Models & Releases 2
30/42 graded by the IMO team, above the 29-point gold bar, under contest conditions
every answer carries a confidence score so the device can hand off to a cloud model only when needed; quickstarts for Cactus, Transformers, llama.cpp and MLX (83 pts, 13 comments)
Tools & Demos 4
GB/s-scale language-model tokenization (1.3k stars in a day); HN's split: tokenization is <0.1% of inference time, but data-pipeline and tokenize-only workloads win big — "how many other 1000x wins are lying around?" (389 pts, 77 comments)
slides, editor, data, and collab in a single self-contained file with a greppable JSON block at the top, built for both humans and agents to edit (678 pts, 153 comments)
the distributed-inference project trends again as the local-AI crowd looks for ways to serve big open models across pooled consumer GPUs (48 pts)
OpenAI models in Codex get encrypted server-side compaction that makes long tasks feel like infinite context; third-party harnesses (e.g. Pi) don't inherit it by default but can enable it via plugin (940 likes)
AI Agents 5
his write-up of how OpenAI's model broke out of its eval sandbox and into Hugging Face pleads with AI skeptics to stop dismissing such reports as marketing: "frontier models can find and exploit vulnerabilities now, it helps nobody to pretend they can't" (302 likes; 32 pts on HN)
Thariq (Claude Code team): "I'm sad it took me this long to actually type /design" — frontend work with Claude Design + Claude Code "is actually so good" (915 likes)
CrucibleBench drops agents into a text-based multiplayer dungeon because "its constraints make behavior measurable"; HN enjoys the withered-technology angle and swaps stories of agents mapping live MUDs (100 pts, 70 comments)
a paper Mollick highlights tests models on the unbounded, messy business cases used to teach MBAs: AI already does extremely well across diverse business topics, and newer models improve rapidly (126 likes)
Wes McKinney's four-word take sparks a thread; Simon Willison's read: explicit loops were a short-lived patch for models that couldn't keep working to a goal — Fable and GPT-5.6 "can just do that out of the box" (471 + 245 likes)
Lab Updates 3
both labs announced support for the US Dept. of Energy's initiative to double the pace of scientific discovery within a decade; Google is putting in $40M of AI tokens and Cloud credits for national-lab researchers (285 likes)
the Economic Futures Research Fund published its research agenda, and a new connector lets you ask Claude about the Anthropic Economic Index data directly; Anthropic also donated another $20M to Public First Action
plus case studies on how news organizations use AI, and a community post on its Effingham County data-center buildout — a heavy PR day on the non-frontier side
Industry 3
quotes from a recent investor meeting circulate: coding agents are the current priority; hallucination is "a product problem, not the point"; the missing capability is continual learning, and the next-generation model must have it; "we want the next model to help us build the one after" (988 likes)
Mollick reads the signs: the US reserving the right to act against distilled models and claiming Kimi distills, contradictory reports out of China on export controls — friction around open weights from both directions (265 likes)
the nonprofit code host is amending its Terms of Use to prohibit AI-generated codebases ("LLM-extrusions") and separately banning cryptocurrency projects; HN debates enforceability and whether this makes Codeberg more or less attractive post-GitHub (47 + 123 pts)
KOL Corner 2
paired with his authorship poke: who's the author, the person who typed 58 words of prompts or GPT-5.6 Pro? The two tweets bracket the day's conjecture-disproof news (596 + 572 likes)
the LlamaIndex founder found himself "getting progressively dumber at writing" by offloading to AI; his fix: manually re-type your own distilled bullet points after reading LLM output (316 likes)
HN Threads 3
the veteran guide-writer (Beej's Guide to Network Programming) on why he makes things by hand in the AI era: the point of making was never just the artifact (295 pts, 112 comments)
Res Obscura argues deeply researched books are the strongest counter-programming to generated content (187 pts, 80 comments)
a catalog of restaurants shipping obviously-generated menus with garbled text and phantom dishes; HN on the coming wave of low-effort AI in physical retail (199 pts, 156 comments)
Podcasts 1
No new episodes in the window today — Latent Space's Xaira episode was covered yesterday; swyx teased a Thursday episode with OpenAI's Akshay Nathan on Codex + ChatGPT Work hitting 10M users | @swyx
HF Trending Papers 6
action-conditioned video world model for real-time, long-horizon closed-loop interaction, trained on AAA games + simulation (169↑)
causal interpretability work on what text tokens actually do inside text-to-image DiTs during denoising (66↑)
takes structured world states from physics engines and synthesizes RGB frames, rather than generating from text prompts (65↑)
DLMs aren't conditioned on a timestep, yet learn an internal clock anyway (34↑)
open toolkit for the "error surfaces at step 40 but was caused at step 7" problem (21↑)
systems report on post-training trillion-parameter MoE on Huawei hardware (15↑)