Frontier models improve at a pace faster than the pace at which most humans are able to keep up. Not all of us are obsessed with pushing the boundaries of our agents and trying new ways to work with them. I often spend 12+ hours a day working with an LLM, and still, when I talk to people, I hear of ways of working I hadn’t thought of. What if we could speed up that pace of learning? What if everybody shared their practices of how they are making the most of AI, and if we could learn from each ot
Nvidia just released a free speech-to-text model that can transcribe up to 40 languages, but it only works on recordings, not live audio, so it’s a big step for meeting and medical transcription tools while still leaving real-time use to rivals like ElevenLabs.
Hi HN, Like many of you, I spend a lot of time coding with Codex and switching between cloud and local models. As a software engineer, something I never understood, and honestly hated—was `/compact` and the way it’s designed. On local models my GPU screams, and after a few hours prefill takes 2 to 3 minutes per turn. On cloud models, as the context grows, the model starts getting noticeably dumber. IMHO the root problem is context bloat and the fact that compaction destroys the quality of the mo
I’ve spent the last year building context for AI analysts in the mid-market. I’ve open sourced a library from my learnings, along the same principles of the Matt Pocock skills. Principles: * Claude / Codex / Cursor etc are great data agents - you just need context to enable them to answer business questions correctly. * Much of the important context lives in analysts heads' * Context must be maintainable (short, human readable - markdown and yaml) * Added context should be load bearing, not repe
This is a project I've been working on for about 8 months now. I worked in legal tech for 3 years and found that one of the biggest hurdles for AI agents and lawyers alike when doing legal research is tracking changes, and correctly cross referencing those changes between different forms of legal documents. W2D builds a SQLite dataset from public sources, and offers agent skills for answering questions about that dataset. The dataset holds multiple revisions of the U.S. Code at each release poin
I was wondering how those classic demoscene effects come out in different languages. It’s interesting to see the required contortions for each language to get there. For example, SQLite has no float32, so it's emulated in SQL. Wasm has no sin(), so the hand-written port builds its own sine table, to name a few. It was built end-to-end by AI agents. Every port is checked pixel by pixel against the C reference by a script in the repo.