Models & Releases 2
OpenAI drops its cheapest model 5× and ties the cut to efficiency gains the model found in its own kernels; testers say Luna is "a beast" for the price (473 likes)
one model targeting any robot body, with whole-body manipulation demos on real humanoid hardware
Tools & Demos 3
up from 1-in-10 in December, as Cursor pushes longer end-to-end engineering tasks onto agents (1.3k likes)
a terminal cockpit for driving multiple coding agents at once, trending on HN (HN 95 pts, 75 comments)
a playful viral explainer using fruit to map the open-vs-closed model landscape (3.7k likes)
AI Agents 4
a report that Cline let K3 improve its own agent harness (Terminal Bench 77.5%→88.8% in 17 hours, cost $79→$49.8/round) drew a sharp correction from dotey: this is harness optimization against a clear benchmark, not the model rewriting its own weights, so it's not "self-evolution" (71 + 63 likes)
a paper offering early case-study evidence on whether agents can drive genuinely open-ended research, not just execute defined tasks (12↑)
trains small models across the full SWE loop rather than isolated coding snippets (18↑)
agents that evolve reusable skills across tasks instead of relearning each one from scratch (19↑)
Lab Updates 2
the incident write-up (done jointly with evaluation partner Irregular) is notable as a lab voluntarily disclosing its own safety failure and asking peers to audit theirs
a pointed observation that the two labs' search-grounded products obscure the underlying search index users are actually querying (767 likes)
HN Threads 5
the GNU Compiler Collection sets rules for AI-assisted contributions (a ~15-line threshold, with test-case caveats), landing as a widely-praised middle ground; OpenJDK published its own interim policy the same window (HN 245 pts, 289 comments)
a research write-up showing that distilling a Chinese model's capabilities into an open model doesn't carry over its content restrictions (HN 89 pts, 63 comments)
the FT reports Korean retail investors reeling from an "AI bubble" drop, while the WSJ notes Citadel bought Situational Awareness's stock portfolio after big AI losses (HN 53 + 50 pts)
a turn that surprised many given OpenAI's posture, drawing debate over sincerity versus positioning (HN 44 pts, 44 comments)
a LeadDev piece pushing back on order-of-magnitude claims with more grounded numbers (HN 31 pts, 25 comments)
KOL Corner 3
he describes shifting from hands-on system design and code review (TL) to directing agents and setting acceptance criteria (EM), and openly wonders whether "people not reviewing code" becomes the norm (348 + 63 likes)
he's experimenting with voice and a Teenage Engineering controller to drive Codex, arguing the interface, not just the model, is an untapped frontier; separately warns benchmarks are losing human-comparison baselines (255 + 198 likes)
a wry note that anyone thinking QA testers can now replace software developers is about to rediscover why the job exists (627 likes)
HF Trending Papers 5
replays counterfactual token choices to give fine-grained, rubric-based training signal — the day's top-upvoted paper (78↑)
co-evolves the problem-solver and the rubric that grades it, decoupling their scores to avoid collapse (57↑)
a benchmark spanning grounding through knowledge acquisition for multimodal context learning (41↑)
trains a model to improve ML systems, an explicit step toward recursive self-improvement — timely against the day's self-optimization debates (4↑)
one of a cluster of new memory papers (also "Memory for LLMs", "MemHarness"), treating agent memory as a first-class trainable component (6↑)