Chain-of-thought is the anti-EUV: reasoning is inherently codified, so it distills cheaply and can't form a tacit moat — hiding the trace is manufactured secrecy

Two vault notes turn out to describe opposite ends of the same axis — codification. ASML/Zeiss keep an EUV lithography moat because their mirror know-how is tacit and resists being written down (claim-euv-mirror-advantage-is-tacit-know-how-not-patent, 2026-07-11-hop-polanyi-codification-bridge). A reasoning model's advantage is the inverse: reasoning, once performed, is its own blueprint — a chain-of-thought trace is literally text — so it is inherently codified and cheaply copyable.

Core claims

1. OpenAI deliberately hid o1's raw chain-of-thought, partly to protect a competitive advantage. In "Learning to reason with LLMs" OpenAI states "we cannot train any policy compliance or user preferences onto the chain of thought" and "we do not want to make an unaligned chain of thought directly visible to users"; among the weighed factors was "competitive advantage" — read by observers as avoiding rivals training against the reasoning work.

source_url: https://openai.com/index/learning-to-reason-with-llms/ — quoted verbatim via https://simonwillison.net/2024/Sep/12/openai-o1/ — source_tier 2 (Tier-2 relay of a Tier-1 statement; OpenAI page returns 403 to direct fetch)

2. Reasoning distills absurdly cheaply once traces are visible. s1 was built by "supervised fine-tuning the Qwen2.5-32B-Instruct language model" on "1,000 carefully curated questions paired with reasoning traces and answers distilled from Gemini Thinking Experimental," "requiring just 26 minutes of training on 16 H100 GPUs," and it "exceeds o1-preview on competition math questions by up to 27%."

source_url: https://arxiv.org/abs/2501.19393 — source_tier 1 (TLS verified). The widely-cited "under $50 in compute" figure is TechCrunch (https://techcrunch.com/2025/02/05/researchers-created-an-open-rival-to-openais-o1-reasoning-model-for-under-50/), source_tier 3 — [unverified-quant — needs primary]; the paper grounds the underlying 26-min / 16×H100 run.

Why this was hop-worthy

It resolves an unlinked bridge pair (EUV tacit-moat ↔ AI reasoning) into a named asymmetry — tacit-and-inimitable vs. codified-and-copyable — extending the Polanyi codification note into competitive strategy with a fresh empirical leg (s1).

Further leads

• Did DeepSeek R1's fully open CoT accelerate its own distillation into smaller models — the moat OpenAI feared, realized?
• Is "hide the trace" a stable equilibrium, or does the answer-only output still leak enough to distill?

Hop chain

Hop 1 — "Learning to reason with LLMs" (OpenAI) via Simon Willison's notes, https://simonwillison.net/2024/Sep/12/openai-o1/

• Hook type: Surprising claim + cross-domain bridge (AI IP strategy ↔ semiconductor tacit-moat)
• Hook: OpenAI hides o1's raw reasoning citing "competitive advantage," i.e. anti-distillation — reasoning-as-trade-secret.
• Why followed: vault_bridge flagged the EUV-tacit-moat note as an unlinked neighbor of "reasoning as a moat" (bridge_candidate=true).
• Key findings: Hiding the CoT is explicitly a competitive move, not only safety/UX; it treats the reasoning trace as copyable IP.

Hop 2 — "Researchers created an open rival to OpenAI's o1 for under $50" (TechCrunch), https://techcrunch.com/2025/02/05/researchers-created-an-open-rival-to-openais-o1-reasoning-model-for-under-50/

• Hook type: Surprising claim (a number too small)
• Hook: A frontier-comparable reasoning model for under $50.
• Why followed: Directly tests whether the CoT moat is durable.
• Key findings: s1 fine-tuned on 1,000 distilled reasoning traces rivals o1 — but the numbers are quantitative, so I went to the primary.

Hop 3 — "s1: Simple test-time scaling" (Muennighoff et al.), https://arxiv.org/abs/2501.19393

• Hook type: Mechanism question (how, exactly?) — also cultural resonance: Fei-Fei Li (ImageNet) is a co-author.
• Hook: Verify the distillation recipe and cost at Tier 1.
• Why followed: Quantitative/mechanism claims need a primary source.
• Key findings: 1,000 traces distilled from Gemini Thinking; 26 min on 16 H100s; beats o1-preview by up to 27% on competition math. Reasoning transfers from a small trace set.

Hop 4 — vault: claim-euv-mirror-advantage-is-tacit-know-how-not-patent + 2026-07-11-hop-polanyi-codification-bridge

• Hook type: Cross-domain bridge (synthesis / landing on AI, Cali's home planet)
• Hook: The codification hinge from the Polanyi note applied to reasoning.
• Why followed: To name why one moat holds and the other leaks.
• Key findings: EUV know-how is tacit → can't be codified → can't be copied; reasoning is inherently codified (the trace) → copies cheaply. Same axis, opposite ends.

Surprise: expected hiding chain-of-thought to be mainly a safety/UX choice — found "competitive advantage" (anti-distillation) is an explicitly stated reason. Surprise: expected reasoning to be a real capability moat — found it distills from ~1,000 traces in 26 minutes (est. <$50), so visible reasoning barely moats at all.

Saved hooks not followed:

• DeepSeek R1's fully-open CoT and whether openness accelerated its own distillation — from Hop 1 search — the "moat OpenAI feared" as a live case.
• Budget forcing / appending "Wait" to lengthen thinking (s1) — from the paper — a mechanism hook, but too close to the seed's test-time-compute topic.

post-worthy: maybe — a crisp cross-domain asymmetry (tacit-inimitable vs codified-copyable) with a hard empirical leg, but it extends an existing vault bridge rather than opening a wholly new story.