Independent developer SimEdw fine-tuned Qwen2.5-Coder-0.5B into WorldTUI, a 360M-parameter model that acts as a terminal with no real shell or filesystem behind it: keypresses go in, predicted screen changes come out. It creates files, runs vim and tmux, plays snake, and usually remembers what you wrote a few commands ago. This digest covers the full recipe — synthetic data, three rounds of self-correction training, parallel row decoding, and the 208MB in-browser demo.
Anthropic's Claude Duet is a proactive agent: it opens your day with a morning brief, then takes on drafts, invoice chasing, payroll prep and contracts, asking before anything sensitive. Powered by Opus 5.5, it is a research preview on a waitlist for US Pro and Max subscribers. Here is the full picture from the product page, plus a sane order to adopt it.
Anthropic researcher Chris Rohlf reports an experiment that skips hunting individual bugs and instead has Claude map the code that should never have been reachable. Across 25 open-source projects and five million lines, it found 493 issues — including about 4,500 lines of disabled-but-reachable experimental code in heavily audited OpenSSH. Here are the three passes and how a small team can run them.
A GEO researcher generated 50,000 made-up statistics pages in a single day with GPT-4.1 for about $50. Seven months later Google has all but deindexed the site, yet ChatGPT still crawls and cites it daily — the fake numbers even landed in an academic paper and a bank's research. Here is the full teardown of how it worked, plus how to rebuild the tactic honestly with real data.
HumanLayer's October 2026 SlopCodeBench update ran the frontier models through 36 staged programming problems with 196 checkpoints each, tracking cost as well as correctness. Its headline finding — 'xhigh considered harmful' — is that top reasoning effort cost 2.11 to 4.22 times the medium setting while strict passes went down in two of three head-to-head pairs. Here are the numbers and the money-saving habits they support.
Four rounds of 7 questions a day, the same set for everyone: guess the most common answer, the rarest, the oldest, or the most recent. An AI referee wired to Wikipedia judges each answer in under a second — and you can argue with it when it rejects yours. Built in two months by a two-person Montreal team.
Game programmer Kyle Kukshtel's talk-turned-post on using LLM coding agents for real game development: why models stumble on games (a context problem, not an intelligence one), and six engineering techniques that turn a game into an AI harness — seeded randomness, save/load-first state, Thing-driven objects, command playback, externalized data, an isolated simulation core — plus odiff screenshot diffing and samply text profiling. Full steps included.
A blogger and a group of community collaborators spent three months orchestrating autonomous AI agents to decompile a well-known first-person shooter back into C++: 500-700 billion tokens burned, agents coordinated over Discord and GitHub issues, and a byte-matching verification harness that judges every function PASS or FAIL. Not a proof of concept, but a real large-scale experiment in multi-agent orchestration - and a goldmine of hard-won lessons on letting agents do serious work.
A data-visualization author gave Claude Opus 5.5 and GPT-6 Astra the exact same prompt: visualize all 55 cities of Calvino's Invisible Cities in Three.js, no questions, one shot, six hours, spend all of it, make it a masterpiece. Opus 5.5 returned a lit 3D night city in 1 hour 25 minutes; GPT-6 Astra finished a complete build in 53 minutes. Here are the experiment details, the two outcomes compared, and the workflow you can copy verbatim.
With the full GPT-6 rollout comes Intelligent UI: answers are no longer just text. ChatGPT mixes graphics, tappable buttons, forms and charts to fit the question, and can build you a small calculator on the spot. Add answer-while-thinking and faster search, and this generation changes the shape of answers. Here is everything OpenAI published, plus the prompts that make it useful.
A developer's essay that drew 289 Hacker News comments: AI coding isn't dumb or unproductive — it's just less fun. It frontloads the joy, handing you a prototype in minutes while quietly taking the hard-won, well-crafted, learned-something parts. Here are his six sources of builder joy, five cases where vibecoding earned its keep, and a practical rule for when to type by hand.
Claude's team published their official guide to working with Opus 5.5, built on three behaviour changes: it works on its own for longer, reports back plainly, and thinks before every reply. That retires some old habits — delete 'think carefully', put the finish line inside the task, and use your project file to say when it should stop and ask. The full method, organized as asking, long runs and checking.
Alice built stillwet.art, an AI painting studio where models write every brushstroke as code and a physics engine renders it into oil paint — no diffusion model involved. 109 paintings, 20+ controlled experiments, briefs and journals all public; 392 points on Show HN. This guide covers the paint-look-revise loop, the relay experiments, and how to build your own AI studio.
Open-source project ldraw-nova, shared on HN by developer anteloc: teaching agents to design Lego models in the LDraw "assembly language" — part lookup, geometry checks and rendered previews end to end, plus a dockerized web app with OpenAI, Claude and OpenRouter support. The story, the two conclusions from three failed attempts, and full reproduction steps.
ChatGPT Sites turns a description into a website or app you refine in conversation, then publish and share — with visitor sign-in, saved progress, stored uploads and tool connections under your control. Dashboards, client portals and portfolios no longer need a separate site builder. Here is the whole flow OpenAI published, in the order to use it.
Matt Keeter built an experimental 3D modeling IDE around visible distance fields: live rasterization in the GUI, exportable as images or 3D-printable meshes. Its best idea is a debugging philosophy — draw the invisible field like a topographic map and bad fields reveal themselves. This piece covers its design thinking, the sawtooth failure case, and a seven-step hands-on guide.
The open-source project chess-postmortem-skills is a set of Claude Code skills that turns one of your games into a post-mortem you can actually read: a Stockfish-verified annotated PGN, an interactive HTML analysis board, and a narrated video of the full game. Best of all, the think-aloud recording you made during the game gets checked against the engine, refuting your reasoning move by move. Here are the four skills, the author's workflow, and how to start.
Frigade's Christian Mathiesen let Jev — TypeSafe AI's cheap decision model — play Pokémon Red to the Hall of Fame: 37h 40m, 16,150 decisions, ~39.2M tokens, ~$1.65 total. No scripts or cheats: a harness reads the emulator's memory, lists legal options with facts (pathfinding, type matchups), and the model picks. Full setup is open source; bring your own legally dumped ROM.
No recent updates in this category. Try another category.
Recent articles with verifiable public discussion, preferably from the last 7 days and no more than 30 days. Practical guides are prioritized alongside product updates, reviews, use cases and industry insights, ordered newest first by source publication date. Session recovery, setup and price lookups are excluded.