{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidzq3qvrdbczrsoydhdls2x4up2rn77uodgtsodm3rjkaqdb2bqhq",
    "uri": "at://did:plc:ayfizlbmvprdzzpxhepnqrns/app.bsky.feed.post/3mhtvsd3snmq2"
  },
  "path": "/f/games/14729",
  "publishedAt": "2026-03-24T20:42:25.000Z",
  "site": "https://jstpst.net",
  "textContent": "I had a 700 words written about \"AI\" in games rendering, and how there are plenty of pre-transformer (pre-LLM) uses where it actually sincerely makes sense to use a neural network in physics and rendering pipelines, and all the cool potential for screenspace shaders implemented in convolutions.\n\nI'll post the draft below but I stopped because (1) LLM generated text makes it discouraging to write text and (2) LLMs make it discouraging to write about neural networks.\n\nDLSS 5 being \"trainable\" is not surprising (duh? any of these should be small-weights small-window shaders that don't take a lot of VRAM) and \"promptable\" (red flag klaxon blaring). \"Promptable\" means they have an LLM there, which, what? Lol? What the hell?\n\nlol",
  "title": "\"DLSS 5 will be trainable and promptable\"",
  "updatedAt": "2026-03-24T20:42:25.000Z"
}