Loading activity…
@scaling01
lead them to paradise Researcher at @EpochAIResearch LisanBench: https://lisanbench.com/ Impressum & Datenschutz: https://lisanbench.com/legal
AI
Anthropic's October 2026 acceleration was about 80% due to AI, METR guesses, though the estimate is highly uncertain.
AI
OpenAI says researchers used an upcoming model in an autonomous pipeline, at a total API cost of $1.2 million.
AI
Kimi K3.1-Pro's open-weights announcement links to a Hugging Face URL under moonshotai.
AI
The search compared predicted Doppler curves with a spectrogram, using flight data for candidate locations.
YouTubeAI
FlashProve aims to verify replies against a provider’s committed model without revealing weights, but proof generation takes substantial GPU time.
AI
AgentTime's team says GPT agents lead its updated evaluation, which combines duration following, runtime forecasting and retrospection.
AI
OpenRouter reportedly processed more tokens in the nine weeks before November 2, 2026, than in all its earlier history.
AI
Chinese models trail Mythos Preview almost seven months later, one argues; another estimates roughly two months if coding counts more.
AI
The v1.0 benchmark spans 11 domains and 206 tasks, with a score that averages performance across domains.
AI
Mythos 5.1 remains available to vetted cyberdefenders and life scientists through Anthropic's trusted access programs.
AI
Google's Argon 4 drew praise on X, while another user said it hallucinated badly in a real codebase.
AI
Cognition says its 132-task coding benchmark assesses whether maintainers would merge AI-generated code changes.
AI
Their post frequency and “shitpost rate” will likely decrease, scaling01 says; their views remain their own.
AI
Epoch AI says its fastest exploit attempt took 12 hours and the exploits cost about $1,000 per compromised version.
AI
Mistral announced the weights after a public preview, but its linked main Hugging Face listing showed “Upcoming release.”
AI
Unlike models with fixed tokenizers, H-Net learns to group raw bytes within the model, Cartesia says.
XAI
One poster suggests OpenAI has the best reinforcement learning—or that its models perform poorly without reasoning.
AI
A post says Carbon scored 88.2% on SEC-bench Pro and is rolling out first to cybersecurity specialists.
AI
Cognition says its 132 tasks were made by repository maintainers; its leaderboard puts Claude Opus 5.5 first at 45.5%.
AI
SemiAnalysis says it can reach over 850 tokens per second per user, even at 2.8 trillion parameters.
