Meta FAIR 发布 Autodata 智能体数据科学家

In today’s issue:

• Meta FAIR introduces Autodata
• Zyphra releases ZAYA1-8B
• SubQ ships a 12M-token frontier model
• Anthropic introduces Natural Language Autoencoders
• Claude Managed Agents adds dreaming and multi-agent
• Printing Press: an agent CLI factory
• Flue agent harness framework launches
• Anthropic adds keyless auth
• AlphaEvolve marks one year of impact
• Goodfire opens a neural geometry series
• Firefox hardened with Claude Mythos

And all the top AI dev news, papers, and tools.

Top Stories

Autodata: An Agentic Data Scientist From Meta FAIR

Meta FAIR (Jason Weston et al.) introduced Autodata, an agentic data scientist that builds high-quality training and evaluation data autonomously. The framing is that inference compute can be converted into model quality if the data pipeline itself is an agent.

• Agentic Self-Instruct loop: A planner-executor agent generates, critiques, and refines training and eval examples in a closed loop, replacing static seed sets with a process that keeps producing harder data as the model improves.
• 34-point weak-to-strong gap: On a CS research QA task, Autodata data opens a 34-point accuracy gap between weak and strong models, a much larger separation than off-the-shelf instruction sets achieve.
• Inference compute as a quality lever: The work reframes synthetic data as the place where inference budget pays off, an angle that lines up with Microsoft’s FaraGen and the broader synthetic-environments thread.
• Why it matters: Pairs naturally with self-improving agent runtimes (Claude Managed Agents Outcomes loop, ACE, AHE), giving teams a credible recipe for the data half of the self-improvement story.

Blog

Read more