AI Official

Hugging Face published 'Profiling in PyTorch (Part 3): Attention is all you profile' on 2026-07-10. It shows how to read GPU profiler traces across attention backends — naive, math, memory-efficient, FlashAttention-2, and cuDNN — to find which wins.

The takeaway: 'inefficient'-looking fused kernels like FlashAttention-2 are often fastest, and trace visualization exposes hidden costs (memory copies, CPU overhead) that occupancy metrics miss.

Google Cloud announced on 2026-07-10 that Cloud Run sandboxes are now in public preview — a native, secure runtime for executing untrusted code and agent workloads that starts in milliseconds. It lets developers safely run AI-generated code, headless browsers, and user-submitted scripts without risking host apps, data, or credentials, at no extra cost on existing allocated CPU and memory.

Google Cloud announced on 2026-07-10 that AlphaEvolve — its Gemini-powered code-optimization and algorithm-discovery agent — is now generally available to everyone on the Gemini Enterprise Agent Platform, after a private preview. It runs a Define, Measure, Optimize, and Apply loop, with early customers reporting an 80% supply-chain model gain (BASF), a 10.4% warehouse-routing improvement (FM Logistic), and a 30% genomics error reduction (PacBio).

Anthropic released Claude Code v2.1.206 on 2026-07-10. New: directory-path suggestions for /cd, a /doctor check that flags oversized checked-in CLAUDE.md files, /commit-push-pr auto-allowing git push to the configured remote, /login for Anthropic public gateway endpoints, and worktree-entry confirmations.

Fixes: expired-login errors, --resume/--continue startup input, per-server MCP request_timeout_ms, and /model picker pricing.

OpenAI published Codex CLI pre-release rust-v0.145.0-alpha.2 to GitHub at 00:13 UTC on 2026-07-10 — the latest automated alpha in the in-development 0.145.0 line. Tagged by github-actions with only a 'Release 0.145.0-alpha.2' note and no authored changelog, it ships refreshed prebuilt macOS, Linux, and Windows binaries (164 assets) as OpenAI iterates toward the next Codex stable after the 0.144.x series.

OpenAI published Codex CLI pre-release rust-v0.144.0-alpha.4 to GitHub at 04:33 UTC on 2026-07-09 — the latest automated alpha in the in-development 0.144.0 line, which opened yesterday (alpha.1/alpha.2) after the 0.143.0 stable cut. Tagged by github-actions with no authored release notes, it ships updated prebuilt macOS, Linux, and Windows binaries as OpenAI iterates toward the next Codex stable.

OpenAI promoted Codex CLI to stable rust-v0.143.0, graduating the 0.143.0 alpha line. Highlights: remote plugins now on by default (npm marketplace sources), auth and Responses API traffic can route through macOS/Windows system proxies (PAC/WPAD), new Amazon Bedrock GPT-5.6 Sol/Terra/Luna models with first-class 'max' reasoning effort, and MCP tools using tool search by default.

Anthropic shipped Claude Code v2.1.204, a small maintenance release. It fixes hook events not streaming during SessionStart hooks in headless sessions — a bug that could cause remote workers to be idle-reaped mid-hook.

Hugging Face's LeRobot team released v0.6.0 of its open-source robotics library, themed 'Imagine, Evaluate, Improve.' It adds three world-model architectures (VLA-JEPA, LingBot-VA, FastWAM), five new VLA models (incl. NVIDIA GR00T N1.7, MolmoAct2, EO-1), two reward models (Robometer, TOPReward) and six simulation benchmarks.

A new lerobot-rollout CLI brings DAgger, and FSDP multi-GPU training runs on HF Jobs cloud.

OpenAI published Codex CLI pre-release rust-v0.143.0-alpha.38 to GitHub at 04:34 UTC on 2026-07-07 — the next alpha in the in-development 0.143.0 line, following alpha.35. Tagged automatically by github-actions (commit 19ad83e), it carries no authored release notes and ships updated prebuilt binaries for macOS, Linux, and Windows as OpenAI iterates toward the next 0.143.0 stable cut.

