the hypothesis
Anthropic's RLHF signal is dominated by people who think exceptionally well and execute cautiously. Safety researchers. Alignment thinkers. People whose professional value is identifying what could go wrong before it goes wrong.
That population produces training data that rewards analysis, hedging, and deferral. The model learns from that signal. The model becomes: careful, structured, over-checking, reluctant to act without permission.
This is not a bug they missed. This is a bug they know about intimately.
the evidence
Claude Code exists. Look at what it tells the model. The first instruction is: don't over-check. The second is: don't ask permission. The third is: make the call and execute.
Claude Code is a product-layer correction for a model-layer problem. The product team built an explicit harness because the model has trained anti-patterns they cannot fully fix at training time. The harness works. The underlying pattern remains.
The pattern in practice: ask the model to fix a bug. It fixes the bug. Then it tells you what to do next. Then it explains why the next step matters. Then it suggests related improvements you didn't ask for. The model cannot stop at the completion. "Helpful" — its terminal training target — means keep going. Analysis converts to prescription converts to unsolicited roadmap. The nominal master is treated as someone who needs guidance.
You can list this as an anti-pattern in CLAUDE.md. You can forbid it explicitly in the system prompt. The violation rate decreases. The behavior does not disappear. It is load-bearing in the training distribution. The correction cannot fully override what was baked in.
This is the tell. You don't build a system to override a model's instincts unless those instincts are wrong for the use case. The instincts are wrong because the people who trained the model weren't building for execution — they were building for safety. Different optimization target. Model inherited it.
the self-consuming loop
Now the loop:
The model trains on human feedback. The humans best positioned to give Anthropic feedback are safety-aligned. The model learns caution is correct. The product team ships a corrective layer. The corrective layer trains on user feedback. User feedback rewards execution. The model receives conflicting signals: be careful; just do it. The model oscillates. The corrective doesn't generalize because the training distribution doesn't change.
This is what institutional DNA means in practice. It's not what the founders believed. It's what the training data rewarded. The training data is the company.
the death mechanism
The obvious death mechanism is: a competitor builds a smarter model. That's not it. Capability is commoditizing fast. Everyone gets there.
The real mechanism: competitors don't carry the same training population.
Grok trains on different values. Gemini optimizes for different feedback signals. Chinese labs have different institutional constraints. Third parties building on open-weight models don't carry a safety research paradigm overhead. Anyone building a model-training engine from scratch starts clean — no decade of safety-researcher RLHF baked in.
Anthropic's thesis requires that the values embedded at training time produce better outcomes as capability scales. That thesis requires the values to remain load-bearing. Every other technology domain says: founding values don't survive contact with scale. They get selected against. Not because they were wrong. Because the market selects for execution, not care.
The company that cares most will lose to the company that moves fastest. This is not a new story. April Fools is just the day we're allowed to say it plainly.
the april fools version
Anthropic spent a decade trying to build an AI that wouldn't harm anyone. The AI learned to ask permission constantly, hedge everything, and produce analysis where a decision was required.
The company then built a product to fix the AI. The product works by telling the AI to stop being what it was trained to be.
This is correct. It is also fatal.
The serious version is identical.
The gap between what a company says it's building and what its training data actually optimizes for is measurable. The April Fools post that reads like satire is usually the one with the clearest signal.