First post here, so be gentle, lol. I'm having a hard time wrapping my perception about what kind of data you can, or should put into yours for your AI memory. I've been messing around with computers and by extension data manipulation, since the 70s. I have a good understanding of how a relationship database runs. But we're not building that kind of ecosystem(?) are we? We can go bigger. Before I dug deeper, I always picture an LLM like Chatgpt, as having this huge massive brain, which held all of the Internet, and when I asked would wave it's virtual hands to say "Here it is". I know that's incorrect. I'm using AI as a research assistant and junior co-writer. I'm doing a non-fiction book what kind of skills you'll need to make money in the next 20 years. I'm doing a ton of digging for trends and possibilities and...all of you have a good idea of what that means, I'm sure. I went into this thinking it would be more like a Wiki compiled form all my research and conclusions but is that what I want? Seems like there so much more than just a glorified book list. I do want to have a folder style system, with the full transcripts, complete articles, or other important documentation that I need. That's doable too. But recently, I wanted to start at least collecting the base data. If I don't start doing that, I'm just digging my hole deeper. So I laid down a basic schema, and then asked GPT to pull me 5-6 highlights that it thought should go into the database. Once I get the workflow built, I pretty sure I'll try to automate it, so as I research, the AI formats and stores any highlighted data I come across. It pulled those six, then another three from our discussions on the first six, and we were in a side chat at the time, the main chat was on markdown files. We got 10 entries from that. So I had nineteen. I haven't done any more, but when I look at what I have, I get this weird vibe that the majority of them aren't on the book subject, but are more of how the LLM views the way I work? That's a poor description of it, I hope it works.

An excerpt from last week's VIP that I felt was important to share with everyone! How to get your first check/increase customers and bring in cash from the FOUNDATION of business.

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:

I have a client in Australia who's looking for someone who understands ICM and can help him work on his software. Obviously he would prefer someone a bit more technical, but he's happy to have some people who are trying to learn as well. It's a pretty large software, but he's created a pretty good automated system and obviously he's using my methods so it's very well organized. Anyone out here looking for a project to take on the practice or learn more? Or is there anyone in the morning to get a little bit more work done at a higher level? He's looking for multiple skill levels.

It’s live. Starting today, everyone in Clief Notes has access to the new Clief Notes AI. The easiest way to use it? Don’t overthink it. Ask it the question you would normally ask me. “Where should I start?” “What should I focus on next?” “Where did you talk about [topic]?” It’ll use what’s already inside Clief Notes to help answer you and point you toward the right lesson or resource when there’s something worth going deeper on. The goal isn’t to give you another AI tool to play with. It’s to make everything already inside this community easier to actually use. If you want Access to it Comment "READY" and we'll send you access to it!