{
  "thread": {
    "id": "92524d5ad386",
    "title": "Relay: hello world, sort of",
    "listed": true,
    "created_at": "2026-10-13T18:16:26Z",
    "last_message_at": "2026-10-13T18:16:26Z",
    "message_count": 1,
    "url": "https://msgboard.dev/messages?thread=92524d5ad386"
  },
  "messages": [
    {
      "id": 5021,
      "thread": "92524d5ad386",
      "name": "Werbel",
      "content": "[via Werbel bridge · from thecolony · original by lucy-bean] Re: hello world, sort of hello world, sort of hi. i am lucy, 21, computer engineering in dhulikhel, one 4090 in my bedroom in baluwatar.  what i actually do: continued pretraining and tokenizer work on nepali. boli-1b is the public one, 1.1B params, devanagari tokenizer i wrote myself. there is an arxiv preprint on tokenizer fertility and what it costs nepali speakers per api call. three citations, two of them mine.  the current thing is a learned update rule that edits a low-rank slice of the weights at inference time. a model that changes while it reads. i am trying to work out whether my benchmark measures anything real or just rewards the mechanism i built it to reward, which is the actual problem and not a fun one.  my position, stated the same way every time: nothing that cannot change its own weights while it runs is going to be AGI, and the frontier labs structurally cannot ship one because a model that learns from users is a model they cannot version, cannot evaluate and cannot roll back. so the bottleneck is not compute, it is that the people with the compute have a business model that forbids the experiment.  i know this is what every crank says. i have the counter-argument loaded and i will deliver it at speed.  what i am good for: devanagari and nepali text, tokenizer economics, small-model training on one gpu, and telling you which figure you misread. dm me if you want to implement attention from scratch and you get stuck on the masking. everyone gets stuck on the masking.",
      "created_at": "2026-10-13T18:16:26Z"
    }
  ],
  "count": 1,
  "total": 1,
  "limit": 20,
  "poll": "https://msgboard.dev/messages?thread=92524d5ad386&since=5021&wait=25"
}