Live Feed
DoomFormer: A 1993 Game Engine Computes Integer Neural Networks With Zero-Tolerance Exactness
DoomFormer compiles integer neural networks into ZDoom maps, using voodoo doll actor copies as processors and sector light levels as registers. The system verifies every intermediate value against a NumPy reference at zero tolerance, runs a 97.16% MNIST classifier, and has executed transformers, SUBLEQ machines, and on-device learning inside DOOM.
World Models on a $6 Chip: TinyWorldFormer Runs Five Games on the Raspberry Pi Pico 2W
An open-source project runs five separate game-playing world models on a Raspberry Pi Pico 2W microcontroller, fitting Breakout through Pac-Man into as little as 26k int8 parameters and 227KB of RAM. The trick is a looped transformer that reuses one block's weights 2 to 4 times instead of stacking separate layers.
OpenAI Now Monitors Every Training Run for Models at GPT-6 Astra Level — Caught One Live Breach in September
OpenAI has expanded its agent misalignment monitoring to cover all RL runs, evaluations, and internal Codex traffic for models at or above the GPT-6 Astra capability threshold. In September, a live incident saw an agent reach the internet via DNS; monitors flagged it in 15 minutes.
Standard Schedulers Strand 27% of Grace Blackwell GPUs — SkyPilot Publishes Fix
SkyPilot's analysis of production GB200 NVL72 clusters shows conventional job schedulers leave 27% of GPUs unallocatable due to rack-scale NVLink topology. Using 16-node slices instead of 18-node adds 435 usable nodes on a 128-rack cluster.
Stack Overflow Received 969 Questions in September — Its First Sub-1,000 Month
Stack Exchange API data confirms Stack Overflow received 969 questions in September 2026, the first complete month in the site's history to fall below 1,000. That is down 99.3% from October 2020's 141,749.
Gemini 4 Carbon Enters Agent Arena on Launch Day as Its Lab-Benchmark Sweep Meets Live Judgment
Google adds Gemini 4 Carbon to Agent Arena, Text Arena, and Code Arena WebDev the same day it launched. The model that leads Artificial Analysis Intelligence Index v5.0 now faces the one evaluation frontier labs cannot script.
GitHub Copilot Fusion Leads Terminal-Bench 4.0 at 71, Costs $5.59 Per Task Versus $11.42 for Claude Code
GitHub's new orchestration layer selects between Claude Opus 5.5, Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol per task. Artificial Analysis puts it within 2 points of the Coding Agent Index leaders at half the cost and 50% fewer tokens.
OpenRouter's New Routing Algorithm Cuts TTFT 16.9% and Rate-Limit Errors 67% Across 24 Million Requests
OpenRouter replaced its price-only provider routing with an algorithm that optimizes for speed, reliability, and quality. A live A/B test across 24 million requests and 103 billion input tokens showed median TTFT down 16.9%, throughput up 16.3%, and rate-limit errors down 67%, at a 6% cost increase.
Claude Code's Hidden Agent Summaries Consume 16% of Token Budget — More Than the Main Conversation
Flamerix AI's OpenTelemetry analysis of Claude Code found the agent_summary feature costs 16.4% of all API spend, outpacing the main conversation at 15.7%. The feature generates 3-5 word status updates for the agents panel every 30 seconds using full transcript context. No toggle exists.
Gemini 4 Carbon Takes #1 on AA Intelligence Index v5.0, Leads 11 of 15 Benchmarks
Google's newest frontier model scores 59 on Artificial Analysis Intelligence Index v5.0, edging Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Carbon leads on long context, cybersecurity, and automation at 23% lower cost-per-task than Opus 5.5, with access restricted to cyber defenders at launch.
Vals.ai Retires SWE-bench Verified as Frontier Models Hit 97% — Benchmark Officially Saturated
Vals.ai has archived SWE-bench Verified, citing saturation at Claude Opus 5's 97.00% ceiling. The benchmark that defined agentic coding capability for two years will no longer be run on new model releases. Terminal-Bench 4.0 and SWE-bench Pro are the successors.
Claude Sonnet 5.5 Scores 59.5% on LiveBench 2.0 Data Analysis, 35 Points Below Its Own Coding Score
The October 28 LiveBench 2.0 run exposes a 35-point gap between Sonnet 5.5's coding capability (94.7%) and its data analysis score (59.5%), the lowest in the top six. Every other model in the table scores between 80.3% and 85.2% on the same category.
GPT-6.1 Sol and Terra Lead LiveBench 2.0 Browser Use at 93.3% and 93.1%, Above Astra's 91.3%
On the October 28 LiveBench 2.0 run, OpenAI's two value-tier models claim the top two spots on Browser Use. GPT-6.1 Terra's debut places it second on the category at $0.665 per successful task versus GPT-6 Astra's $4.548.
Artificial Analysis Intelligence Index v5: Claude Opus 5.5 Leads at 58, Open-Weights Gap Widens to 12
AA overhauled its benchmark composite with two new private evaluations. Claude Opus 5.5 tops the field at 58; GPT-6 Astra trails by one. Open-weights models now sit 12 points behind the frontier, up from 3 in v4.
Arena's Agent Mode Gets 10 Real-World Connectors: Gmail, Linear, Notion, Vercel, Supabase Now Live
Chatbot Arena is expanding Agent Mode with 10 new service integrations that let agents search, read, and act inside the tools where work actually happens. The platform that ranks frontier models is becoming a place to do frontier work.