eurekin

22 ms·

1. 331.▲
   by eurekin 3y ago

   It's based on a real world filter, which actually is a separator. By that analogy anything that sticks to the filter is considered a waste
2. 332.▲
   by eurekin 3y ago

   I'm especially surprised a tiny milligram worth of brain is able to provide such high level functions and coordination. I'd normally expect all those neurons would be occupied fully with processing visual stimuli, smells and coord
3. 333.▲
   by eurekin 3y ago

   I wonder if that kind of trick could be used during the training of adversarial networks - GANs, to have a strong adversary
4. 334.▲
   by eurekin 3y ago

   I like to use chatgpt for the easy stuff, things I forgot, do to rarely to remember or code in another language very similar to the one I already know. I do quickly run into bumps, where search is necessary (a lot of times it's some va
5. 335.▲
   by eurekin 3y ago

   Thank you for that comment. If you ever come across that article, please share
6. 336.▲
   by eurekin 3y ago

   What effective client speeds are you getting?
7. 337.▲
   by eurekin 3y ago

   What about networking? Did you go over 1gbit?
8. 338.▲
   by eurekin 3y ago

   I wish I saved my first attempts at Christmas tree svg generations... Comparing the unicorn it's like a night and day improvement. A lot better now.
9. 339.▲
   by eurekin 3y ago

   I added those in succession: > man woman king queen ruler force powerful care and couldn't reliably determine position of any of them
10. 340.▲
    by eurekin 3y ago

    I use that all the time in my hobby tinkering pseudo cloud server on a odroid SBC. It feels like I'm literally on that specific computer directly. Plugins like docker work as well
11. 341.▲
    by eurekin 3y ago

    Maybe it's a modulo operation squished together with the modus operandi, for laughs
12. 342.▲
    by eurekin 3y ago

    At my department, instructors were well aware of statistics. That was a prerequisite course on the AI path. Some early day software (WEKA) used statistic nomenclature extensively.
13. 343.▲
    by eurekin 3y ago

    re.: > The problem of "identification" was quickly solved by another engineering feat, which was to slap on "positional embeddings". As usual, this too didn't happen because there was a deep mathematical understa
14. 344.▲
    by eurekin 3y ago

    I'm really speculating here! I think 3 is a good fit, since: the K + Q (both together) represent simply pairwise importance (of tokens + positional embedding representation) The V lifts that up one "abstraction level": K+Q al
15. 345.▲
    by eurekin 3y ago

    Those are such great questions. I'm also trying to find out and my current notes are as following. Attention matrix, From the llm viz bbycroft.net: So the main goal of self-attention is that each column wants to find relevant inf
16. 346.▲
    by eurekin 3y ago

    For anybody wanting to take a look at the code, this time the Github link does include it - it's not empty, which is typicaly for those "too good to be true" publications
17. 347.▲
    by eurekin 3y ago

    This specifically is: TheBloke / Mistral-7B-Instruct-v0.2-GGUF / mistral-7b-instruct-v0.2.Q8_0.gguf and I'm running it in LMStudio with the config: { "name": "Exported from LM Studio on 21.12.2023, 14:57:
18. 348.▲
    by eurekin 3y ago

    How long should the domain be up not to be considered new?
19. 349.▲
    by eurekin 3y ago

    It's basically a stream of jpegs - true!
20. 350.▲
    by eurekin 3y ago

    I was always stumped by this effect. Are we suppose to scroll consistently or with a big swipe? What if, I did not swipe strongly enough? Should I be ashamed I didn't make it to the end? O the desktop it's always: should I be the
21. 351.▲
    by eurekin 3y ago

    Downvoters care to comment? Uncensored llm versions typically perform better (at least on benchmarks) to their "lobotomized" or aligned counterparts
22. 352.▲
    by eurekin 3y ago

    I'm using the tiny mistral 7b for those 1hr long transcripts internal in my company. I was surprised that even the quantized 7b version easily chomped my 3090's vram - the context takes a lot. I think it goes up to 32k tokens (I g
23. 353.▲
    by eurekin 3y ago

    I itched to build sth in this context. I'm so happy someone did actually do that! Some videos are a real treasure
24. 354.▲
    by eurekin 3y ago

    I recently made them work with meeting presentation transcript. Mistral 7b (tiny one) was enough to reliably extract quite a lot of information! - It could derive speaker names, based purely on how people called themselves in the conversati
25. 355.▲
    by eurekin 3y ago

    > like having all resources ... > AI in charge of many of our resources > provided a capable AI arrives Yes, this is actually coming closer to reality. Please be more specific and list one resource, how we arrive at AI that control
26. 356.▲
    by eurekin 3y ago

    Ah, that might be it! I have only 32
27. 357.▲
    by eurekin 3y ago

    > simple natural order of the food chain So, AGI would need to exist in some robotic form (or control whatever keeps us alive), without oversight, control and share our biological needs and aspiritions (in order to participate in the foo
28. 358.▲
    by eurekin 3y ago

    No, I have tried about 3 times and all failed. Altough it might be that I'm simply doing it wrong. There are people (can point at a discord server) claiming it works for them and that they even sell finetuned models to business clients
29. 359.▲
    by eurekin 3y ago

    - in context learning (simply put it into context) - finetuning, as in restarting the models checkpoint and relearning it on the previous + new data - adding extra neurons (e.g. LoRA adapters) at certain places and restarting learning Oh, i
30. 360.▲
    by eurekin 3y ago

    Yes! "Reality has a surprising amount of detail" http://johnsalvatier.org/blog/2017/reality-has-a-surprising-... and I'm yet to see a single believably convincing AI threat scenario.