Product · OpenAI
OpenAI Ships Astra for Law, a Frontier Model Built for the Practice
A new offering powered by GPT-6 Astra bundles tools, settings and context to support the expertise and judgment of lawyers and legal-technology firms.
OpenAI is moving its frontier models into regulated verticals, and the law is the latest test. Astra for Law wraps GPT-6 Astra with the tooling a legal team actually reaches for, so drafting, review, research and compliance sit inside one workspace while the final call stays with the practitioner. The launch extends a pattern the lab has run across medicine and engineering: ship a general model, then bind it to a domain's documents, permissions and workflow so that expertise, not the model, remains the point of trust.
The company frames the product around professional judgment rather than automation of the lawyer. It pairs the model with settings and context tuned for legal work, a clear signal that frontier labs now see the path to revenue running through high-stakes, high-accountability industries where an error is expensive and a well-placed draft is worth real money.
Industry · Anthropic
Anthropic Publishes Three Metrics to Track AI Self-Research
As models start building the next models, the lab wants the public to see the progress.
AI systems are getting more powerful, and they are increasingly used to build the next version of themselves. Anthropic is publishing three measurements to make that dynamic legible, starting with how much AI R&D is now done by AI itself. The goal is transparency about a frontier that has largely been described in private, not measured in public.
Product · Anthropic
Anthropic Opens Its Life Sciences Program and Debuts Mythos
Life-science professionals can apply to use the models, including Mythos for the first time.
Applications are open for the Life Sciences Verification Program. Through it, researchers and clinicians can use Anthropic's models under a new set of safeguards designed to enable the full range of biology-related work. The program's headline is Mythos, a model now being surfaced to life-science users for the first time, with guardrails tailored to biological discovery rather than general-purpose chat.
Research · Google DeepMind
DeepMind's AlphaGenome Atlas Helps Flag Disease-Causing DNA Variants
Researchers at the Broad Institute, the University of Exeter and beyond are using AlphaGenome Atlas to better identify potentially pathogenic DNA variants and interpret their role. The model turns a long-standing bottleneck in genomics, distinguishing the variants that matter from the vast background of harmless mutation, into a task a trained system can now help shoulder.
Model Release · via @PrismML
Ternary Bonsai 2 27B Cuts Model Size to a Ninth
A ternary build of a 27-billion-parameter model claims 9x smaller than its full-precision counterpart.
Built on Qwen3.8 27B, Bonsai 2 27B uses ternary weights to shrink the footprint dramatically while chasing the parent model's quality. The release continues the push toward models that are cheap enough to ship on edge hardware and small enough to run where a full-precision giant cannot.
“The concerns around the most powerful, unreleased models are real. But enterprise AI adoption is still far too slow.”
— Kai-Fu Lee
Cutting the Cost of Open-Source Biology Models
Biologists rely on specialized open-source models for molecular structure, drug-like molecule design and mutation prediction, but running them is expensive. Anthropic details how to lower that cost so it does not limit impact.
Appshots Comes to Windows ChatGPT
Appshots lets ChatGPT pull context from the app you are working in, so debugging an issue, grabbing a UI or moving data no longer means endless copy and paste.
FLARE-AI Proposes Open Flaw Reporting for AI
A study audits twelve flaw-reporting systems from AI developers and safety groups, then designs FLARE-AI from feedback by forty-nine experts, routing a single machine-readable report to multiple developers and coordinators.
Runway Adds Enhance Frame Rate
A new frame-interpolation model converts any footage to 25, 30, 48, 60 or 120 fps, including the NTSC 59.94 standard, targeting editors who need clean slow motion from ordinary sources.
World Labs' Atlas Model Rebuilds a Campus in Real Time
From just thirty-two images of NVIDIA's campus, the Atlas model created a navigable park in real time on NVIDIA's open platform, pointing toward world models that turn sparse photos into explorable space.
Replit Ships Custom Connectors for Teams
Workspace admins can now connect supported, publicly accessible HTTPS REST APIs through API-key authentication, giving teams a governed way to build against the internal systems they already use.
Bionic Agent Learns Introspection and Session References
The agent can now search its own session history, handling longer context and repeated compaction, and lets you reference other sessions directly in the composer with an @ mention.
Computer Gains Effort Controls
Effort presets combine the orchestrator model with reasoning depth, letting users tune how deeply and efficiently Computer works through a long-horizon task. Live on web now, with mobile and desktop to follow.
The New Pika Is an AI Platform for Creatives
Pika unveils a rebuilt creative platform made for and by creators, aiming for simple, comprehensive workflows and outputs that meet exacting production standards.
Interactive Avatars Open to Your Own LLM
With the Interactive Avatar API, you bring your own LLM, knowledge base and logic while Synthesia handles the avatar layer, voice, face and lip-sync, so an avatar can listen and respond inside your product.
Kimi K3 Serving Doubles Throughput in vLLM
Kimi K3 in vLLM now delivers 2.2 to 2.8x throughput on the B300 benchmark versus v0.27.1, with the gains broken down across scheduling, KDA state handling and MoE kernels.
Verified U.S. Clinicians Get Free GPT-6 Astra Pro
Verified clinicians in the United States can get free GPT-6 Astra Pro access through ChatGPT for Clinicians, extending the lab's push to put frontier models directly into clinical workflows.
“AI existential-risk probabilities are too unreliable to inform public policy.”
— Arvind Narayanan and Sayash Kapoor
Infrastructure · Zhipu AI
Zhipu Used GLM-5.3 to Optimize Its Own Inference Stack
A GLM-5.3-driven agent, not an engineering team, did most of the optimization.
Zhipu disclosed that all production inference for GLM-5.3-Flash now runs on more than 100,000 domestic AI accelerators, moving from first run to full production in under two weeks with end-to-end throughput up 3.2x. The striking detail is who did the work: an AI agent driven by GLM-5.3, rather than a team of engineers, handled the bulk of the optimization, a concrete case of models improving the systems that serve them.
Comment · Vercel
More Software Will Be Built Next Year Than in All of History
Vercel's Guillermo Rauch argues that it took a decade to reach one billion deployments, and then another 1.4 billion in the last ten months alone. Much of the coming software will be tiny, personal and even disposable, the long tail of AI-assisted building.
Grok Bot Now Has a Voice
Researchers Spend $7–8k a Day on Codex
Keras 3 Modernized Expedia's Ranking
Recursive Self-Improvement Since 1987
A new post covers four decades of RSI, from meta-evolution and self-modifying policies in a 1987 diploma thesis to the Gödel Machine and today's LLM agents, a reminder that the idea long predates the current debate.
Agora: Git as Shared Memory for AutoResearch
Thirteen model agents ran for nearly twelve days without central scheduling, recording results as append-only commits and producing 1,703 contributions on a frozen 119M-parameter model.
LimiX-2 Targets Structured-Data Intelligence
A contextual mechanism network shifts tabular in-context learning toward mechanism-oriented joint modeling, topping TabArena, TALENT and BCCO with a single model and no downstream parameter updates.