Tag: Artificial Intelligence

Anthropic dropped their June 2026 Economic Index today and buried inside the survey data is something that should be making headlines: Over a third of respondents (9,700 actual Claude users, linked to real usage data) believe AI will be capable of handling most or nearly all of their work tasks within the next year. Not “some tasks.” Not “help me write emails.” MOST of their work. And here’s the part nobody wants to talk about: the people who delegate the most to AI are the MOST optimistic about their job prospects. Meanwhile entry-level workers are the ones most worried about displacement. Senior devs and managers? Thriving. Junior colleagues? Everyone in the survey is more worried about them than themselves. The data also shows AI autonomy is measurably higher on Claude Code than o...

Most of the open vs closed talk here is about whether you can run the thing on your own hardware. fair, that's the obvious draw. but the part i think gets slept on is that open weights mean you can actually post train on top of the base, not just run inference. With a closed api you're renting intelligence. you can prompt it, you can rag around it, but you can never make it yours. you cant fine tune the actual weights for your domain, you cant distill it down, you cant freeze a version and own it forever. You're permanently downstream of whatever the provider decides. I saw some post about people post training their own models on top of glm-5.2 now that its open weight, and that framing stuck with me more than the benchmark numbers did. a frontier-ish base you can legally build on changes ...

Solo project, sharing it here for the AI angle rather than the subject matter. I fine-tuned Llama 3.1 8B (QLoRA, single T4) on the complete works of a 19th-century author whose corpus is fully public domain. The interesting problem wasn't the domain — it was trying to get a small model to cite its source (book, chapter, item) on every answer instead of just asserting things confidently. What I learned, which might be useful to others doing domain fine-tunes: - Teaching the *format* of citation is easy. Teaching *correct* citation is hard. The model reliably produces "Source: [Book], chapter X, item Y" — and the concept is usually right, but the exact number is often wrong. It learned the shape of grounding without the precision. - That gap is exactly why I run the production version as RAG...