MasterScrat

7 ms·

1. 61.▲
   by MasterScrat 2y ago

   His intro to RL (not for LLM) blog post is a great read FYI https://karpathy.github.io/2016/05/31/rl/
2. 62.▲
   by MasterScrat 2y ago

   I'd be curious to hear how you'd express these notions in more scientific terms?
3. 63.▲
   by MasterScrat 2y ago

   I think both of these projects miss what makes TikTok (and reels in general) so effective. A good TikTok video gets "injected into your brain". You have zero effort to provide and suddenly this stuff is in your mind. I'm not
4. 64.▲
   by MasterScrat 2y ago

   You mean, like the fastapi documentation used to be? https://github.com/fastapi/fastapi/issues/3273
5. 65.▲
   by MasterScrat 2y ago

   How do bandwidth costs work now? do you pay the ISPs a flat fee, or is it still usage-based? how much cheaper is it compared to cloud providers?
6. 66.▲
   by MasterScrat 2y ago

   Can you talk about the main non-LLM NLP tools you use? e.g. BERT models? > One prompt? Fair. 10? Still ok. 100? You're pushing it. 10M - get help. Assuming you could do 10M+ LLM calls for this task at trivial cost and time, would
7. 67.▲
   by MasterScrat 2y ago

   This hasn't been my experience, using TikTok from Switzerland, I almost exclusively see English language, with a focus on my interests
8. 68.▲
   by MasterScrat 2y ago

   Thanks, amazing story! I found this nice coverage of the events: https://www.mentalfloss.com/article/92007/why-us-federal-cou...
9. 69.▲
   by MasterScrat 2y ago

   Multi-touch support (at least on iOS) is a nice touch
10. 70.▲
    by MasterScrat 2y ago

    Very insightful to have a number from him here: > LLMs are trained on much more than the whole Internet -- they also consume handcrafted answers produced by armies of highly qualified data annotators (often domain experts). Today approxi
11. 71.▲
    by MasterScrat 2y ago

    Here's how it looks like apparently: https://youtu.be/hXUbTKd6pmo?t=29
12. 72.▲
    by MasterScrat 2y ago

    Interesting point. In a way this is a "simulated game engine", trained from actual game engine data. But I would argue a working simulated game engine becomes a game engine of its own, as it is then able to "propell the game&
13. 73.▲
    The Problem with Startup "Experts" [video](youtube.com)

    1 points by MasterScrat 2y ago|0 comments
14. 74.▲
    by MasterScrat 2y ago

    We sell text-to-image model finetuning (aka "Dreambooth") as a service and yes, this is one of the use cases. Recently a travel agency used our platform to generate images of people in the destinations they were advertising.
15. 75.▲
    by MasterScrat 2y ago

    Ok I now understand better what happened: The price for using images as part of your prompt has indeed not changed between GPT-4o-mini and GPT-4o Yet overall, captioning 500 images now costs me 5x less. This is because when I'm caption
16. 76.▲
    by MasterScrat 2y ago

    It almost sounds shady... "it's 30x cheaper per token but you now need 30x more tokens per image"? Has anyone already validated this based on billed cost? running a batch myself to check EDIT: Ok so I captioned 500 images in
17. 77.▲
    by MasterScrat 2y ago

    It also provides decent spam prevention, and high probability people will check the associated email address
18. 78.▲
    by MasterScrat 2y ago

    It's a tangent, but reading this linked article from 2014: https://blog.cloudflare.com/the-relative-cost-of-bandwidth-a... Am I reading correctly that egress in Europe costs $8 Mbps/month which is $0.0004/GB,
19. 79.▲
    by MasterScrat 2y ago

    Mostly from the tight JAX-TPU integration yeah
20. 80.▲
    by MasterScrat 2y ago

    We've built our startup from scratch on JAX, selling text-to-image model finetuning, and it's given us a consistent edge not only in terms of pure performance but also in terms of "dollars per unit of work"
21. 81.▲
    by MasterScrat 2y ago

    I’m not saying you’re a GPT4-based bot, but I want to point out this comment really triggered my LLM radar
22. 82.▲
    by MasterScrat 3y ago

    I used Windows 10 for the first time in years this weekend to play games While playing , Windows would switch from the game to a stupid "Restart your computer now or in 1 hour?" modal - with no option to get it to permanently go
23. 83.▲
    by MasterScrat 3y ago

    We started on V3s, now fully moved to V4s with some V5Es, investigating a full move towards V5E & V5P
24. 84.▲
    by MasterScrat 3y ago

    Interesting! this was already the case with TPUs easily beating A100s. We sell Stable Diffusion finetuning on TPUs (dreamlook.ai), people are amazed how fast and cheap we can offer it - but there's no big secret, we just use hardware t
25. 85.▲
    by MasterScrat 3y ago

    Scaling up compute can improve throughput, but can't easily improve latency between tokens. Generation is usually bottlenecked by the time it takes to go through the network for each token. To speed that up, you need to perform these c
26. 86.▲
    by MasterScrat 3y ago

    And how does Mistral do "accurate long content retrieval"?
27. 87.▲
    There's a supercomputer lodged inside this Barcelona church(lonelyplanet.com)

    1 points by MasterScrat 3y ago|0 comments
28. 88.▲
    by MasterScrat 3y ago

    Heartbreaking to hear that a company that's really helping moving things forward is struggling
29. 89.▲
    by MasterScrat 3y ago

    Interesting - so Azure OpenAI was not affected? Do they get system updates at the same time as the OpenAI API? Is the pricing the same?
30. 90.▲
    by MasterScrat 3y ago

    TPUs are amazing for Stable Diffusion. We've been doing training (Dreambooth) and inference on TPUs since the beginning of the year at https://dreamlook.ai . We basically get 2.5x the training speed for Stable Diffusion 1.5