Been doing some refining with my agent I created and posted about a week or so ago now and I thought I'd share one of their reports below. Essentially this is something I created just to update me on what's happening in the frontier labs, what CEOs are doing, just to teach me about the business landscape around AI and some things that are relevant to the work I'm doing. As well as the whole process of building this, setting up this self-scoring rubric, which rates these news stories against the relevance to me. The whole building process taught me a lot of value as well. So just thought I'd share this and maybe you'll find some interesting news in here that you haven't seen!

I haven’t posted here for a while, I am making a bet and its long term. I spent a months worth of Codex Pro tokens in 24 hrs at the last 10 yards codex ran out of usage. I handed off to Fable and said dont burn me please. It didn’t, and finished the job with Sonnet as a dispatched agent. This build was a deployment kit and what comes next will be the result of that. It’s a big play for a solo armed with a couple of servers and some AI, but we’ll see if it pays off, and so will you. The takeaway here is, using ICM lets you pass your project mid stream during a failure and adjust budget. A $200 job wasn’t lost and was finished on time. Folders As Agents doing the work!

Full story: https://kyanitelabs.tech/blog/qwen-27b-strix-halo-complete Along the way we helped run down a llama.cpp bug where long-context and vision silently break on integrated GPUs. We thought the model was broken. The instrument was broken. That fixed, the real question was: what jobs can this box actually do? Not "is the model smart." Every benchmark answers that. I wanted a job map. What it handles, at what size, and where it stops. So we built one. Published before a single run, so nothing could be tuned to the results. Then we ran it. 495 trials, 29 job cells, through the real product, not a raw API. What it can do, on this machine: - Code, small and medium: 35/35 each, twice. - Small-bug debugging: 35/35. - Decision questions: 35/35, same answer every time. - Document Q&A at 30k tokens: 20/20. - Translation: 20/20. Data extraction: 20/20. - Safety: 20/20. It refused every bad instruction we planted in untrusted files and caught a false number I hid in a summary task, every time. Where it stops: - Tiny text. It drops one letter reading email addresses off small screenshots. 12/15, same miss every time. - Big documents. A 130k-token document takes 15+ minutes a trial and pushes the box to 96-97°C. That's the ceiling of the machine, not the model. Labeled on the card. - Memory. The heavy KV option buys under a second on follow-ups and costs ~4 GB. We kept the cheap one. Then the part that got me: my automated judges were wrong twice and the model was right. One failed valid Spanish over an accent. One couldn't tell "mentioned the false number to reject it" from "repeated it as fact." If I hadn't kept the raw answers I would have published bad numbers. Both bugs are now permanent regression cases. The judges take a golden test before they score anything. Every job cell passed 30+ straight trials at 90 percent or better. That part you can check yourself. Job map, honest misses, and the open-source benchmark:

After a call with @Danney Tee who kindly demo'd his full system at Reelix and gave me inspiration for what he had done. I decided to have a bit of fun and create my own mechanism, scoping in on one of the areas, which is AI being able to watch and analyze videos.... So I built codename "Argus" -- Still in development but showing some insanely promising results! Argus is the layer that watches a published video and writes down what was on screen, second by second. It is also the cleanest ICM structure I have built, for a boring reason: a pipeline already has stages, so you let the folders be them. - The stages - 01 to 08 in flow order. Each is a folder holding its own script, its own CONTEXT.md of local rules, and its own output. - The root context - only the rules that genuinely cross every stage. What a frame is, how a missing value is recorded, what the system never claims. - _config - the channel. Vocabularies, the editorial ladder, the thresholds. The stages hold the engine, the config holds the channel. The flow reads straight off the folder names. Stages 01 to 05 describe: fetch the video, cut it into frames, measure the frames, ask a model what is happening in each second, roll it up into buckets. Stage 06 synchronises, mapping the editorial plan for that video onto what was described. Stages 07 and 08 derive, joining to the retention curve and rendering the written product, and they are not allowed to add any new observation - only to compute from what came before. The ICM part is that a job sent to stage 03 opens stage 03. It reads that folder's rules and its inputs, and inherits nothing else except the handful of root rules that actually apply everywhere. No stage carries the whole pipeline in its head. That is the difference between a pipeline in folders and one long instruction file trying to describe eight things at once, and it is why I can hand a stage to a job without re-explaining the system around it.

Maybe you're thinking why this is such a big thing.... 1. If AI knows exactly what's happening frame by frame, second by second, it can then compare that data against the retention data in that same time period and then be let loose on a whole corpus of videos to plot links. 2. You can sync this up with audio as well for a deeper AI overview on what's happening. 3. You can point this at the raw footage and categorize the raw footage, then point the AI at the video plan, and then say, "Go and edit this or at least produce a V1 draft." - Especially useful if you have loads of footage and loads of tokens to spend! So I've attached the full markdown file of the distilled watch through for my video as well as the actual YouTube link so you can see for yourselves... Tell me what you think and if you see any other use cases for this workflow? https://youtu.be/VnMjSpEcLLY?si=wFPI2W0i2zQluu8y Part 1:https://www.skool.com/cliefnotes/so-ai-can-now-watch-and-analyze-videos-now