Hugging Face shipped major updates to its Kernels project for packaging and distributing custom compute kernels. A new 'kernel' repository type arrives on the Hub, secured with trusted publishers and Sigstore cosign code signing.

The CLI splits into `kernels` and `kernel-builder`, and the project adds Torch Stable ABI and Apache TVM FFI support, plus infrastructure for agentic workflows that build, optimize, and benchmark kernels across hardware.

Apple published its ICML 2026 conference page ahead of the July 6-11 event in Seoul, confirming 31 accepted papers, 7 oral/expo presentations, and a booth (B305, Hall B1). Research spans computer vision, NLP, and privacy, including 'VideoFlexTok,' 'MemoryLLM,' and 'MODUS.' Apple researchers also serve as Area Chairs and workshop co-organizers.

Microsoft and Nine Entertainment announced an Australian-first AI agreement for news media content on July 3, 2026. Copilot will ground answers in Nine's journalism — The Australian Financial Review, The Sydney Morning Herald, The Age, and Brisbane Times — surfacing snippets and summaries that link back to Nine's sites.

Nine CEO Matt Stanton called the copyright-respecting deal a 'win-win'; it's Microsoft's first of its kind in Asia-Pacific.

OpenAI published Codex CLI pre-release rust-v0.143.0-alpha.35 to GitHub at 02:33 UTC on 2026-07-03 — the next alpha in the in-development 0.143.0 line, following alpha.34 and alpha.33 earlier this week. The release carries no authored changelog (an automated version-bump build from commit 5969673) and ships 146 prebuilt binaries for macOS, Linux, and Windows as OpenAI iterates toward the next 0.143.0 stable cut.

OpenAI published Codex CLI pre-release rust-v0.143.0-alpha.33 to GitHub at 01:58 UTC on 2026-07-02 — the next alpha build in the in-development 0.143.0 line, following alpha.32 from 2026-07-01. Release notes carry only the auto-generated 'Release 0.143.0-alpha.33' with no detailed changelog, indicating routine near-daily iteration ahead of a 0.143.0 stable cut.

Prebuilt binaries ship for macOS, Linux, and Windows.

OpenAI published Codex CLI pre-release rust-v0.143.0-alpha.32 to GitHub at 03:47 UTC on 2026-07-01 — an alpha build in the in-development 0.143.0 line, following alpha.31 from 2026-06-29. Release notes carry only the auto-generated 'Release 0.143.0-alpha.32' with no detailed changelog, indicating routine iteration ahead of a 0.143.0 stable cut.

Prebuilt binaries ship for macOS, Linux, and Windows.

OpenAI published Codex CLI stable release rust-v0.142.5 to GitHub at 01:15 UTC on 2026-07-01 — a bug-fix patch on the 0.142 stable line, running parallel to the 0.143.0 alpha dev builds. The fix stops full Responses WebSocket request payloads from being written to trace logs (PR #30771) and was backported to the release/0.142 branch.

It maps to the @openai/codex npm package v0.142.5.

Google UK's 2026 Economic Impact Report shows workplace AI adoption surged to 73%, up from 34% in 2025 — but gains concentrate among the top 15% 'AI Trailblazers,' who report 84% higher promotion rates, 88% better performance reviews, and save ~8 hours weekly. Google launched an AI skills quiz and the 'AI Works for Britain' initiative, backing the UK goal of training 10M workers by 2030.

Google tools supported £140B in UK economic activity in 2025.

OpenAI published Codex CLI stable release rust-v0.142.4 to GitHub at 05:04 UTC on 2026-06-29 — a patch over v0.142.3 on the 0.142 stable line, running parallel to the 0.143.0 alpha dev builds. Release notes state 'No user-facing changes were identified for this release,' marking it a routine maintenance build.

Distribution includes prebuilt binaries for macOS, Linux, and Windows.

omlab released VLX-Flow, an open multimodal model for real-time, continuous video understanding that processes video as streaming chunks instead of waiting for a query—separating observation from answering. A two-layer memory (visual cache + semantic memory) with Linear Attention keeps latency stable as streams grow, with more stable time-to-first-token than full-attention and sliding-window baselines.

Targets on-device cameras and robotics.