An industry pulse of what is happening in AI.
Sim2Reason trains LLMs inside MuJoCo physics simulators, zero human annotation. Generate scenes, auto-label QA pairs, RL-train on synthetic data. Zero-shot: +5-10% IPhO, +17.9% JEEBench, +4.4% MATH 500. Outperforms models trained on curated real-world QA pairs. CMU + Lambda.
Blackwell chips: $4.08/hr, up 48% from $2.75 just 2 months ago. CoreWeave +20% & extended minimum contracts from 1 to 3 years. Anthropic limits Mythos to ~40 orgs. OpenAI CFO: "We're making some very tough trades...because we don't have enough compute." The age of abundant AI is over.
Google launches Android CLI in preview: a terminal interface for agent-driven Android development. 70% fewer LLM tokens for project setup and 3x faster tasks in internal tests. Includes Android Skills (SKILL.md files) and Android Knowledge Base. Works with Claude Code, Codex, Gemini CLI, and others.
AI Mode in Chrome now opens web pages side-by-side with your search panel -- no more switching tabs to follow a link. A new plus menu lets you add open tabs, images, and PDFs as context for follow-up questions. Canvas and image creation tools are accessible from the plus menu anywhere in Chrome. Rolling out in the US today.
For a long time, GPT has been a monkey's paw. Claude has been a genie. Opus 4.7 changes that: now "substantially better at following instructions." Prompts for earlier models may now produce unexpected results. Anthropic built a genie. Today they shipped something closer to a paw.
Opus 4.7 in Claude Code defaults to a new xhigh effort level (between high and max). Fixed Extended Thinking is gone -- adaptive thinking lets the model decide when to reason. It calls tools less and spawns fewer subagents by default. Specify tasks fully in the first turn.
Codex update: background computer use, in-app browser for frontend iteration, image generation via gpt-image-1.5, 90+ new plugins (Atlassian, GitLab, CircleCI, Microsoft Suite, Superpowers), memory across sessions, automations that can schedule work across days or weeks, and proactive task suggestions.
A 21GB quantized Qwen3.6-35B-A3B on a MacBook M5 beat Claude Opus 4.7 on the pelican-riding-a-bicycle SVG benchmark -- the flamingo-on-unicycle backup test went to Qwen too. This benchmark used to roughly track general model quality; that correlation may now be broken.
The proof-of-work analogy for AI cybersecurity is wrong. Bug-finding is intelligence-capped, not compute-capped. Run an inferior model infinite times and it still won't find the OpenBSD SACK bug's multi-step chain. "More GPU wins" is the wrong frame. Better models win.
Qwen open-sources Qwen3.6-35B-A3B: a sparse MoE (35B total / 3B active) that rivals much larger dense models on agentic coding. SWE-bench Verified 73.4%, Terminal-Bench 2.0 51.5%. Natively multimodal. Open weights on Hugging Face.