p1esk

67 ms·

1. 751.▲
   by p1esk 3y ago

   It’s not 8x86B. Total number of parameters is 314B. Perhaps it’s 8x39B to fit on a single 8xA100 (40GB) server?
2. 752.▲
   by p1esk 3y ago

   Please give me an example where GPT4 reasons worse than an average adult college educated human. Any topic, STEM or otherwise.
3. 753.▲
   by p1esk 3y ago

   The DSL: https://openai.com/research/triton
4. 754.▲
   by p1esk 3y ago

   Is that half life temperature dependent?
5. 755.▲
   by p1esk 3y ago

   I’m pretty sure GPT4 would wipe the floor with you in any debate (including this one).
6. 756.▲
   by p1esk 3y ago

   Inference world is gradually switching from INT formats to FP formats. FP8 is already supported in modern hardware, and FP4 support is coming. In my experiments I get better perplexity in language models with FP4 than with INT4.
7. 757.▲
   by p1esk 3y ago

   I just looked at the paper: the answer is no, floating point is not supported.
8. 758.▲
   by p1esk 3y ago

   For floating point? Are you sure?
9. 759.▲
   by p1esk 3y ago

   No, but if they express interest I will try to help them.
10. 760.▲
    by p1esk 3y ago

    I plan to.
11. 761.▲
    by p1esk 3y ago

    Thanks, good to see a decent paper on this. I really hope AI will speed up this research. Are CI and Alcor still the only two companies doing human cryopreservation? Shouldn’t they be doing some of this research using customer money (chargi
12. 762.▲
    by p1esk 3y ago

    The link you posted says the last improvement to their process was done in 2004. So in the last 20 years, the vitrification has not gotten better. Is it because no one has been doing any research, and if so, why not? If you look at any othe
13. 763.▲
    by p1esk 3y ago

    Sadly this is becoming more common everywhere. Companies (and individual people) just don’t care about long term prospects. This includes SWE industry: if people switch companies every two years on average, why care about things like tech d
14. 764.▲
    by p1esk 3y ago

    You might be confused: to them, people like you and the OP are bad actors. What you mean by “bad actors” are their paying customers.
15. 765.▲
    by p1esk 3y ago

    Has there been any progress in cryonics in the last 30 years?
16. 766.▲
    by p1esk 3y ago

    Where did you pull this “taxonomy” from?
17. 767.▲
    by p1esk 3y ago

    The difference between our opinions is I see 2024 as the time just before the knee of the exponential progress curve, and you see it as a long way to go to get to the knee. I realize how saying "this time it's different" migh
18. 768.▲
    by p1esk 3y ago

    A superior race (AI or aliens) might also see some purely aesthetic value in individual humans, like butterfly collectors.
19. 769.▲
    by p1esk 3y ago

    How can you possibly regulate 2?
20. 770.▲
    by p1esk 3y ago

    Me neither.
21. 771.▲
    by p1esk 3y ago

    I thought we were talking about why would someone be concerned with recent progress in AI. That’s the question I tried to answer.
22. 772.▲
    by p1esk 3y ago

    Two likely scenarios: 1. Bad people will use powerful models to do bad things. 2. Many jobs will be automated, and many people will get laid off.
23. 773.▲
    by p1esk 3y ago

    I definitely see value in ants.
24. 774.▲
    by p1esk 3y ago

    GPU time.
25. 775.▲
    by p1esk 3y ago

    Is there any evidence that such a probabilistic model can run better than a state of the art model? No, but they got 15M seed funding anyway.
26. 776.▲
    by p1esk 3y ago

    Pretty much.
27. 777.▲
    by p1esk 3y ago

    AI has no evolutionary pressures imposed on it Are you sure? https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboar... Just wait until we start putting these models inside robots, or world simulators, and
28. 778.▲
    by p1esk 3y ago

    If AI does ever become sentient, it will probably reason and decide about eliminating large swaths of human population in a similar fashion as we reason and decide about eliminating large swaths of ant population.
29. 779.▲
    by p1esk 3y ago

    Yes, but it’s heavily constrained in what it’s allowed to do through ‘alignment’, and it’s not that smart. If someone releases a pre-alignment version of GPT-5 into the wild, that would be scary.
30. 780.▲
    by p1esk 3y ago

    To solve self-driving we need two things: 1. Deep and detailed understanding of how the world works. We are just starting to make real progress there (GPT-4), and more work is needed [1]. 2. Reliability. A model should make significantly fe