Reddit post predicts Qwen-3.8-FN agents will sweep ARC-AGI-3 Kaggle

A Reddit poster says Qwen 3.8 FN solved a work problem requiring local-model privacy with the same skill as a frontier cloud model. The task involved tracking students and supervisors for an MSc dissertation module.

Community inference engines and quants run the 125B–177B MoE model on setups as small as 8 GB VRAM plus 32 GB RAM, with one report hitting 900 t/s prompt processing and 30 t/s generation. capy-flash v0.7.1 runs the 125B model in 64 GB RAM at 234 tok/s on two 16 GB cards.

BNY Global CFO Dermot McDonogh said he is monitoring the bank's AI spending to avoid overspending on related tools, in an interview with Bloomberg's Paul Allen and Haidi Stroud Watts. He also discussed the bank's growth and business strategy outlook.

OpenAI released dots at its late-September developer day; it's available in ChatGPT desktop and mobile to Pro-tier subscribers and up, with plans starting at $100/month. Enterprise product lead Alexander Embiricos said much of dots' functionality was already reachable via the Codex coding app, but required more hand-holding.

The New York Times Style section profile, by Alexandra Alter, centers on Mayer struggling with a task for her daughter Marielle's 10th birthday before turning to her new AI startup Dazzle. Daring Fireball's John Gruber criticizes the piece's framing of Mayer.

A growing patchwork of AI security benchmarks each measures different variables, and none substitutes for testing models against your own code and workflows, Dark Reading reports.

A developer who started with ChatGPT and moved to Gemini 1.5 Pro says they regret not trying Claude sooner for AI-assisted coding. The post describes directing the model on functions and design systems rather than pure vibe coding.