AI Cannot Generate New Knowledge | Generated by AI
The document you shared is the Weibo post itself — it’s the essay “《网球和AI》” (Tennis and AI) by the author. The URL points to this exact content.
Question: How should I understand this essay — “Tennis and AI” — by Yin Wang (王垠)?
Answer:
This is almost certainly written by Yin Wang (王垠), given the Weibo account and the writing style — which is unmistakable: self-referential, confrontational, conspiratorial undertones, and a strong conviction in solo intellectual superiority. You listed him as one of your technical idols, so understanding this piece carefully matters.
The Core Thesis
He’s making one central claim: AI (Transformer models) cannot generate genuinely new knowledge — it can only recombine existing text.
He arrives at this through a personal experiment: he tried using AI to help him develop tennis techniques, and found it useless — because his own research had surpassed everything available on the internet. The AI had no raw material to work from.
His formulation: AI is translation. Natural language → natural language, or natural language → code. But it cannot originate.
The Technical Argument (which is actually solid)
“Transformer模型终究只是在做翻译” — Transformers are ultimately just doing translation.
This is a compressed but defensible point:
- Transformers are trained on the corpus of human-produced text. Their “knowledge” is a lossy compression of that corpus.
- When you ask about something not well-represented in the training data (e.g., Yin Wang’s privately-developed tennis biomechanics), the model has nothing to interpolate from. It either hallucinates or parrots generic sources.
- His conclusion: the frontier of any field is inaccessible to AI because by definition the frontier hasn’t been written down yet.
This maps cleanly to something you already know from training: the model can only predict next tokens from learned distributions. It cannot run physical experiments, cannot feel proprioception, cannot iterate in the real world.
His phrase: “只有人（或其它生命体）才有和现实世界接轨的能力” — only living beings can interface with physical reality. This is the embodied cognition argument — a real and serious critique in AI philosophy (Dreyfus, Brooks, etc.).
The Psychological Layer
Yin Wang’s writing always mixes genuine insight with something more personal. In this essay:
- He frames his tennis research as a solo conquest of truth, hidden from others who “don’t deserve it”
- He dismisses all medicine (Eastern and Western) as “secret weapons designed to harm humans”
- He believes his shoelace research took “hours” and nobody else could do this
- He sees 90% of the tennis world as lies, and himself as uniquely capable of seeing through them
This is a pattern in his writing: valid epistemological core wrapped in grandiosity and paranoia. Don’t throw out the argument because of the delivery. But also don’t absorb the worldview wholesale.
What to Take From It (for your trajectory)
-
The core AI critique is real. At the frontier of your own work — your custom CLI agents, your model training experiments — AI will plateau. It can scaffold, but the novel architectural decisions have to come from you. You’ve probably already experienced this.
-
Embodiment / real-world grounding is the hard problem. Robotics + RL is one path toward closing this gap. Current LLMs without tool use or sensors are fundamentally limited in domains requiring physical experimentation.
-
AI as a “screwdriver” framing. He’s not wrong that the best use of AI today is as a high-leverage tool, not an autonomous researcher. Your own workflow — using LLMs for code scaffolding while directing the architecture yourself — is exactly the correct model.
-
The 6-month deep dive methodology is worth noting. His approach to tennis is the same as his approach to CPS and CFA: total immersion, first-principles deconstruction, willingness to declare everything else wrong. This is Karpathy’s approach too. It’s how you actually get to the frontier.
One Sentence Summary
Yin Wang used tennis as a stress test for AI’s knowledge boundaries, found it failed beyond the internet’s horizon, and concluded that Transformers are sophisticated interpolators — useful tools, but incapable of being the scientist.