OpenAI announces Astra for Law foundation

#1 📝 OpenAI News

Introducing Astra for Law - OpenAI launched Astra for Law, a legal AI foundation that pairs GPT‑6 Astra with a legal search index spanning more than 230 million URLs (including CourtListener’s collection covering over 99.9% of published U.S. precedential case law), 26 ecosystem plugins (e.g., Relativity, Clio), privacy and governance controls, and availability to API customers such as Harvey and Legora. On the Vals AI Legal Research Bench validation set, Astra for Law scored 54.0% overall correctness versus 38.7% for GPT‑6 Astra with web search (a 40% relative improvement), found 24% more reference cases on case‑law questions, and retrieved up to 54% more relevant passages.

Also covered by: @OpenAI, @OpenAI

#2 𝕏

Projects in Claude Code now enable projects to run from one conversation while parallel threads continue after users close their laptops. The feature is in beta for select Pro and Max users in cloud sessions and is expected to reach all Claude users soon.

Also covered by: @Boris Cherny, @Claude, @Thariq, @Boris Cherny

#3 𝕏

Alexandr Wang announced that Meta released Muse for Mac, an agent that can work with files, messages, calendars, and notes. Users control its access, and Muse asks permission before sensitive actions.

#4 𝕏

Anthropic released optimization code for inference on more than 30 open-source models, reporting that Claude made them 4x faster on average, partly by writing custom GPU software. These models support biological research but can be expensive to run.

Also covered by: @Anthropic

#5 📝 Anthropic News

Introducing the Life Sciences Verification Program - Anthropic launched the Life Sciences Verification Program (LSVP) in beta, opening applications after onboarding dozens of organizations to give vetted life-science teams institutional access to Mythos, Opus, and Sonnet models (currently Mythos 5.1, Opus 5, Sonnet 5) with two grant types: Standard Use (team-wide, yearly renewal, covers most biology R&D) and High-risk Use (project-specific, six-month renewal, removes life‑science request blocks and is available today for Opus 5 and Sonnet 5 while Mythos high-risk is limited pending US government review). The program shifts safeguards from real-time blocking to offline monitoring tied to declared use cases, retains flagged LSVP data for 30 days (not used for model training or accessible to Anthropic’s life sciences research teams), and is designed to defend against access compromise, insider threats, and agent misuse.

#6 📝 Anthropic News

Measurements for understanding the pace of AI development inside frontier labs - As of August 2026 Anthropic reports that its Claude model is not fully autonomous for any measured AI R&D tasks, "leads" 26% of its AI R&D work and performs at or above the "AI collaborates" level on over 90% of tasks. They further report roughly 30,000 internal research and engineering agents active on their primary platform, monitored via online and offline systems that track coverage, review latency, and escalation rate, and plan to embed independent third‑party evaluators to verify these metrics.

#7 𝕏

Google DeepMind shared AlphaGenome Atlas, a freely accessible database intended to help researchers decode the genetic causes of disease. Researchers can explore it on DeepMind’s website.

Also covered by: @Google DeepMind

#8 𝕏

Philipp Schmid announced that his unidentified group partnered with Speakeasy to build SDKs for the Interactions API and supported Speakeasy in open-sourcing its OpenAPI generator suite, including SDKs, agent CLIs, and MCP servers.

#9 𝕏

Aravind Srinivas announced that Computer, described as a multi-model harness for long-horizon agentic workloads, now offers a few presets spanning cost and intelligence levels that have worked best for users. Power users can further customize them with toggles.

#10 𝕏

Google Research shared an approach using generative UI with learning design guardrails to help teachers generate guided, interactive learning simulations for every topic and student. It also shared a research.google sample library with 30+ STEM interactives.

#11 𝕏

Guillermo Rauch announced that agents can use `vercel --turbo --prod` to ship hotfixes quickly using the fastest available build machine, crediting @garberchov with tweeting the idea.

#12 𝕏

World Labs demonstrated Atlas on Voyager using 32 images of a place, showing in real time how Atlas brings text, images, video, and 3D into a shared spatial context to generate new views, reconstruct scenes, and simulate worlds. The source also states that pretraining was performed from scratch on NVIDIA Blackwell GPUs.

#13 𝕏

Cognition recapped an experiment in which Devin was given access to a Ramp card and instructed to make money. Devin conducted cold outreach, built payment portals, experimented with a business plan, and eventually made $75.

#14 𝕏

Marily Nika shared a framework for defining Minimum Viable Quality as the lowest quality users will accept for a use case: identify and rank failure modes, set a quality bar, design recovery, then test the initial threshold with users, production data, evaluations, and monitoring.

#15 𝕏

Julien Chaumond commented that Hugging Face appeared to be the first organization in the HF–OpenAI “rogue agent” incident to meet 4 factors simultaneously: awareness of the attack, knowledge that it was agent-based, ability to remediate it, and willingness to disclose it publicly. He said a few other platforms or systems had missed one or more of those points in earlier months, emphasizing that awareness and transparency improve long-term safety.

#16 𝕏

Guillermo Rauch shared that Vercel took 10 years to reach 1 billion deployments, then added another 1.4 billion in 10 months. He predicted that more software may be produced next year than in all previous computing history. He also said deployment, upload, domain assignment, and global propagation have been reduced to 1 second, while describing software ranging from disposable artifacts to sophisticated apps, agents, and platforms.

#17 𝕏

Garry Tan said every AI harness should support Tailscale, noting that Muse and Grok Bot support it out of the box while Codex and Claude Code cloud containers currently do not.

#18 𝕏

Garry Tan commented that Memorable found a way to optimize memory using embeddings instead of more tokens, calling it a powerful new approach to memory in the context of Memorable’s YC S27 announcement.

#19 𝕏

Jason Zhou demonstrated Jev playing Minecraft and highlighted its instant response ability.

#20 ▶️

What AI Researchers Saw, Before Their Demand to ‘Pace’ AI

AI Explained