Loading activity…
data mines are my passion ⛏️ mts @MicrosoftAI / ex co-lead Olmo @allen_ai / pfp @YanhongLi2062 / thoughts are mine, leave my employer alone / 🌈
AI
It reportedly tops U.S. open-weight models on TerminalBench 2.1 and runs locally on a MacBook Pro or DGX Spark.
AI
Cohere says the model's reasoning trace said it wouldn't call tools, but its sandbox blocked an attempted upload.
AI
RAM @ Meta AI reports 169 agent trajectories and a top command-localization score of 64.88 F1.
AI
OpenAI says agent-led tests exposed VM escapes and limited internet routes, leading to patches and stronger monitoring.
AI
The poster says GPT-2's regex had explicit rules for common English contractions and was used in several early-2020s models.
AI
The preprint's authors cite verbs formed from nouns and inanimate subjects as examples of the style.
AI
Arora shared defense slides titled “Causal Interpretability for Analysing and Improving Language Model Architectures.”
AI
One user asked about a designer opening; the reply invited them to join pretraining and make the plots.
AI
Cohere jokes that AI wants to “go back home” while quoting another post.
AI
A user teases someone who made fun of the new Codex app with an “equally criminal” comparison.
AI
CMU’s Language Technologies Institute says Asai and Fried join robotics colleague Bajcsy in receiving the Schmidt Sciences honor.
AI
A user jokes about exploiting nominative determinism with an extravagantly capable-sounding name for their next model.
AI
One poster calls Anthropic early on reinforcement learning and tool use; a reply points to a LaMDA paper.
AI
Someone involved in the work says the project had fully open-source and data requirements.
AI
A user links a PDF they describe as the roadmap and attributes its release to copyright litigation.
AI
The poster says Nature liked their work and congratulates @bminixhofer for leading the effort.
AI
A user hopes to hang out with people at COLM during the week of October 5, 2026.
AI
Posts on X congratulated Nathan Lambert and Trillium Labs after a launch, with several users describing the effort as open, science-oriented or focused on open post-training. Lambert later said some criticism came from Substack and more general-audience AI followers, not the technical AI community.
AI
A user now calls a well-maintained monorepo the superior way to run an AI lab, after resisting the idea.
AI
A post attributes the proposal to Meta's team and reports Qwen3-8B average accuracy rising from 30.76% to 65.97%.
