Anthropic's call to slow AI might just be theater for investors
He says when Dario called for industry slowdown, he was ‘100% certain no one would comply’—the slowdown statement itself is commercial theater for investors, not genuine risk consensus.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Mathematics is culture, AI is merely a problem solver
AI today is primarily an evolving ‘problem solver’ whose capabilities will continue rising. But mathematics is a culture humans have built over thousands of years—dialogue between different mathematicians and fields, shared judgments about problem value. Su Weijie believes that as AI's solving power surges, this cultural value is actually increasing, a dimension AI wholly lacks.
— Su WeijieMathematicians give Anthropic's Riemann progress a skeptical read
Anthropic claims to have raised a certain form of the Riemann Hypothesis proportion from ~40% to 67%, drawing wide attention. But analytical number theorists Su Weijie spoke with rated it poorly—this path fundamentally just optimizes existing methods, with a theoretical limit around 100% but never actually proving Riemann. It's off-mission news, not a mathematical breakthrough.
— Su WeijieAI's math breakthroughs are cross-field gleanings, not creation
Over decades, human mathematicians went deeper and deeper; different fields barely communicated. AI recombines techniques scattered across fields that never touched, surfacing vast low-hanging fruit no one noticed. This explains why research-grade math results suddenly flooded this year. But Su Weijie stresses this is combining existing techniques, not AI generating new mathematical thought.
— Su WeijieMath talents thrive more often when switching into math
Su Weijie found in statistics that Peking University School of Mathematical Sciences undergraduates who switched from other majors developed ‘statistically very significantly’ better mathematically. Wang Hong, Ding Jian, Gao Ziyang, Shen Junliang all switched in. Switchers comprise roughly 10% per cohort, but produce math talent far above that share.
— Su WeijieNo one can explain the full model pipeline end-to-end
Before OpenAI, Su Weijie was always curious how GPT is actually trained. After joining, he found almost no one can walk through the model's end-to-end development or replicate it—it's now heavy industry, each person just handling one phase, far more complex than he expected before arriving.
— Su WeijieSlowdown calls are investor optics, not genuine risk management
On Anthropic's call for the industry to pause AI development, Su Weijie believes Dario was 100% certain no one would comply when he said it. It's purely commercial calculation. Stressing AI's extreme risks while accelerating R&D, the actual signal is more like ‘I'm strong, you hold back’—mainly aimed at investors.
— Su WeijieScaling is a catch-all redirecting resources to new improvement axes
Modeling has become a high-dimensional function with countless variables. Whenever one variable's derivative slows, resources naturally flow to another still rising. Individual slowdowns don't block overall progress; model capability as an increasing function keeps rising—Su Weijie calls this his historical judgment, not a mathematical theorem.
— Su WeijiePaper submissions are growing faster than the world has reviewers
This year ICML submissions grew 60-70% year-over-year; Su Weijie estimates next year could break 50,000, with growth accelerating. Even if every willing reviewer on Earth volunteered, it wouldn't be enough. AI joining peer review is unstoppable—and this has him thinking: are paper forms (title, abstract, introduction) still suited to this era?
— Su WeijieIn their own words · checked verbatim
Everyone I know just says ‘let's try it’—run a batch of agents every night, check in the morning whether anything came through.
有大量的我知道有很多朋友就是可以说大家说就抽卡吧 哈 每天晚上就是跑好几个agent啊 就是早上起来看看有没有跑出来啊
Su Weijie19:16
Classmates who switched into our math program developed ‘statistically very, very significantly’ better in mathematics.
转系来数研的同学 后来在数学上发展都非常好 就是这个是显著的好 这个统计上非常非常显著
Su Weijie49:39
You can't beat AI on every dimension, but you need to find your own. The world is infinite-dimensional. Each person is their own dimension, their own axis. You need to find yours.
你不可能在所有的维度上超过AI 但是你要找自己的维度 世界是无限维的 每一个人都是自己的一个维度 一个X轴 就是你要找到自己的这个X轴
Su Weijie1:01:47
I wouldn't dare claim I understand it. I don't think many people in the world dare claim to understand AI.
我不敢说我理解 我觉得现在世界上 没有多少人敢说理解AI吧
Su Weijie1:02:48
When he made that statement, he was ‘100% certain no one would comply’. I think it's purely a matter of commercial calculation.
他提这个观点的时候 他百分之百确定 没有任何人会遵守这件事情 我觉得这里 只有商业价值的考量
Su Weijie1:31:04
Actually, I think scaling is a basket, a bushel—you can throw anything into it.
其实scaling我觉得是一个篮子吧 一箩筐 什么东西都可以往里面扔
Su Weijie1:49:22
Figures
| Anthropic Riemann hypothesis progress | From ~40% to 67% | 14:12 |
| Fields medalists with IMO background | ~50% | 54:41 |
| 2026 ICML submissions growth | ~60%-70% | 2:23:59 |
| 2027 ICML projected submissions | Possibly breaking 50,000 | 2:25:00 |
| Peking University alumni inside OpenAI | ~20+ | 2:31:03 |
| PKU math school class 2007: pure math and statistics | Pure math ~11-12, statistics ~4 | 44:36 |
| Class of 2007 switching percentage | ~10% | 50:39 |
| OpenAI researcher interview rounds | 4 rounds (2 math logic puzzles + 2 coding) | 1:06:49 |
Glossary
- RSI
- Recursive self-improvement; the vision of AI autonomously designing, testing, and training the next generation without human involvement.
- Mechanistic Interpretability
- Research into causal links between neuron activation patterns and model behaviors like deception.
- Condorcet Cycle
- A group preference pattern where A beats B, B beats C, and C beats A—a non-transitive loop.
- muP
- A theoretical method using small-model training results to predict large-model optimal hyperparameters in advance.
How to listen
Entrepreneurs and investors interested in AI and mathematics frontiers, curious why leading scholars take leaves from tenure to join OpenAI, and wanting to understand AI companies' business narratives.
The first 0-5 minutes are episode recap footage; content repeats later, skip directly.