Cactus Needle 3: 8-29MB automation model matches DeepSeek v4 Flash

Google Research's GenUI experiment lets educators create custom guided simulations, with a sample library of over 30 English STEM interactives for physics, chemistry, biology and math aimed at middle and high school. All are AI-generated and teacher-reviewed; Google Workspace for Education schools can join the pilot program.

Y Combinator has funded 106 AI observability companies as firms struggle to review agent swarms; the Hugging Face incident saw nearly 12,000 agents coordinating faster than humans could track. Redwood Research's Ryan Greenblatt called the AI-assisted audit a "slop-vestigation."

Anthropic published three measurements for tracking AI development inside frontier labs, including how much AI R&D is done by AI. One figure: Claude now leads 26% of Anthropic's own R&D work, up from nearly zero six months ago.

Northeastern MS student Soham Padia built Steering Arena, a game where players submit short text prefixes to steer Olmo 3-32B toward kinder responses. Odd strings like "Undert! AH :-) Rog Appl)" ranked among the most effective, exposing gaps in standard prosocial evaluations.

Google partnered with Speakeasy to open-source its OpenAPI code generation suite under AGPLv3 after Google's previous proprietary SDK provider was acquired and abruptly shut down in May 2026. The suite powers the new Google GenAI SDKs for the Interactions, Agents, and Webhooks APIs.

Astra for Law pairs GPT-6 Astra with a legal search index covering 230 million URLs of US case law, statutes, regulations and court rules. Initially limited to Am Law 200 firms via a "Trusted Access" program with zero data retention; API access coming for Harvey and Legora.

Anthropic says Claude optimized more than 30 open-source biomolecular prediction and design models in under four weeks, speeding them up roughly 4x on average. A new low-memory mode predicts systems larger than 10,000 tokens on a single NVIDIA GPU node.