4 ms·
Now that LLaMa is out of the bag, there is no stopping the scammers. They can just run LLaMa on their own GPUs, fine tune it to whatever nefarious purpose they
by gamegoblin 4y ago
Now that LLaMa is out of the bag, there is no stopping the scammers. They can just run LLaMa on their own GPUs, fine tune it to whatever nefarious purpose they have, etc.
Not only can scammers do this, but they will prefer to do this purely for cost reasons. Scammers have to turn a profit, too. And nation state actors will do it for traceability reasons.
So the really bad actors won't use OpenAI anyway. So OpenAI should stop lobotomizing their service for the non-nefarious actors who are just trying to make the next generation of software.
Sure, go ahead and put subtle steganographic fingerprinting in the OpenAI output, so long as it doesn't affect quality. This will hinder the low-tech bad actors who can't run LLaMa themselves, and won't hurt the good actors.
But it's the most infuriating thing to try to build on the GPT4 API when its "helpful assistant" chat type mindset occasionally breaks the workflow.
- arpowers 4y agoThey will be using this angle as a moat, models are after all just a big file of assorted weights. They are losing sleep trying to figure out how to keep their lead
- isoprophlex 4y agoYeah and I wonder how futile that is. Looking at what happened with stable diffusion, there's no way this can be kept under wraps. Weights WILL leak and trigger a cambrian explosion of open source innovation. Their moat, if any, will be in the boring things: regulatory capture, integration, whatever you call Oracle's business model of "pay through the nose and blame us if something goes wrong". And at that point it'll be just any boring regular tech company.
- circuit10 4y agoI’m not sure, it probably depends on the scale of the scam but running your own infrastructure will be quite expensive if you’re not using it a lot
- isoprophlex 4y agoYou can trivially get compute and gpu resources that scale to zero when inactive.
- circuit10 4y agoYou’re probably right, I was thinking of a not I made (not for scams) that only replies once every hour or so and for that it would probably be impractical but if there are multiple people at a time it would make more sense
- isoprophlex 4y agoIf you want to just dick around a bit check out https://www.banana.dev/ https://www.banana.dev/ They offer serverless GPU deployments, pretty hassle free (compared to using AWS or azure)
- tarruda 4y agoApparently can fine tune and run Alpaca-lora 7B using a single RTX 4090: https://github.com/tloen/alpaca-lora https://github.com/tloen/alpaca-lora A few days ago before it was taken offline I did some tests I did with Alpaca-stanford. If I can run a GPT-3 level LLM in my machine, I won't even bother looking at GPT-4. Yes GPT-4 is impressive, but GPT-3 and Alpaca are already good enough IMO.