Live Feed
Google Open-Sources TorchTPU at PyTorch Conference: Native PyTorch Now Runs on Ironwood, Amazon Does the Same for Trainium
At PyTorch Conference 2026, Google and Meta released TorchTPU, making TPU a first-class PyTorch device. Amazon launched TorchNeuron for Trainium the same day. vLLM and SGLang get native TPU backends; Hugging Face Transformers v5 adds device_map='tpu' support.
DeepSeek Open-Sources DeepInfer: The Production Inference Framework Behind Its Commercial Services
DeepSeek releases DeepInfer on GitHub, the inference framework powering its large-scale production deployment. The package bundles six custom CUDA kernel libraries, expert parallelism load balancing, asynchronous scheduling, DSpark speculative decoding, and deterministic batch-invariant sampling. Initial support covers DeepSeek-V4.1-Flash.
Claude Opus 5.5 Has the Best Raw Score on HealthBench Pro but Drops to Sixth After Length Penalties
Artificial Analysis adds HealthBench and HealthBench Professional leaderboards across frontier models. Claude Opus 5.5 scores 71.4% raw but falls to 59.9% after length adjustment, losing more than 11 points. GPT-6 Astra leads the adjusted leaderboard at 66.0%. On general HealthBench, Claude Sonnet 5.5 tops at 62.1% while DeepSeek V4.1 Flash matches within 1.2 points at $0.005 per task.
NVIDIA Nemotron 3.5 Super VL Enters Vision Arena: 89.9% GPQA, 71.1% SWE-bench on One H200
NVIDIA's 120B/12B-active vision-language model joins Vision Arena, Image-to-WebDev, and Code WebDev on October 19. Strong on reasoning and coding benchmarks; tau2-bench retail, airline, and telecom scores all below frontier agentic baselines reveal where the capability doesn't follow.
AA-PersonalityBench: 36 Frontier Models Scored on 30 Traits Including Warmth, Conviction, and Slop
Artificial Analysis released AA-PersonalityBench on October 19, scoring 36 models on 30 pairwise personality traits rolled into 8 indexes. Claude Fable 5.1 scores 94th percentile for warmth. A dedicated Slop index penalizes hollow filler language. Frontier models split into two distinct personality clusters.
OpenAI's MultiLangBench: GPT-6 Astra Hits 87.4% Across 20 Languages, GPT-6.1 Sol Reaches 97.9% on COBOL Migration
OpenAI released MultiLangBench and a COBOL migration evaluation on October 19. GPT-6 Astra scores 87.4% average across 20 languages building a SQLite-compatible engine from scratch, up from 26.8% for GPT-5.6 Sol. GPT-6.1 Sol reaches 97.9% accuracy on COBOL migration at high reasoning effort.
VIDRAFT Darwin-180B-RSI Claims Open-Weight GPQA Diamond Lead at 94.4%
Korean startup VIDRAFT released Darwin-180B-RSI, a 180B open-weight model that tops MMLU-Pro at 88.1% and GPQA Diamond at 94.4% on AIhoster's independent board — results use majority voting over up to 16 samples, not single-pass evaluation.
Delta Won't Let AI Book Your Tickets. JPMorgan Won't Let AI Move Your Money. The Web Is Closing.
Yelp, Delta Air Lines, JPMorgan Chase, Charles Schwab, Amazon, and quietly Cars.com and Amtrak are blocking AI agents from their platforms. Cloudflare expanded its block settings on September 15 to cover Googlebot-class crawlers. Meta is partnering with retailers. OpenAI, xAI, and Anthropic have no clean answer.
Claude Wrote a Formally Proven Lean Kernel. It Runs 31% Faster Than the Official One.
Lean v4.35.0 ships con-leche, the first Lean proof-checker kernel written in Lean and proven consistent with set theory. Claude (Fable and Opus) implemented it under supervision at the Lean FRO. A companion kernel in Rust was then formally verified by Aeneas. The trigger: AI agents found two soundness bugs in the official kernel in August.
MiniMax M3.1 Flash Revealed as Space Bunny: 3x Faster, 70% Cheaper, Open Weights October 23
MiniMax confirmed that Space Bunny, the anonymous model that processed 200 trillion tokens on OpenRouter, is an early build of M3.1 Flash. Ships at $0.15/M input with 1M multimodal context and adjustable thinking.
Mistral's Timestral-209M Tops All Three Time-Series Benchmarks on CPU Without a GPU
Mistral released Timestral, a four-model family of time-series foundation models, with the 209M checkpoint ranking first on GIFT-Eval, TIME, and fev-bench over Google's TimesFM-3 and Amazon's Chronos-2 — running at 140ms per forecast on a standard server CPU.
Grok Voice 3.0 Takes #1 on Speech-to-Speech Index at 86.9, Runs Millions of Tesla and Starlink Calls Daily
SpaceXSI's Grok Voice 3.0 leads Artificial Analysis's composite Speech-to-Speech Index at 86.9, ahead of GPT-Live-1 Astra at 83.2 and Gemini 3.8 Live at 82.6, while already handling production voice traffic at Tesla and Starlink at an unchanged $0.08 per minute.
StepFun Delivers Step 5 Preview Weights on Deadline Day, Ships Technical Architecture Blog
StepFun published 600B open weights for Step 5 Preview on October 15 as promised, despite a brief false-obituary moment when a half-uploaded repository returned 404. The release comes with an architecture deep-dive and a commercial threshold license that activates at 100M users or $20M revenue.
Meta Ships Muse Spark Open: 816B Parameters, Runs on Two Mac Studios, US Open-Weight Record
Meta released the open weights of Muse Spark on October 15, delivering on a commitment that was uncertain at the model's September launch. The 816B-total, 42B-active MoE is natively multimodal and runs entirely on two Apple M5 Ultra Mac Studios.
Anchor Action Index Cuts Computer-Use Agent Token Costs by 100x
Anchor launched a shared memory of completed web workflows that lets agents skip the reasoning loop entirely. Instead of rethinking every click from scratch, agents replay a proven path as deterministic code — cutting token costs by 100x.