PyTorch 2.14 Ships Faster Apple Silicon and 100x GPU Speedups
PyTorch 2.14 lands with a new CUTLASS-based GEMM backend, in-place fault tolerance for distributed training, and native linear algebra on Apple Silicon.
20 stories
PyTorch 2.14 lands with a new CUTLASS-based GEMM backend, in-place fault tolerance for distributed training, and native linear algebra on Apple Silicon.
Google's agentic IDE ships a bundle of upgrades spanning genomics skills, deep-reasoning multi-agent workflows, generative UI artifacts, and a rebuilt terminal.
Cognition brings its two-agent Fusion harness to Devin CLI and Desktop, pairing a frontier planner with a cheap executor for up to 39% savings.
Google's WikiSkill paper gained 15 points of accuracy from a notebook its agent couldn't read, and its skills still need a transfer test
A self-distilled fine-tune of Qwen3.8-Flash-Next cuts reasoning tokens by 26% and lifts speculative decoding acceptance from 69% to 90%.
NVIDIA Labs released SoL-Pi, a coding agent extension that cuts token cost by roughly one-third using four mechanisms discovered by auto-research loops.
Comfy-Org has repackaged the MAP team's YuE2 music model into single-file safetensors, and ComfyUI now ships native nodes to run it.
Sakana AI's Fugu Max and Fugu Ultra v2 orchestrate pools of open models to beat frontier LLMs on cost and capability simultaneously.
A community maintainer used the Heretic abliteration tool to strip refusal behavior from OpenBMB's MiniCPM5-2B, cutting refusals from 99/100 to 4/100 while keeping tool use and reasoning intact.
Google Research introduces ToolGrad, an answer-first framework that generates tool-use training data with a 99.8% pass rate at lower cost.
Cursor's new Projects feature replaces one-off chats with a persistent coordinator agent that delegates to subagents, syncs context, and runs recurring work.
NVIDIA and USC present HorizonRelight, a diffusion transformer method that keeps lighting stable across long videos by propagating context between chunks.
OpenAI opens up the same agent harness that powers Codex, letting developers spin up long-running cloud agents with a single API call.
Anthropic ships a live session viewer and a server-side auto permission mode for Claude Managed Agents, letting the model decide when to pause for human approval.
OpenAI launches a vertical ChatGPT for banks that bundles Daloopa, PitchBook, and LSEG data with GPT-6 Astra reasoning and firm-specific templates.
humans& releases Persimmon, a 550B user model built to simulate real people in group chats, fooling AI judges 20 percent of the time.
Runway details how it turns diffusion video models into causal, autoregressive streamers using teacher forcing and on-policy distillation.
Together AI ported its ThunderKittens kernel framework to NVIDIA's Vera Rubin NVL72, hitting 22.4 PFLOPS on NVFP4 GEMMs and rivaling cuBLAS.
A math benchmark built to resist AI just fell. GPT-6 Astra cracked the final Tier 4 problem, closing out a 98 percent run in 14 months.
OpenAI's full-duplex voice model lands in the API at $0.05 per minute, letting agents listen and speak simultaneously while delegating reasoning to backend models.