Alt. Profile @Th4tGuyII

  • 0 Posts
  • 36 Comments
Joined 25 days ago
cake
Cake day: June 11th, 2024

help-circle
  • So providing a fine-tuned model shouldn’t either.

    I didn’t mean in terms of providing. I meant that if someone provided a base model, someone took that and but on of it, then used it for a harmful purpose - of course the person modified it should be liable, not the base provider.

    It’s like if someone took a version of Linux, modified it, then used that modified version for a similar person - you wouldn’t go after the person who made the unmodified version.


  • SB 1047 is a California state bill that would make large AI model providers – such as Meta, OpenAI, Anthropic, and Mistral – liable for the potentially catastrophic dangers of their AI systems.

    Now this sounds like a complicated debate - but it seems to me like everyone against this bill are people who would benefit monetarily from not having to deal with the safety aspect of AI, and that does sound suspicious to me.

    Another technical piece of this bill relates to open-source AI models. […] There’s a caveat that if a developer spends more than 25% of the cost to train Llama 3 on fine-tuning, that developer is now responsible. That said, opponents of the bill still find this unfair and not the right approach.

    In regards to the open source models, while it makes sense that if a developer takes the model and does a significant portion of the fine tuning, they should be liable for the result of that…

    But should the main developer still be liable if a bad actor does less than 25% fine tuning and uses exploits in the base model?

    One could argue that developers should be trying to examine their black-boxes for vunerabilities, rather than shrugging and saying it can’t be done then demanding they not be held liable.














  • I agree on them being safe - when rules are properly adhered to, they’re extremely safe, similarly to air travel. People only suspect their safety because when they do fail, they tend to fail spectacularly, again similar to air travel.

    Having said that, they may be efficient to operate, but they are by no means efficient to build. They cost a lot of resources, and have a 10 year lead time - plus you need to worry about the cost of waste storage and decommissioning.

    So sure, nuclear is better than fossil fuels, but you’re just kicking the nonrenewable can down the road.

    That time and resources would be far better spent on renewables, because that where humanity is gonna have to go long-term no matter how well any other alternatives work.





  • The TL;DR for the article is that the headline isn’t exactly true. At this moment in time their PPU can potentially double a CPU’s performance - the 100x claim comes with the caveat of “further software optimisation”.


    Tbh, I’m sceptical of the caveat. It feels like me telling someone I can only draw a stickman right now, but I could paint the Mona Lisa with some training.

    Of course that could happen, but it’s not very likely to - so I’ll believe it when I see it.

    Having said that they’re not wrong about CPU bottlenecks and the slowed rate of CPU performance improvements - so a doubling of performance would be huge in this current market.