It can be a small skill.
The last thing I learned to do was whistle. Never could whistle my whole life, and tutorials and friends never could help me.
So, for the last month or two, I just sort of made the blow shape then spam-tried different “tongue configurations” so to speak – whenever I had free time. Monkey-at-a-typewriter type shit. It was more an absentminded thing than a practice investment.
Probably looked dumb as hell making blow noises. Felt dumb too (“what? you can’t whistle? just watch”), but I kept at it like a really really low-investment… dare I attract self-help gurus… habit.
Eventually I made a pitch, then I could shift the pitch up a little, then five pitches, then Liebestraum, then the range of a tenth or so. Skadoosh. Still doing it now lol.
(Make of this what you will: If I went the musician route my brain told me to, then I would’ve gotten bored after 1 minute of major scales. When I was stuck at only having five pitches, I had way more longevity whistle-blowing cartoonish Tom-and-Jerry-running-around chromaticisms than failing the “fa” in “do re mi fa”.)
So, Lemmings: What was the last skill you learned? And further, what was the context/way in which you learned it?
I got olama and WebUI working privately / locally and I’m able to insert documents into it with persistence and query them.
Nice, AI with half of the suspicion removed.
Does it save you a lot of time, what do you use it for? I have a somewhat old GPU but have been considering something like this to comb manuals. Does it have a file size constraint?
I have two projects for it right now. The first is shoving my labyrinth of HOA documents into it so I can answer quick questions about the HOA docs or at least find the right answer more effectively.
The second is for work, I shoved a couple months of slack, some Google docs, some PDFs all about our production product. Next I’m going to start shoving some of GitHub in there. It would be kind of nice to have something that I could ask where is the shorting algorithm and how does it work and it could give me back where the source code is in any documentation related to it.
The HOA docs I could feed into GPT, I’m still a little apprehensive to handover all of our production code to a public AI though.
I’ve got it running on a 2070 super and I’ve got another instance running on a fairly new ARC. It’s not fast, But it’s also not miserable. I’m running on the medium sized models I only have so much VRAM to deal with. It’s kind of like trying to read the output off a dot matrix printer.
The natural language aspect is better than trying to shove it into a conventional search engine, say I don’t know what a particular function is called or some aspect or what the subcompany my HOA uses to review architectural requests. Especially for the work stuff when there’s so many different types of documents lying around. I still need to try some different models though my current model is a little dumb about context. I’m also having a little trouble with technical documentation that doesn’t have a lot of English fluff. It’s like I need it to digest a dictionary to go along with the documents.
HOA docs didn’t even cross my mind, that’s resourceful.
Has the AI been particularly accurate, and does it cite where it found your information? With more technical stuff it’s always confidently wrong
ty for the response btw
It tells me what document in the collection it used, But it doesn’t give me too much in the way of context or anything about the exact location in the document. It will usually give me some wording if I’m missing it and I can go to the document and search for that wording.
I’m just one person searching a handful of documents so the sample size is pretty small for repeatability, so far, if it says it’s in there, it’s in there. It definitely misses things though, I’m still early in the process. I need to try some different models and perhaps clean up the data a little bit for some of the stuff.
Using the documentation as source data It doesn’t seem to hallucinate or insist things are wrong, it’s more likely to say I don’t see any information about that when the data is clearly in the data set somewhere.
YW on the responses I’m having fun with it even if it’s taking forever to get it to dial in and be truly useful.
That’s pretty smart, using it for legal documents. If the accuracy is high, it might be nice to just copy paste any tos or whatever to get the highlights in plain language (which imo should be a legal requirement of contracts in general, but especially ones written by a team of bad faith lawyers intended for people they don’t expect to read it and deliberately written to discourage reading the whole thing).
We’re a long way from trusting it to do something critical without intervention.
AI would be good at looking at an X-ray after a doctor and pointing out anomalies. But it would be bad to have it tell the doctor that everything looks fine.