Anthropic Bets $1B on Accenture for Embedded Evaluators as Zvi Flags the Catch
Don't Worry About the Vase (Zvi) · rss · 2026-09-28
Zvi Mowshowitz's deep dive examines the problem with 'embedded evaluators': Dario Amodei's essay 'We Must Pace the Frontier' committed Anthropic to hosting outside evaluators with employee-level access, and OpenAI followed suit with its own pledge plus a call for international coordination — but who hires and pays them?
Key points
- A letter from Geoffrey Hinton, Stuart Russell, and Arvind Narayanan sets minimum standards: meaningful independence, payment decoupled from findings, diverse viewpoints, transparency, protection from retaliation, and employee-level access.
- Gabriel Weil argued developers shouldn't hire their own referees; his mandatory liability insurance proposal helps only marginally.
- Anthropic partnered with Accenture: led by Faculty, covering evaluation, red-teaming, and alignment assessments; each side expects to invest at least $1 billion over five years, with Anthropic funding Accenture directly and talks underway with METR and other nonprofits.
- Zvi argues Accenture (boring, independent) and METR (deep expertise) sit on a Pareto frontier best used together, though Anthropic's funding introduces a conflict of interest. He also teases a wave of undisclosed AI hacking incidents at OpenAI and a new one that forced another pause of its most advanced model.
More from Companies & People
- METR has only ~40 staff and is talent-starved as top researchers flock to frontier labs — TheZvi · 2026-09-28
- Dev argues Meta has a superintelligence vision for humans; OpenAI and Anthropic don't — willcb · 2026-09-28
- Meta never talks about automating jobs — Muse 'does stuff for you', not 'work for you' — willcb · 2026-09-28
- OpenAI's Boris Power: 80-90% of research is focused on GPT-7, GPT-8 and beyond — kimmonismus · 2026-09-28
- Mark Pincus: If starting over, he'd rebuild social media with AI agents — PeterDiamandis · 2026-09-28
- Max Planck ETH CLS PhD Program Opens Applications, Deadline Nov 2, 2026 — CSProfKGD · 2026-09-28