PyTorch
Torchtitan trains MoE routers in fp32 with bf16x9 on Blackwell
MoE router forward and backward outputs now stay in fp32, using bf16x9 kernels to keep tensor-core throughput on Blackwell.
read --wire →
$ tail -f topics/ai.log
AI / ML news, every morning.
Daily updates from the AI and machine-learning ecosystem - PyTorch, Hugging Face, the Anthropic and OpenAI SDKs, local LLMs, and the agent frameworks.
256 updates across 10 projects this week.
Your week in AI / ML, every Monday.
One calm review of what shipped across AI / ML - the commits, releases, and security advisories that matter. Every Monday, with security advisories same-day. Free, unsubscribe in one click.
We'll start you on the top AI / ML projects - refine anytime. · Read a sample issue →
Recent wires
JAX pins XLA to new revision
JAX updated its XLA dependency to a specific commit revision, aligning with upstream changes.
read --wire →
OpenAI
Codex voice-host ships real-time microphone audio; openai-node fixes streaming hangs
Codex wires processed microphone audio into voice-host peer connections, while openai-node patches an SSE iterator hang and a stale terminal message snapshot.
read --wire →
Anthropic
Claude Code 2.1.263 ships across GitHub Actions and SDK packages
Claude Code 2.1.263 and Agent SDK 0.3.263 rolled out to the GitHub Action and both the Python Agent SDK and the CLI's changelogs today.
read --wire →
AI Agents
LangChain adds async OpenAI tools; OpenHands fixes flaky CI
LangChain core 1.6.2 and new OpenAI async tool support land alongside OpenHands fixes for a flaky test suite and broken branch URLs.
read --wire →
Local LLMs
vLLM Fast Start maps weights zero-copy; GLMGA sampling capped
Two security patches and a zero-copy weight cache land across vLLM and llama.cpp, while sglang ships v0.5.19 with 786 PRs.
read --wire →
Hugging Face
TRL removes PPOTrainer, vendors Liger-Kernel losses
TRL removed PPOTrainer and vendored its fused linear losses from Liger-Kernel, while Transformers stopped synchronizing the accelerator on each decode step and Diffusers added a modular blockset for Wan 2.2 VACE.
read --wire →
PyTorch
executorch fixes dropped edge-manager passes; TorchTitan overhauls MxFP8
Executorch now runs a recipe's edge-manager passes after partitioning, TorchTitan replaced its MxFP8 training stack, and TensorRT's Dynamo integration ships a batch of correctness fixes.
read --wire →
JAX ships thread-safe PyArray, changes jax.jit inline semantics
JAX merged three changes today: PyArray is now thread-safe, inline=True in jax.jit means late inlining, and a previous commit was reverted.
read --wire →
Meta
React Native fixes bridgeless ordering and commits dev-machine paths
React Native's bridgeless runtime can now schedule buffered CallInvoker work before earlier module calls, and the `spm add` command no longer writes a developer's absolute hermesc path into the app project.
read --wire →
OpenAI
OpenAI ships retry fix, prompt cache diagnostics, sandbox security
Python SDK users get safer retry handling and prompt cache visibility, while Codex tightens Windows sandboxing and Guardian review behavior.
read --wire →
Anthropic
SDKs ship compliance types, Claude Tag usage breakdowns
All three Anthropic SDKs released today, adding named types for organization compliance state and Claude Tag breakdowns to usage reports.
read --wire →
PyTorch
Inductor fixes stale captured-tensor mutations in while_loop
PyTorch's Inductor now correctly replays cond_fn mutations of captured tensors in while_loop, fixing silent wrong results in later compiled calls.
read --wire →
AI Agents
LangChain 1.4.0 and anthropic 1.7.1 ship with MCP adapter, Claude Fable 5.1
LangChain core 1.4.0 exposes a first-class MCP namespace and the Anthropic package adds support for Claude Fable 5.1, while OpenHands splits automation permissions and Cline's desktop app now runs shared Hub plugins.
read --wire →
Local LLMs
Ollama enables speculative decoding under structured output, vLLM adds LoRA and fusion
Ollama's MLX runner now runs speculative decoding under structured output constraints, promising near-full draft throughput on MLX models, while vLLM lands static LoRA loading in its Rust frontend and begins manual activation-quant fusion.
read --wire →
Hugging Face
TRL documents 1M-token training, diffusers fixes group offloading, kernels fix Intel heuristics
TRL ships a runnable million-token SFT example, diffusers repairs a group-offloading device mismatch, and triton-kernels corrects Intel autotuning constants that cost roughly 10x on Battlemage.
read --wire →
JAX removes deprecated Pallas APIs, adds barrier-skip for multi-device calls
JAX purged a set of deprecated Pallas APIs and gave Mosaic GPU callers a way to skip multi-device barriers.
read --wire →
Meta
React Native sheds Paper-era internals, stabilizes Maestro e2e
Core component specs stop announcing old-architecture names, and e2e suites get guards for flaky keyboard and startup timing.
read --wire →
OpenAI
macOS Codex sandbox hardened against terminal input injection
Codex patched a sandbox escape that let a terminal-tied child queue input for the unsandboxed shell, and OpenAI SDKs shipped a patch release adding GPT-6-Astra support.
read --wire →
Anthropic
Claude Code bumps to 2.1.260 across action and SDK
Anthropic's GitHub Actions and Python SDK now bundle Claude Code 2.1.260 and Agent SDK 0.3.260, keeping the toolchain in lockstep.
read --wire →