◈ AILog

EdgeWise runs Qwen3.5-9B locally, delegating only key steps to a cloud LLM. Inference cost falls 62% on SWE-Rebench, 34% on Terminal-Bench 2.1.

Speculative Forcing freezes converged video tokens early: 4-step quality at ~1.7 steps per token on LongLive, over 1.5x faster on one H100.

Shanghai AI Lab's Intern Lumina U2: a 16.9B MoE diffusion LLM, 1.4B active params, unifying image, video, and 3D understanding with image generation.

OmniCompass probes 50 multimodal models on when, where and what in video; open weights trail closed models, widest on audio-visual tasks.

NVIDIA open-sources Nemotron 3.5 Super VL: a 120B MoE with 12B active params, distilled from Nemotron 3 Ultra, frontier accuracy at a fifth the size.

Meta released Muse Spark Open: 816B / 42B MoE, 1M context, two Mac Studios. First since Llama 4. JobBench and MCP-Atlas lead; Together AI at 1.25 per million input.

Apple's Machine Learning Research team published a real engineering report on October 15 showing how MLX and Thunderbolt 5 RDMA can tensor-parallelize Meta's newly open-weighted 816B-parameter Muse Spark Open across two M5 Ultra Mac Studios, hitting over 36 tok/s on a four-machine cluster.

A joint randomized controlled study from Carnegie Mellon, MIT, Oxford, UCLA, and UC Berkeley, covering three experiments with hundreds of online participants, finds that using an AI tool for as little as 10 minutes significantly erodes people's persistence and accuracy on hard problems: once the AI is removed mid-task, the AI group's performance collapses while the never-used-AI control group holds steady or even improves. Co-author Michiel Bakker, an MIT assistant professor, argues the right response is not a ban but a "tutor path" — AI that scaffolds learning rather than handing out direct answers.

System1-Arena: 30 AI models on 92k questions, 138 closed loops. Decision models match GPT-6 Luna, calibrate better; loop success drops 32.9% to 3.1%.

Meta open-sourced Muse Spark Open, which it calls the most capable US open-weight model: 816B parameters, 1M context, runs locally on two M5 Ultra Macs.

Microsoft-led LensErase removes any object, text or watermark from one image plus one mask, pretrained on 101M pairs for large gains on benchmarks.

KAIST's KV4-RL caches RL rollouts in INT4 KV with INT8-protected sinks: 1.80x speedup on Qwen3-8B GRPO, within 0.5pp of BF16 accuracy.

Tsinghua trains each document's KV cache at test time, weights frozen: 96.7% RULER accuracy at 32x compression vs 40.0% for KVZip+, code open-sourced.

ZJU's ATAS tops HF Papers: persistent scene and event memories keep AI video consistent across shots, leading all four backbones tested. MIT licensed.

Gemini 4 Argon tops DualEntry's accounting AI benchmark at 91.1%; open-weight GLM-5.3 Flash ties Grok 4.7 for second at 88.1%, at $0.006 per task.

NetBSD raised 120,282 USD in September, double its goal, by banning AI code since 2024 as Linux adopts LLMs. The surplus shows a paying anti-AI base.

Anthropic merged Project Glasswing into a three-tier Cyber Verification Program (Defense / Red Team / Specialized) with safeguards relaxed by tier.

xAI released Grok Voice 3.0 on Oct 14. Artificial Analysis rated it #1 on the Speech-to-Speech Index at 86.9 — ahead of GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6) — with 98.5% on Big Bench Audio, 97% on Full Duplex Bench, $0.08/min pricing, sub-second latency, and millions of daily calls already in production across Tesla and Starlink.

Linus used an LLM for 10 folio-lookup kernel patches. Hot folios defer release after RCU; data accesses under RCU skip refcounts. Order-of-magnitude speedup.

阶跃星辰 10 月 15 日开源 Step 5 Preview:600B MoE、27B 激活、1M 上下文,输出低至 2.7 美元/百万 token,仅 Kimi K3 的 18%。表格与 Kimi K3、GLM 5.3 在编程/Agent 上拉平,距 GPT-6 Astra、Claude Opus 5 一档。