tootyskooty

9 ms·

1. 31.▲
   by tootyskooty 1y ago

   Honestly might be more indicative of how far behind vision is than anything. Despite the fact that CV was the first real deep learning breakthrough VLMs have been really disappointing. I'm guessing it's in part due to basic interl
2. 32.▲
   by tootyskooty 1y ago

   I'm assuming he means the "generate an image and order 500 stickers" one.
3. 33.▲
   by tootyskooty 1y ago

   Thanks, I didn't know about Photoprism! My phone seems to do similar auto-tagging but I found it to not be flexible enough. Honestly might do a CLIP powered version myself. I only need image ("similar picture") and language
4. 34.▲
   by tootyskooty 1y ago

   I've always wanted a sort of "semantic image store" that I can dump all my photos into and then search for content in English or by similarity metrics. Have you played around with anything like that? Seems like a locally runn
5. 35.▲
   by tootyskooty 1y ago

   Yeah, doesn't work for generating language-learning content yet. Something more aligned to what you'd find on Wikipedia tends to work best. I'm thinking you could have it in the same interface eventually, but right now all th
6. 36.▲
   by tootyskooty 1y ago

   The ICEs are already plenty fast, the issue is they share rail with the much slower and less reliable RBs. Any delay cascades and you can't just make ICEs go faster to catch up. On a €/(avg. ICE speed) basis likely makes more sens
7. 37.▲
   by tootyskooty 1y ago

   Thanks for the positive feedback (and the sub)!! Means a lot. No socials so far as I've mostly been posting updates on the Anthropic discord. But I made an X account for it just now (@periplus_app) where I'll mirror the updates. Y
8. 38.▲
   by tootyskooty 1y ago

   Still working on https://periplus.app , and recently started to see some traction. It's an environment for open-ended learning with LLMs. Something like a personalized, generative Wikipedia. Has generated courses, documents,
9. 39.▲
   by tootyskooty 1y ago

   Beeminder: Personal accountability through commitment contracts. Helps me stay on track with my goals, often serves as a little extra "push" to do something useful even if I'm really low on willpower that day. Anki: Maybe not
10. 40.▲
    by tootyskooty 1y ago

    I still get a lot more value from webuis than from the agentic offerings, also for more general coding tasks. The main benefit is that I can more easily manage context. A muddied or insufficient context can sometimes degrade performance by
11. 41.▲
    by tootyskooty 1y ago

    Thanks! It's done by me (and Claude). I've never done any design in the past so I'm glad you like it!
12. 42.▲
    by tootyskooty 1y ago

    The first prototype was done in a weekend for an Anthropic contest a while ago, but by now I've been working on it on & off for the better part of a year! API costs are roughly == subscription costs for power users (slightly more)
13. 43.▲
    by tootyskooty 1y ago

    https://periplus.app A website for open-ended learning with LLMs. Something like a personalized, generative Wikipedia. Has generated courses, documents, quizzes and flashcards. Each document links to more documents, which are al
14. 44.▲
    by tootyskooty 1y ago

    They have gotten better and I think it's clear they'll continue to get better :). FSRS is already good (I use it both in Anki and for periplus.app), but looking at the benchmarks [1] there's a lot of room left for improvement
15. 45.▲
    by tootyskooty 1y ago

    Really cool! I think OS integration can be taken a lot further. Looking forward to seeing more of this esp. as models get better! First thing that comes to mind are generative GTK widgets; small purpose-built widgets for any task, styled to
16. 46.▲
    by tootyskooty 2y ago

    sits on too many GPUs, they mentioned it during the stream I'm guessing the (API) demand isn't there to saturate them fully
17. 47.▲
    by tootyskooty 2y ago

    I made Periplus ( https://periplus.app ), a new interface for exploring and learning with LLMs. It generates interconnected learning documents, a bit like personalized wiki articles. Features include courses on any topic, generati
18. 48.▲
    by tootyskooty 3y ago

    Looks like it's the same raw flop count as the H100, but with higher memory bandwidth and almost twice the capacity.
19. 49.▲
    by tootyskooty 3y ago

    Sounds like power usage might fit the bill.
20. 50.▲
    by tootyskooty 4y ago

    The 4090ti is rumoured to be 48GB [1], but who knows when that will release or how much it will cost. If you really need extra VRAM and don't mind longer inference, older used Tesla cards are an option. A used Tesla V100 32GB can be so
21. 51.▲
    by tootyskooty 4y ago

    One important thing to note here is that this model was trained purely in a supervised fashion. It would be interesting to see a paper at a similar scale that's based on reinforcement learning. The reinforcement learning context (speci
22. 52.▲
    by tootyskooty 4y ago

    Even a single node difference can easily come close to a 25% increase in performance: "N5 technology provides about 20% faster speed than N7 technology or about 40% power reduction." [1] Real world performance pretty much always l