All AI news

AIUnsloth now supports OS-level sandboxing on Linux via bwrap, on Mac via seatbelt, and on Windows via Microsoft's MXC. Per-tool-call latency is under 100ms across all three, and its software-style sandboxing with regex AST checks adds about 3ms. The Windows integration was built in collaboration with Microsoft.

AIPeriodic Labs' Liam Fedus and Ekin Dogus Cubuk explain why scientific discovery differs from math and coding, and why experiments remain the ground truth. They describe reinforcement learning grounded in physical experiments, AI-driven materials characterization, and the view that failed experiments can be valuable training data. The transcript was truncated before the discussion of giving lab instruments "140 IQ" was completed.

AIGoogle Research is presenting EnvHarness, a flexible framework that enables co-evolution between LLM agents and their training environments, at the #COLM2026 Google booth #107 today at 11:00 AM PT. The post notes that static environments limit agent growth, and EnvHarness is described as a plug-in architecture that dynamically reshapes environment behaviors to improve reinforcement learning and adaptability.

AIOpenClaw released its v2026.9.9 patch, adding support for GPT-6.1 Sol in Codex and Claude Haiku 5.5. The update also improves recovery from failed updates, fixes missing iMessage replies, and resolves scheduled-job issues. The release credits 90 contributors.

AIGoogle Gemma has published a developer guide for EmbeddingGemma 2, with code snippets to help developers start searching beyond text. The post directs readers to the full guide on the Google Developers Blog.

AILauren Tan (@poteto) asks a bot to monitor interesting ways people use Grok Bot on X and send a daily 9am Slack digest. The request also asks for the coolest relevant ideas plus suggestions for improving her workflow.

AIVoice Arena's Monsoon ASR dataset fine-tuned Whisper Medium on Bengali FLEURS, reducing LLM word error rate from 85.27% to 7.65%. The corpus spans 100,000 hours across 50 languages, and Voice Arena says more than 80 organisations have asked to license it since its launch a week ago.

AIAWS published a reference architecture for running multiple teams on one Amazon SageMaker HyperPod EKS cluster, with each team isolated in its own Kubernetes namespace. The design combines AWS IAM Identity Center for authentication, per-team SageMaker AI domains, HyperPod Task Governance for fair resource allocation, and namespace-level cost allocation for per-team spend visibility.

AIJason Liu says he lacked the vocabulary, not the ambition, to design objects in CAD for 3D printing, so he learned printing, part efficiency, and shape terminology. He argues AI is like magic that still requires learning the right spells and incantations, and that it has made him want to understand the subject more deeply.

AIJetBrains has released Mellum2.1, a 12B mixture-of-experts thinking model with 2.5B active parameters, under Apache 2.0 on Hugging Face. Post-training reinforcement learning in real software repositories raised SWE-bench Verified from 2.0 to 47.0, according to JetBrains' self-reported results. Qwen3.5-9B still leads on SWE-bench Pro, GPQA Diamond and AIME, and GGUF builds start at 7.0 GB for local use.

AINatura AI's Interface is a ring that lets users press and hold to speak requests to AI agents such as Claude Code, Codex, or Hermes, then release to send them. The post argues that screenless interfaces may define the next phase of agent use, since handing work to agents is currently slowed by pulling out a phone. Early-adopter pricing is $99, with shipping slated for January.

AIJabil senior director Thomas Brown says humanoid robots are entering a phase of tens of thousands of units, where manufacturability, cost structure, and quality become central. He says Jabil works with developers to cut costs for scale, while compute and memory prices remain a pain point, and that humanoids make sense in factories and warehouses while mobile arms still suit high-speed tasks.

AIGoodfire Research describes probe-based cyber monitors for Kimi K3 and GLM 5.3 deployed on a production inference stack. The probe filters suspicious exchanges before an LLM judge reviews them, reaching about 93% recall at a 5.5% benign-session interruption rate at roughly 50x lower judge cost. In FAR.AI's red-teaming, the monitor reduced universal jailbreaks to zero across 140 tested strategies.

AIVercel says StepFun's flagship Step 5 Preview, built for agentic coding, research, and finance, is now live on AI Gateway. The model offers a 1M-token context window, accepts text and image input, and uses a 600B-parameter mixture-of-experts design with 27B parameters active.

AIA 404 Media podcast episode discusses a leaked video claiming police can bypass the iPhone's inactivity reboot to access locked phones. The episode also covers a debate over an LLM "torture" robot prison and a New Mexico case involving ChatGPT.

AIAnthropic says Claude subscribers can claim API credits by going to Settings › Billing on claude.ai and linking or creating a Console organization. New subscribers become eligible to claim after 7 days.

AIAnthropic's credits can be used with the Messages API, Claude Managed Agents, and the Agent SDK. They cannot be used for interactive Claude Code sessions, but they also apply in third-party harnesses that accept a Claude API key.

AIAnthropic now provides monthly Claude Platform API credits to Max and Team subscribers: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team. The credits can be used on any Claude model, including in code or third-party harnesses.

AIArena has announced a Series B round co-led by Lightspeed and Khosla Ventures, with participation from Salesforce Ventures, 01 Advisors, Dell Technologies Capital, and Endeavor Catalyst. Existing investors Andreessen Horowitz, Felicis, a16z's AMP, QuantumLight, and The House Fund also supported the round.

AIArena links readers to a full technical report on its blog and to a full rankings page in its Alignment Index. The post itself gives no model names, scores, or findings, so the substance is only in the linked materials.

AIAnthropic is committing $150 million to the Genesis Mission to deepen its support for scientific research and technological discovery. The company also plans to make Claude and its technical support available to more than 15 federal agencies.

AILangChain says developers can now build agents that find real products, prepare purchases, and pay, with a person approving each order. Users request purchases in Slack and approve orders through Stripe's Link agent wallet, built on MPP and Managed Deep Agents.

AIOracle is using ChatGPT Work and Codex to turn specialist knowledge into fast, repeatable workflows across recruiting, engineering, and operations. The source does not provide figures, timelines, or specific results beyond the headline's claim that days of work can take minutes.

AIJulien Chaumond, Hugging Face's account, mocked the idea of needing US or Chinese rivals when European partners exist. The remark responds to Berlin-based search engine Ecosia dropping French AI partner Mistral in favor of open-source models, including those from China.

AIHelm.ai said it signed $70 million in commercial contracts for its foundation models for physical AI over 12 months, spanning global automotive OEMs, Tier 1 suppliers, and industrial automation companies. The Redwood City, Calif.-based company said it has projects bound for production in autonomous vehicles, mining, and construction, and that it is on a path to break even. CEO Vladislav Voroninski said its models are trained on unsupervised "deep teaching" and are environment-agnostic.

AISantiago says physical consistency is the most important and hardest feature of a world model, and that many generated videos show objects defying gravity. He reports that Seedance 2.5 is currently the best among the evaluated world models. The post links to a physics evaluation benchmark in which eight video world models reached a top score of 57.76/100.

AIWillow Care Inc., maker of the AI dictation app Willow Voice, chose CoreWeave for its forward-deployed support rather than compute alone, according to co-founder and CTO Lawrence Liu. Liu said CoreWeave's reinforcement learning infrastructure lets Willow focus on eval alignment, while Willow fine-tunes its own speech recognition model and pairs it with a compact post-processing LLM. He said inference demand is growing faster than training as dictation use climbs.