# Dario Amodei

> 1983– · AI Researcher, Co-founder of Anthropic
>
> **Recorded contribution:** Co-founded Anthropic; Claude; AI safety research; scaling laws; RLHF

## How to use this dossier

Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory.

## 1. Historical orientation

Physicist-turned-AI researcher Dario Amodei worked at Google and OpenAI on model scaling and safety before co-founding Anthropic in 2021, where he became CEO. His co-authored research includes empirical scaling laws and safety-oriented work, while Anthropic developed the Claude model family and Constitutional AI. The durable contribution is research and institution-building; corporate roles and model claims remain time-sensitive.

## 2. The problem inherited

As language-model capability rose with compute and data, developers lacked reliable forecasts, evaluations, and governance controls proportionate to possible misuse and catastrophic failure.

## 3. The central contribution

Amodei co-authored influential neural scaling work and co-founded Anthropic around frontier-model development and safety-oriented training and deployment practices.

## 4. Reconstruct the mechanism

1. Measure loss and capability across controlled changes in model size, data, and compute.
2. Fit empirical trends and use them to allocate a training budget.
3. Train a general model, then shape behavior with human or model feedback and explicit principles.
4. Gate deployment with evaluations, monitoring, security controls, and escalating safeguards tied to measured capability.

## 5. What changed downstream

- Scaling laws made parts of model development more forecast-driven.
- Anthropic became a major frontier lab and influenced debate over responsible-scaling policies and model governance.

## 6. Attribution, limits, and uncertainty

- Scaling and alignment papers are multi-author work; Anthropic's models depend on large teams, data suppliers, annotators, compute vendors, and users.
- Empirical laws can change outside measured regimes, and company-authored safety claims require independent scrutiny; feedback training does not guarantee truth or alignment.

## 7. Reconstruction lab

Train the same small model at three parameter/data budgets, fit a loss trend, and test one extrapolation. Add a behavior rubric and document where model feedback could hide correlated error. Hold training compute fixed while varying parameters and data, then distinguish an empirical scaling fit from a universal law. Test a capability metric that changes discontinuously even as loss improves smoothly. For model feedback, create a rubric error shared by generator and evaluator and show how iteration reinforces it. This connects scaling and safety: predictable aggregate loss can guide resource allocation, but it does not establish that specific behaviors, values, or hazards improve predictably.

## 8. Evidence trail

- [Scaling Laws for Neural Language Models](https://arxiv.org/abs/2001.08361) — arXiv
- [Constitutional AI: Harmlessness from AI Feedback](https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback) — Anthropic
- [Dario Amodei](https://en.wikipedia.org/wiki/Dario_Amodei) — Wikipedia contributors

---

*Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*