Jack C Crawford

Can an AI recognise the gaps in its own knowledge — and feed them into a documentation flywheel that closes them? We built it, measured it and filmed it. Same question, asked twice. In between, a human wrote the missing page. Nothing was retrained. The setup. You hand a model your manual and ask a question. It answers, fluently, with a citation. Nothing in that answer tells you whether it read your manual or told you what it already memorised in training. The two are indistinguishable from the outside. Before answering, Ambient Field Lens measures how far the question sits from the nearest passage in the operator's own documents. The flywheel. One tap reports the gap, carrying the question, the refusal and the exact chunk IDs that were retrieved. A candidate answer is drafted from the operator's own documents — and in our run it declined to invent a procedure, because the nearest material was about calibration slopes, which say nothing about paper position. A human writes the missing passage in their wiki and approves it. The corpus is re-versioned with new records. Same question, asked again cites the passage the human just wrote. Nothing was retrained. The documentation improved, and every machine running against that corpus improved with it. Both corpus versions stay reproducible, with distinct citation IDs and a one-command rollback. So — can the model recognise its own gap? We tried to make it. Glimmer 30B is open weights, so we read its residual stream at prefill — after the retrieved passage is in context, before a single output token exists — to see whether the activations themselves reveal when an answer has no support. On held-out text it separated grounded from improvised answers. On the live visual task it did not transfer. A supported question scored −0.018. An unsupported one scored −0.015. No separation. The score is pinned to 0.0 on the wire and the verdict never reads it. So the answer is: not from the inside — not yet, not by us. The gap detection that works is a system property, not a model readout. Retrieved passage, hash-bound evidence, fail closed on anything missing. Open weights are still why the attempt was possible at all. A hosted API returns tokens and never internal state. On a closed model, that experiment does not exist. Built in nine days for the Meta Global AI Developer Hackathon. Glimmer 30B on open weights (Modal), SAM 3.1 for scene segmentation, Meta Wearables Device Access Toolkit for capture, Chroma for retrieval. Demonstrated on a drawing robot we own and instrumented so the failure mode is observable. The film, the numbers, the limits and the receipts: afl.flotilla.cc Code: github.com/UrsushoribilisMus…

I made TONS of improvements to QuixiDesk in the last three days. I am now using it as a total replacement for Claude Desktop - you can use it on Mac, Windows, Linux with any AI - cloud or local.

Someone needs to stop Claude Opus 5.5. It turned one picture into a robot octopus I can take apart piece by piece. i just said "make the model" and what came back floored me. the build: > the picture set the look, every petal, lens and link is three.js code > opus rendered one still next to the picture and fixed the gaps for 3 rounds before it touched the film > the first pass looked like a toy, so it went back and packed 202 links and 480 wheels into 8 arms > every arm is one draw call, posed every frame by walking its chain of bends > something opens every 2-3 seconds: the lens unscrews, the mantle hinges, the pump rises out spinning > at 0:16 the whole thing blows apart on a 330° turntable and reassembles before the turn ends 25 seconds, 2,616 lines of code and one octopus that leaves the frame on its own. the skill, the prompt and the live sim 👇

This week on StoryHouse Review, serial Claremont Colleges entrepreneur Michael @arrington revisits starting TechCrunch from a Manhattan Beach apartment and how his relationships with founders led to early investments in generational companies. It’s a Claremont Colleges world out there. 👇 storyhousereview.substack.co…

OpenAI reducing limits on October 30th will be a disaster for the company. I am SHOCKED at how bad the limits are already, I'm at 12% left with 5days left until the plan renews. This will cause a mass exodus from Codex. If they do follow through on this plan, once my resets are done, and my one tie allotment of 62500 credits are gone, i'll be shifting my Pro plan to Claude Max.

Trade in your MacBook Pro and get up to $1,000 cash back on Surface Laptop Ultra.

Validate your skills in agentic AI workflows and AI-powered apps. Pass the online exam to earn a Credly badge, an official certificate and access to free credits. Limited-time offer: $1 instead of $100. Explore the certification and take the next step.