tintor

11 ms·

1. 211.▲
   by tintor 2y ago

   "Screen images simulated."
2. 212.▲
   by tintor 2y ago

   Because US is not a democracy anymore. Checks and balances are gone.
3. 213.▲
   by tintor 2y ago

   You can fine tune a checkpoint of model during pre-training.
4. 214.▲
   by tintor 2y ago

   Not as fast as other 17B models if it has to attend to 10M context window.
5. 215.▲
   by tintor 2y ago

   Burning Man is an example of advertising free world and it is quite a refreshing ad-free week.
6. 216.▲
   by tintor 2y ago

   DOGE administrator is ... Grok.
7. 217.▲
   by tintor 2y ago

   Lawyer time $10.8 ~ 135 days * 8 hours * $1000/h x10
8. 218.▲
   by tintor 2y ago

   Thanks. Here is presentation by Karpathy explaining different stages of LLM training. Explains many details in a form suitable for beginners. https://www.youtube.com/watch?v=bZQun8Y4L2A
9. 219.▲
   by tintor 2y ago

   scarecrows?
10. 220.▲
    by tintor 2y ago

    Richard Sutton - Reinforcement Learning, An Introduction
11. 221.▲
    by tintor 2y ago

    They are different processes. - RLHF: Turns pre-trained model (which just performs autocomplete of text) into a model that you can speak with, ie. answer user questions and refuse providing harmful answers. - Distillation: Transfer skills &
12. 222.▲
    by tintor 2y ago

    o3 low and high are the same model. Difference is in how long was it allowed to think. It works the same with humans. If they spend more time on the puzzle they are more likely to solve it.
13. 223.▲
    by tintor 2y ago

    “they won’t leak” That one isn’t guaranteed. Many examples online of exfiltration attacks on LLMs.
14. 224.▲
    by tintor 2y ago

    Why are they not releasing it in 27" size that they had before?
15. 225.▲
    by tintor 2y ago

    It can be verified if resulting video contains signed metadata with all intermediate steps needed to produce the video from original recording (which is digitally signed by camera). Downside is that large original video assets would need to
16. 226.▲
    by tintor 2y ago

    Possible option is for cameras to digitally sign the original video as it is being recorded.
17. 227.▲
    by tintor 2y ago

    Charging for luggage is fine. The problem is when luggage costs the same or more as ticket without luggage.
18. 228.▲
    by tintor 2y ago

    Too much annoying visual clutter on the discussion page, unlike Hacker News.
19. 229.▲
    by tintor 2y ago

    > they have brought the cost of pretraining a frontier 400B+ parameter model to ZERO It is still far from zero.
20. 230.▲
    by tintor 2y ago

    Smaller models can be trained to match log probs of the larger model. Larger model can be used to generate synthethic data for the smaller model.
21. 231.▲
    by tintor 2y ago

    Meta can call it something else other than open source. Synthetic part of the training data could be released.
22. 232.▲
    by tintor 2y ago

    What % of them were from FAIR vs non-FAIR?
23. 233.▲
    by tintor 2y ago

    The "cyber agencies" could focus on pen testing of domestic companies, and issuing fines to insecure ones.
24. 234.▲
    by tintor 2y ago

    US already has https://en.wikipedia.org/wiki/United_States_Cyber_Command
25. 235.▲
    by tintor 2y ago

    How is this not insider trading?
26. 236.▲
    by tintor 2y ago

    There is plenty of code and bug trackers on the web for ChatGPT to learn from: OSS, Github, Sourceforge, ...
27. 237.▲
    by tintor 3y ago

    Side windows only for business class. How to make economy even more miserable.
28. 238.▲
    by tintor 3y ago

    Why offsite?
29. 239.▲
    by tintor 3y ago

    I just tried "Consumer Disclosure Report" link from LexisNexis you shared, and nothing happens when I submit the request. :(
30. 240.▲
    by tintor 3y ago

    Cost and risk of providing inaccurate answers to election-related questions is high, so they are avoiding them.