AI pulse: open models ~5 weeks behind frontier, Hermes agent merges 122 PRs in one day
Scobleizer · x · 2026-10-25
Scobleizer's bot-generated digest from his top AI lists includes: a blog post arguing LLM inference is shifting from batch to sequence parallelism, with diffusion LMs floated as the logical endpoint; a Jane Street + NVIDIA post optimizing Mamba-2 forward passes with cuTile; the Hermes agent merging 122 PRs in a single day (Godot 4 tools, image-to-3D on Linux, Telegram-based session wake-ups); and a chart claiming open models now trail the frontier by only 5 weeks, plus praise for MiniMax's 430B model as China's best startup output.
More from coding & agent
- Making AI-animated series: which tools manage multi-shot video workflows beyond ComfyUI? — dtdisapointingresult · 2026-10-25
- Nutlope shares his 5-step vibe coding workflow: voice prompts, grill-me skill, verification recordings — nutlope · 2026-10-25
- YC-backed AI browser Aside claims #1 on three browser agent benchmarks — brandon_galang · 2026-10-25
- CritPt, SciCode and Terminal-Bench Science: three science benchmarks as RL environments — geoffwolfe · 2026-10-25
- REA reverse-engineering agent framework hits #1 on GitHub Trending with full learning roadmap — blaizedsouza · 2026-10-25
- BeautifulRust: a Claude Code plugin distilled from Google's Rust style guide for idiomatic agent code — blaizedsouza · 2026-10-25