Searching yarn

Twts matching #training
Sort by: Newest, Oldest, Most Relevant
xuu
txt.sour.is
 (2y ago)
↳ In-reply-to » Home | Tabby This is actually pretty cool and useful. Just tried this on my Mac locally of course and it seems to have quite good utility. What would be interesting for me would be to train it on my code and many projects 😅

Most of the can run locally have such a small training set they arnt worth it. Are more like the Markov chains from the subreddit simulator days.

There is one called orca that seems promising that will be released as OSS soon. Its running at comparable numbers to OpenAI 3.5.

https://youtube.com/watch?v=Dt_UNg7Mchg&feature=share9

⤋ Read More
abucci
anthony.buc.ci
 (2y ago)
↳ In-reply-to » I played with nlpodyssey/verbaflow: Neural Language Model for Go today a little bit today.... First I had to download a ~2GB file (the model), then convert that to a format the program verbaflow understands which came out to roughly ~5GB. Then I tried some of the samples in the README. My god, this this is so goddamn awfully slow its like watching paint dry 😱 All just to predict the next few tokens?! 😳 I had a look at the resource utilisation as well as it was trying to do this "work", using 100% of 1.5 Cores and ~10GB of Memory 😳 Who da fuq actually thinks any of this large language model (LLM) and neural network crap is actually any good or useful? 🤔 Its just garbage 🤣

@prologic@twtxt.net You more or less need a data center to run one of these adequately (well, train…you can run a trained one with a little less hardware). I think that’s the idea–no one can run them locally, they have to rent them (and we know how much SaaS companies and VCs love the rental model of computing).

There’s a lot of promising research-grade work being done right now to produce models that can be run on a human-scale (not data-center-scale) computing setup. I suspect those will become more commonly deployed in the next few years.

⤋ Read More
abucci
anthony.buc.ci
 (2y ago)
↳ In-reply-to » @prologic hmm, dunno about the recency of that line of thought. I suspect though that given his (recent or not) history, if someone directly asked him "do you support rape" he would not say "no", he'd go on one of these rambling answers about property crime like he did in the video. Maybe I'm mind poisoned by being around academics my whole career, but that way of talking is how an academic gives you an answer they know will be unpopular. PhD = Piled Higher And Deeper, after all right? In other words, if he doesn't say "no" right away, he's saying "yes", except with so many words there's some uncertainty about whether he actually meant yes. And he damn well knows that, and that's why I give him no slack.

@prologic@twtxt.net It’s a fun challenge to see how many words you can say without expressing any ideas at all. Maybe this GPT stuff should be trained to do that!

⤋ Read More
carsten
yarn.zn80.net
 (2y ago)

I was listening to an O’Reilly hosted event where they had the CEO of GitHub, Thomas Dohmke, talking about CoPilot. I asked about biased systems and copyright problems. He, Thomas Dohmke, said, that in the next iteration they will show name, repo and licence information next to the code snippets you see in CoPilot. This should give a bit more transparency. The developer still has to decide to adhere to the licence. On the other hand, I have to say he is right about the fact, that probably every one of us has used a code snippet from stack overflow (where 99% no licence or copyright is mentioned) or GitHub repos or some tutorial website without mentioning where the code came from. Of course, CoPilot has trained with a lot of code from public repos. It is a more or less a much faster and better search engine that the existing tools have been because how much code has been used from public GitHub repos without adding the source to code you pasted it into?

⤋ Read More