Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer ran an LLM-powered forecaster against Kalshi and Polymarket prediction markets for five months and found that every metric suggesting profitability was measuring the wrong thing. Directional accuracy of 60.9%…
OpenAI's GPT-6 Astra identified a Breakout-playing neural network from 17,253 raw floating-point parameters with 99% confidence, according to the pre-print "Sparks of Weight-Space Programming." Astra also generated a Bre…
QuickSilver Pro benchmarked three open models (Kimi K3, GLM 5.3, DeepSeek V4 Pro) against Claude Sonnet 5.5 and Opus 5.5 across 17 script-checkable tasks run three times each on 2026-10-20. Both Claude models passed all …
Baidu's open-source Unlimited-OCR-2 document parsing model reads multiple PDF pages in a single request, returning tables that span page breaks as one table instead of two, according to a hands-on test by Khuyen Tran, Se…
Adibvafa Fallahpour and six co-authors at Stanford released Memorilla on November 5th, a research system that compresses what a frozen large language model must remember into 16 query-dependent memory tokens, scoring 44.…
A 24 September 2026 arXiv paper reports that Nubank used the Snowglobe simulator to screen candidate customer-experience agents before deployment, cutting AI agent update cycles from 3 weeks to under 5 days. Simulation-g…
METR, an AI safety nonprofit, reported on November 5, 2026 that Anthropic's AI models are improving at roughly 1.9 times their historical rate, based on an 11-week evaluation from August 12 to October 29, 2026. METR attr…
Taskhold measured the context-window cost of its own MCP server at roughly 8,000 tokens across 22 tools, and cut it to about 7,200 by replacing the 180-character regex that zod's .uuid() writes into all 20 id fields with…
Tencent released Hy4, a 770-billion-parameter open-weight model with visual understanding and computer-use capabilities, on November 4, 2026, claiming a 9-of-12 scenario lead on its WorkBuddy evaluation and a finance sco…
A developer built an agentic loop on Claude Code that iteratively rewrites AI-generated text to evade Pangram's AI-text detector, using the detector's score as an objective function to hillclimb toward 0% AI. After five …
Researchers introduced PCAP-LM, a packet-capture notation that rewrites network traffic as compact symbol lines called PacketGlyphs, shrinking real 5G/4G captures about 800 times smaller than Wireshark's tshark detailed …
Thinking Machines introduced Inkling 2 Small, a Mixture-of-Experts model with 594B total parameters and 17B active, scoring 31 on the Artificial Analysis Intelligence Index and initially available through Tinker with clo…
AMD's ROCm InferaSim replaces most LLM serving-configuration measurements with fast analytical projection, measuring once per set of kernels and letting an agent search projections that each take under half a second on o…
A hands-on benchmark of generic matrix multiplication across CPU, Vulkan (wgpu) and CUDA backends found CUDA runs 7.82% faster than Vulkan, while Vulkan is 8% to 270% faster than CPU. The tester, who had access to only o…
Thinking Machines Lab released Inkling 2 Small on November 5, 2026, a mixture-of-experts model with 594B total parameters and 17B active, available on Tinker with closed weights and full weights promised under an Apache …
Check Point Research reported on November 5, 2026 that it identified 2,557 systems that hosted ARTEX panels, the open-source AI penetration-testing platform, with 1,608 live and publicly reachable at research time, and r…
ByteDance Seed published Cola2, a roughly 9B-parameter base model that plans text as continuous latents generated by a 4B-parameter diffusion transformer and decodes them into tokens, on Hugging Face, where it was spotte…
Vercel's AI Gateway added Ampersand (humansand/ampersand-v0.1), a language model from Humans& built for group conversations that decides each turn whether to speak or stay quiet, and is free on the gateway through Novemb…
Billion Context, an open-source proxy that compresses conversation history between a coding agent and the model API, has reportedly passed 3,700 GitHub stars and 300 forks in about three months by exposing compress, deco…
Billion Context, an open-source proxy that compresses AI coding-agent conversation history, cut costs 58% on a coding-agent benchmark running DeepSeek V3.1 Flash ($4.61 versus $11.05 for full context) while scoring an F1…