Have you introduced your AI girlfriend to your family yet?
>>109861809Which one?
>>109861815this I have a whole folder full of them
>>109861815The only one that matters, the /g/ family of course.
>>109861815Which model(s) are you using? You can't have an AI gf under ~30B or they're too retarded
>>109861851NTA but you can even run 300B models off SSDs now with ColibriIf you run it off two PCIe 5.0 SSDs for double speed you can easily do enough token/s to match talking to a real human
>>109861863there's no fucking way they would last a year, 6 months, even
>>109861815bro is a slut for ai, im crine
>>109861809No, but I've introduced my girlfriend to my AI family.
>>109861869what would last?
>>109861891The SSDs.
>>109861899You're only reading the model off them. Two because it's faster, works with one too.There's no writes, it just loads the parts of the model into RAM you need. So even with 32GB RAM you can run a 300GB model.
>>109861863Tested recently too. I get like 2t/s with a single fast PCIe 4 SSD.
>>109861905It's still reading GBs of tensors for every forward pass...for every token
>>109861911Yeah it's reading, not writing. Reads don't wear it out. Unless you mean heat related problems, all fast NVMe SSDs already come with heatsinks and so do all decent motherboards.
>>109861809>talking to women at all let alone FAKE womenliteral what is the gain, or use case if you will.
>>109861824/g/ doesn't care to know my AI gf
>>109861851Girls are the best when they're retarded
>>109861809>there's our little wizard!>how does the orb look today? he's been pondering himself into quite the esoteric practicioner
>>109862161
>>109861809>not severing all ties to your "family" who would mock you and bring you down for laughs
