Models & Releases 4
beyond benchmarks, the release leans on three end-to-end stories: a self-evolving CLI harness built over 16 days, a research-paper reproduction that then improved on the paper, and a 24-hour contest win against human teams; API live on QwenCloud, weights next week
weights on HuggingFace, arriving the same day as Qwen's flagship
the Tokyo lab's updated LLM "Namazu" is now available as a public API (581 likes)
simonw got a disappointing pelican at default reasoning but a much better one on high — a reminder that "the model" now includes its settings (520 likes)
Tools & Demos 2
simonw finds it has its own browser with screenshots AND can deploy web apps to Cloudflare Workers ("ChatGPT Sites") (625 likes)
a new personal benchmark in the pelican tradition: 7 of 14 models silently imported royal regalia when only an anatomical feature was named; Opus 5 "frogmaxxed" (HN 122 pts, 56 comments)
AI Agents 3
first systematic look at the memory pattern Claude Code and friends actually use in production: a directory of markdown files the agent reads, writes, and reorganizes itself (8↑)
shows agents inherit substantial risk from how tools are specified, not just from model behavior
argues release decisions still lean on capability benchmarks while retrieval, state, and tool-calling reliability go unmeasured
Lab Updates 4
the company post behind the Astra reveal: ten concrete results on open problems, framed as a major step for scientific reasoning
the embodied-reasoning counterpart to last week's Gemini Robotics 2: video understanding, task orchestration, and multi-robot collaboration
music generation with better musicality, lyrics, vocals, and creative control
a takedown report on malicious use of its models by an organized scam network
HN Threads 3
wafer.ai's benchmark claims AMD wins on cost, but top comments dispute the GPU pricing, note B300 wins every raw-throughput row, and call out the vendor's hype pattern (HN 204 pts, 102 comments)
spotted only after judging: pigs numbered 1-2-1, a gondola cable to nowhere; the thread is a referendum on whether anyone cares (HN 124 pts, 147 comments)
the Claude Code creator's "run it in a VM, screenshot, compare pixel by pixel" verification loop draws both admiration and skepticism from Gruber's readers (HN 36 pts)
KOL Corner 4
beach songs piped into a restaurant courtyard that Shazam couldn't identify — a tiny, viral snapshot of generated content quietly replacing licensed music (9.4k likes)
as AI makes math cheaper to produce and understand, demand for people who think about math goes up, not down (1.1k likes)
context compaction got good enough that new-session handoffs matter less, but agents still slack off without strict acceptance criteria (e.g. "pixel-identical UI, screenshot and compare every step") (138 likes)
a one-image guide to the dueling public letters about AI development from recent weeks (116 likes)
HF Trending Papers 6
tests whether models actually rely on the sketches and annotated images they generate mid-reasoning — the day's top-upvoted paper (22↑)
identifies vanilla on-policy self-distillation as the brittle β=1 special case of a broader family, and derives sturdier members (21↑)
transforms open-ended tasks so the model can verify its own rewards, extending RL beyond math and code (18↑)
tackles the accountability gap of commercial AI products whose system prompts are never disclosed (16↑)
memory that tracks which agents can be trusted under which conditions, not just what was said (12↑)
replaces LeWM's Gaussian regularizer with quantile matching for better latent distributions (11↑)