AEF-1 Standard Emerges for Third Party Evaluators; XAI, OpenAI, and Anthropic Cosign
The AI Evaluator Forum (AEF), formed in December 2025, has published AEF-1, a proposed baseline for independent third-party AI evaluations. This standard covers access, conflicts of interest, funding, recusal, and transparency. Anthropic is unilaterally committing to embedded third-party evaluators, with Dario Amodei outlining plans for employee-like access to verify safety practices and assess training pipelines. A significant debate has emerged regarding whether labs should 'pace capability progress' or focus on 'specific mitigations and containment,' with figures like Bilal Chughtai advocating for pacing and others emphasizing control and governance. The newsletter also highlights the hardening discipline of harness engineering for AI agents, new desktop coding agents, and the increasing importance of orchestration choices over raw model quality. DeepSeek-V4.1-Flash is noted for its cost-performance, and new multimodal features and robotics foundation models are discussed.
Original sources
Newsletter Latent Space
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.