Comment by wormius | Hacker News Reader

Comment by wormius | Hacker News Reader

Parent Full thread

I only have a Ryzen 5 3600x and a 7650xt, and downloaded AMD's LLM tool. It runs, and chats. I have thought about training it on the 4500 pages of LJ entries I saved to PDF (I know I'd have to extract the text, and then spell check that puppy. or not. Maybe I should just leave my misspells if it's a doppleganger).

But it seems it'd take a few weeks of running nonstop to train? Everywhere I'm reading says it's hard to determine the actual time required.

I would think 4500 pages of text would be a large enough volume to train. (assuming 500 words per page? If not more)

What's a good minimum number of blog entries/words to train a thing. I would think the more the merrier, but I ain't spending non-stop processing for weeks at a time.

I hate AI, I have no idea why I even want to try this, but... The data is there and it's local to my machine not shared with upstream, so I figure it'd be fine to mess with. Is my estimate in the ballpark?

Depending on the finetuning tool you're using, you can just start the training run, and then it shows you how long it'll take. Like give it 5 mins to stabilise, then see the estimated duration.

Axolotl is a good finetuning tool if you need one.

If you hate AI, why do this?

If you just want to do it, why not just do it?

I should clarify - I'm ambivalent on a lot of the current structures (political, economic) that this is being created in. It's less about "AI" and more about how it's being shoved down our throats. I have ethics concerns (even though I understand and am slightly sympathetic to the arguments that "training" an AI on data is like a human studying previous works - though I don't think that's what's really happening). I have concerns about this being done "at scale" (especially seeing what Nvidia is doing to build giant massive data centers - though - as with most tech "it gets cheaper/cleaner/more efficient" so I can understand over time this could go down).

Being local reduces some of these issues... But even then I'm not sure I trust the big actors the renege on the original deals and start hoovering what they can.

That's why I'm willing to play with this locally on my own end just for funs while not necessarily buying into the hype, or trusting the players/creators that they have the best intent besides more exploitation of natural, human, and data resources.

Sometimes it's ok to have complex opinions, and admittedly I was a bit strong in my condemnation above, but also brevity in comments is a virtue, better to underexplain at first than overexplain and clarify if needed. (He says as he spends multiple paragraphs over-explaining in a followup)

Not OP but I’d personally would rather be an informed hater instead of a believer hater.

Reply on news.ycombinator.com