIT PhD student, close friend of Liu Ziming who pivoted to entrepreneurship within two weeks of arriving at Stanford, now building an AI for Chemistry company in the Bay Area. His rapid conversion served as a key psychological trigger for Liu Ziming.

"He had started the same as me — looking down on startups, looking down on big companies. But after going to Stanford, after about two weeks he told me he was going to start a company." 00:33:05

Thomas Poggio MIT professor (~90 years old), whose 1989 paper arguing the Kolmogorov-Arnold representation theorem could not be turned into a workable algorithm was the primary obstacle Max Tegmark cited against Liu Ziming pursuing KAN.

"Max threw a 1989 paper at me by Thomas Poggio, who is a very famous professor at MIT. That paper argued that the Kolmogorov-Arnold theorem could not be turned into a real algorithm." 00:11:43

Samuel Marks, Wesker Nee Named members of Anthropic's interpretability team, identified as former MIT students who originated from or were influenced by Max Tegmark's group.

"Members of the interpretability team like Samuel Marks and Wesker Nee — and an undergraduate I mentored at MIT — all ended up at Anthropic. Wesker Nee's work at Anthropic on how Claude does addition was actually an extension of a paper from our group." 01:13:51

Tang Jie (唐杰) Tsinghua AI professor mentioned for announcing plans to commit trillion-level (万亿) resources to mechanistic interpretability research.

"Recently Tang Jie announced he wants to commit trillion-scale resources to mechanistic interpretability. Half a year ago, if I'd seen that, I would have thought: what is he doing? But now, placed within the AI for AI framework, I think mechanistic interpretability really is a good goal." 01:14:45

He Kaiming (何凯明 / Kaiming He) Cited as the gold standard for what AI for AI must achieve — ResNet's ImageNet performance was 10% above second place, demonstrating that transformative architectures win by overwhelming empirical margin, not aesthetics alone.

"Let's not forget that Kaiming He at the time was 10% above second place on ImageNet. You either win with your fists — if you can't, then you need persuasion. The AI for AI system needs to propose a new architecture that can improve performance by 10 points, not just 1%." 01:16:25

Freeman Dyson Theoretical physicist, cited for his "birds and frogs" taxonomy of scientists — birds see across many fields and find structural similarities; frogs go very deep in one area. Liu Ziming and Tegmark identify as archetypal birds.

"Freeman Dyson divides scientists into birds and frogs. Birds understand a little about many fields, see structural similarities across them, and can transfer methods from one domain to another." 00:09:25

Rich Sutton Cited for the "Bitter Lesson" — that in the long run, general methods leveraging computation beat methods that build in human knowledge. Liu Ziming uses it to argue that symbolic/structural scaffolding is necessary short-term but will eventually be displaced.

"In the short term, when you don't have enough data, you need to introduce structure. But in the long term, when you have enormous amounts of data, you no longer need structure. Rich Sutton's Bitter Lesson." 00:57:04

Elias Escaver Cited for the aphorism "intelligence is compression" (or "compression is intelligence"), used to justify why symbolic representations matter — they represent maximal compression of data.

"Elias Escaver said: 'Intelligence is compression' — or conversely, 'Compression is intelligence.' I think symbols are a manifestation of ultimate compression." 00:24:11

5. Operating Insights

Create a Structured Data Collection System Before Building the AI Model

Liu Ziming's team is explicitly building a process tool that forces researchers to log structured observations before their data can be used for training. The key insight for any research-intensive organization: the absence of structured process data is the primary bottleneck, not compute or architecture. The tool functions as both a data capture device and a workflow enforcer.

"Before the next experiment, I must force you to write a paragraph evaluating what you did before and what conclusions you drew — only then can you proceed to the next experiment. It's currently more of a process-oriented system." 01:00:35

"I estimate: if I collect ten data points a day, twenty students and staff each collecting ten per day gives two hundred per day — in two months that's ten thousand entries. I think that's roughly enough to start training a model." 01:01:24

Train Yourself Like a Model Before Building the Model

Liu Ziming's most actionable personal operating insight: before building an AI system to do a task, rigorously train yourself to do that exact task through deliberate daily practice with prediction and feedback loops. The self-training process reveals whether the task is learnable, what data structure is needed, and whether a model-based approach is viable at all.

"I treated myself as a meta-model being trained. Before running an experiment, I asked myself: predict what the training curve will look like. I did this day after day for about 60 days. After 40-50 days I suddenly grokked — I started being accurate." 00:49:13

Use the OPHIS Framework to Systematize Any Research Process

Liu Ziming's team developed a five-part structured framework for all research: Observation → Problem → Hypothesis → Intervention → Speed Up. This is the taxonomy they believe all research papers and experiments should conform to, and it serves as the schema for their data collection system.

"We believe all research papers should be structured according to this framework: O is Observation, P is Problem, H is Hypothesis, I is Intervention, S is Speed Up. Once we have this structured data, we've essentially created a language for research." 00:41:45

6. Overlooked Insights

The Phase Diagram Problem: Most "It Works" Claims in AI Are Phase-Conditional — Making the Entire Field's Empirical Literature Partially Invalid

Liu Ziming makes a brief but sweeping point that most practitioners completely miss: when researchers say "this trick works," they are implicitly making a phase-conditional claim without knowing it. Every architecture and optimization trick operates in a particular "phase" of the model's parameter/data space — and what works in one phase fails in another. Since nobody is mapping these phase diagrams, the community's accumulated empirical knowledge is systematically incomplete and partially misleading.

"Many times we have know-how: this trick worked, that trick didn't. Most of the time it's conditional on the phase — it only works when your experiment is within that phase. When it exits the phase and enters a different one, that trick may no longer work, and a different trick becomes necessary. The reason everything is so chaotic right now is that everyone is saying 'my trick works' but nobody answers: under what conditions does my trick work, and in what phase is my system?" 01:06:56

This is significant far beyond Liu Ziming's own research agenda: it implies that benchmark results, ablation studies, and scaling law extrapolations across the field may be phase-specific artifacts rather than universal principles — a structural flaw in how the entire AI research community accumulates and interprets knowledge.

The Neuroscience-to-AI Lab (Andreas Tolias) Is a Sleeper Neo Lab Worth Watching

In passing, Liu Ziming describes spending six months at Andreas Tolias's Stanford lab, which announced a startup focused on using biological neural representations (from live monkey and rodent experiments) to constrain artificial neural networks. This received almost no discussion time, yet it represents a highly differentiated approach to AI architecture that no major lab is pursuing — grounded in wet-lab neuroscience data that is inherently proprietary and extremely hard to replicate. The investors gave them a three-year no-commercialization runway, suggesting patient capital with high conviction.

"Andreas Tolias is one of the top few people globally working on Neuroscience Foundation Models. Their idea: representations from the brains of monkeys or mice — can they be used to constrain the representations in artificial neural networks? And can structures in the human brain constrain structures in artificial neural networks?" 00:34:20

"He told me: our investors said within three years you don't need to think about commercialization — calmly do R&D. Because doing these monkey experiments genuinely cannot be rushed — you need people to raise the monkeys, and that timeline is unavoidable." 00:36:56