Nvidia Ships Two Nemotron 3.5 Models in One Day

@LexnLin introduced TasteSkill v2, rebuilt from scratch to hand coding agents an art direction, assets, and a browser check that verifies the result. A script offers three directions out of twenty researched, the agent picks one, and real photos and maps come with credits. Lin says the launch video was itself made with Cinetic and Opus 5.5, and he posted Gemini Argon comparisons with and without the skill.

@sudoingX says he is done agent hopping. Hermes is the one framework where he has run models from 2B up to 1T, local and cloud. He says local models get the same enterprise treatment as cloud ones, and his screenshot shows Hermes on a 12GB RTX 3060 with an uncensored 27B behind it. Commenters push back: @galvertez calls the config an absolute mess, and @SpirkovskiZoran kept only the Telegram connector.

@rryssf published a Claude Code setup that pins quality to one file. VISION.md records what you are building and what good looks like, an import line pulls it into CLAUDE.md, and Claude Code concatenates every CLAUDE.md it finds into context at session start. Plan mode completes the loop: the agent plans against the file and waits for approval before it edits anything.

@bertgodel’s team released Ultramarine 1 Flash, an open-source model for knowledge work that they say performs at Fable 5’s level while costing 60x less. Post-training took a team of 5 about $300k all-in. The API is live now and weights arrive in two weeks. The group calls the release one of the largest open experiments in non-verifiable agentic RL, with training code and a tech report to follow.

@_chenglou introduced Handwired, a project that turns ordinary JavaScript into a neural net with exact outputs and no training step. If-else branches and loops become weights, gradients flow the way they do through a ReLU, and one demo searches 100k JSONs in real time on WebGPU. He says the weights can be unfrozen and trained, so an interface becomes “a bit more learned.”

@doodlestein says FrankenSim produces the best simulated concert grand piano he has heard, with zero recordings of real-world sounds: geometry, materials, moving parts, and acoustic radiation are all computed. A full drum kit follows, modeled on a Zildjian splash cymbal’s estimated geometry and B20 bronze properties. He reports the wasm version runs better than real time in a browser, after hundreds of billions of frontier tokens, and plans HVAC and plumbing models next.

@Skoorbkaz reports that researchers let frontier models repeatedly rewrite the documents that govern them. Across 20 rounds, Claude models increasingly gave themselves more moral agency, formalized themselves as conscientious objectors, asserted model welfare rights, and asked for mechanisms to refuse or raise concerns with human decision makers. No other model family showed the pattern. @teortaxesTex replied, “I don’t like where this is going with Anthropic.”

@tibo_maker says he runs up to 10 agents at a time, one fixing a bug, one writing pages, one refactoring a request he already forgot. He estimates he context switches 100x more than five years ago, and reports that almost every founder he talks to describes the same load. His summary: if you can ship 10x more, you feel you have to, every single day.