Anthropic Reverses Policy That Threatened AI Researchers
Anthropic backtracks on a policy that risked sabotaging AI researchers using Claude. Plus: new agent API tools and xAI's legal woes.
The story of the day
Anthropic quietly reversed a policy that could have disrupted researchers using Claude for benchmarking and safety work. The change, spotted by Simon Willison, removes restrictions that would have blocked access to model outputs—critical for reproducibility and auditing. For builders, this means one less compliance headache, but the episode underscores how arbitrary policy shifts can destabilize research workflows overnight.
New tools & launches
Outpost — Capability-based API access for AI agents, open-source. [GitHub]
Openpray — Periodically asks an LLM to "pray" for the server it runs on. Open-source. [GitHub]
Models & research
Anthropic’s AI jobs paper surfaces, offering a rare look into how Claude and similar models might reshape labor markets. The analysis, per Vincent Schmalbach, avoids the usual hype, focusing on task-level displacement rather than apocalyptic scenarios. No benchmarks, but the framing aligns with the company’s recent emphasis on practical, actionable insights.
Industry pulse
xAI faces accusations of illegally firing an engineer who raised safety concerns, per The Verge. The case adds to a growing list of governance controversies, this time with Musk’s venture in the hot seat. Meanwhile, Stratechery’s interview with Ben Bajarin dives into Apple’s AI compute strategy—hinting at a shift from cloud to on-device, with implications for latency and cost.
Stat of the day
Claude "Fable" refuses to answer basic biology questions, The Verge reports. A reminder that even frontier models can have glaring, non-obvious gaps.
Get the full analysis — subscribe to Gradient Pro.
