{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiexyqu7mzazbgcmowjhfgz7rvyohgitmu73mgarjg4ide2yf3b25e",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mkbcxyl6mct2"
  },
  "path": "/t/a-g-i-and-you-how-much-of-your-own-work-can-you-automate/1379707#post_2",
  "publishedAt": "2026-04-24T18:24:23.000Z",
  "site": "https://community.openai.com",
  "textContent": "Heya VB, hope you’re well.\n\nI’ve actually stumbled on a task I cannot automate **at all.** I’m at hundreds of hours of conversation across multiple models—ChatGPT/Codex, 5.4, 5.5, ClaudeSonnet/Opus 4.6, 4.7. We’re writing a system ontology; and what we’re doing amounts to one very , very long philosophical conversation about JSON schema. I’m running out of context every day of the last two months… and I thought this would take a week tops.\n\nA few weeks (or months now) ago, I tried firing up some subagents to try hammering out the full ontology in a single step. We went through about five iterations from various [very detailed] prompts before it became clear that we needed to go one-by-one for a given “domain” of the overall ontology. We ended up with “a good starting place.” That’s it.\n\nNow, we’re at hours and hours of discussion for each DOMAIN schema. It’s actually very exciting to have a topic that is so challenging for frontier models and little old me.\n\nI have as much professional judgement on this matter as anyone—it’s really more the attention span for the work that would be lacking.\n\nWe have been producing (and working extensively on) a single skill / subagent pair, and paying minute attention to every phrase we create. At this point, we have over 50 Diff Coins—our term for a single JSONL that’s intended to train a model. So, we’re not formally fine tuning yet, and we have a dataset that COULD be used to fine tune, but I just have the model look through the extensive research before we begin work.\n\nWe are definitely **becoming more efficient** , but wow, I’m shocked. This isn’t “waste” conversation either, many of our steps might alter a definition slightly or need a fidelity sweep through the entire ontology, (like 5000+ lines of JSON Schema, and 7 documents with 7 more to go.)\n\n## Automation and Augmentation\n\nThis task is definitely “ai augmentation.”\n\nGiven the experience with it, my first real try at fully annotating a dataset for fine tuning, for example—I am going to be very wary of any automation outside of my expertise.\n\nThat said, i also think one of the main barriers between Automation and Augmentation is the lack of precisely this sort of Ontology.\n\nI’m working under the hypothesis that a completed ontology (a json structure that shows how everything conceptually fits in a system) will greatly enhance efficiency and all other attempts at full Automation.\n\nThe actual results of an AI building with this ontology “in mind” have been very promising.\n\nBut if my experience is typical of people designing ontologies, then the barriers to entry with regard to thought, time and compute are considerable.\n\n## The Hard Part\n\nClearly this is the hard part. lol I’m still shocked by “how hard,” even as I become increasingly “pleasantly surprised” by how previous moves are starting to compound into “model experience.”\n\nBut, the “extra” work, geeze. I’m working in a limited range of documents, and it’s STILL hours reading, and re-reading stuff a model wrote, or I wrote, or was added or appended to some concept—and rewriting language so it matches exactly what I have “in mind”, which is absolutely necessary for avoiding down-stream hallucinations which will immediately propagate.\n\nWhat about annotating the data set? Geeze. What about automating the dataset in the first place?\n\nBut all of this would be called \"review, correction, integration, accountability, domain knowledge, outside of domain knowledge, taste, edge cases, research for both familiar and unfamilliar topics, integrating new things (like the Agent SKILL protocol, MCP, or A2A just this morning), designing skills and subagents organically while we work (which has been one of the greatest assets so far), annotating data.\n\n## Cost Subsidization\n\nWhy, thank you for asking! I actually think this type of work would be very worth subsidizing because it’s “compute and attention” heavy, but we’re ultimately trying to make downstream operations more efficient.\n\nSo, in theory, I would be more willing to pay full price for compute with a [philosophically sound] ontology helping the Models understand “where things belong”. To be perfectly frank, a single turn with Codex GPT 5.5 Extra High (thinking) burnt my 5 hour context limit all at once.\n\nThough, you know, as I think about it, what about subsidizing “non-wasteful workflows.” Like, are you making full use of your conversations, outputs, harvesting mistakes as a dataset for small model training (as you go): Give that dev a subsidy!\n\n(I think the term “tokenmaxxing” is embarrassing. If you’re making efficient use of compute, and running out of tokens, that’s one thing. But if you’re just burning compute to pad a statistic…)\n\n## Practically\n\nMy goal is to work with businesses to develop their own ai-first system. I can say, without a doubt, that the hardest part will be getting the experts to sit down and annotate their own datasets.\n\nI just throw up my hands and tell them the accuracy WILL ONLY come from them doing THEIR homework (no way am I annotating a set whose topic I don’t understand. Nope. Not after all this. I’ll build it, I’ll make it easy to annotate, but annotate myself? Nope.)… but I’m betting that if they have a complete ontology to start with, “full auto” will be achievable far sooner.",
  "title": "A(G)I and you: how much of your own work can you automate?"
}