kirill5pol

7 ms·

31.
by kirill5pol 3y ago
Mind sending me an pm? (my email's in my bio) I'd love to try it out with some of the podcasts/nvim chatgpt responses you mentioned and add it to your account!
32.
by kirill5pol 3y ago
Well that music video lyrics work is a very unexpected pleasant surprise! :) Thanks for catching that! Will fix that!
33.
by kirill5pol 3y ago
This is awesome! Did you hand select all 1000? I wanted to hand select for the categories to get some better starting recommendations but it was taking too long so I was a bit lazy and just scripted it…
34.
by kirill5pol 3y ago
Any particular type of content you’re thinking of? I’m currently working on adding podcasts since they’re pretty similar to YouTube videos, but I’m sure with some tuning I could see if it works for other things
35.
by kirill5pol 3y ago
Yes!! I definitely spend a little too much time on YouTube myself, would you mind sharing what kind of content (categories) you watch in a pm? I’d love to get some better ideas of what kind of material to tune the question generations on
36.
by kirill5pol 3y ago
I have a fairly simple implementation of the SM-2 algorithm, and just making the assumption that if you answered the question correctly then you have “perfect recall”. This isn’t exactly correct but I have been using it myself and seems to
37.
by kirill5pol 3y ago
Yeah it’s not ideal yet the suggestions under each category I just YouTube’s own category labels so often it didn’t give the best results, but it’s something I’m working on!
38.
by kirill5pol 3y ago
So far it’s basically a heuristic for the video having both captions and the token count for generation be reliably under the limit, but I am working on making it work for arbitrary length videos! I did some tests for 2-3 hour podcasts and
39.
Show HN: Anki/Duolingo-like app using educational YouTube videos(platoedu.org)
289 points by kirill5pol 3y ago|97 comments
40.
by kirill5pol 3y ago
I think a good way of explaining #2 is “weight” in the sense of a weighted average
41.
by kirill5pol 3y ago
In figure 3, it shows that Falcon and Pythia are much less susceptible to the lack of an attention sink than Llama or MPT… seems they work almost as well by just doing naïve windowed attention
42.
by kirill5pol 3y ago
One big one is speculative decoding for larger models, the more “accurate”* the smaller model, the more the speed up on the bigger model * as in matches the token that the larger model would output
43.
by kirill5pol 3y ago
The last part is interesting! What kind of use case would the users prefer to have it slower?
44.
by kirill5pol 3y ago
Just from the abstract, this is primarily for batching inference, for batched inference using GPUs gives an order of magnitude speed increase so probably not something that usually makes sense to do on CPUs…
45.
by kirill5pol 3y ago
For local yes, but at data center level the parallelization is still often worth it.
46.
by kirill5pol 3y ago
Yep, had the same thoughts a few years ago too. Basically if you accept mathematical realism as an axiom, pretty often you get to this idea
47.
by kirill5pol 3y ago
Yes but they’re slow enough on normal hardware for that 5-10x to be painful…
48.
by kirill5pol 3y ago
From their numbers 3 hours with 32 A100 80GBs. From lambda cloud: 3 hours * 4 ~22$/hr for 8x A100 ~= $265 So yeah not too expensive even for a native fine tune (obviously this ignores all other costs other than the GPUs)
49.
by kirill5pol 3y ago
The article linked here talks about GPT4 being a mixture of experts, which is exactly what you’re describing
50.
by kirill5pol 3y ago
That hasn’t been true for a while. Falcon 40B seemingly outperforms LLaMA 60B according to the OpenLLM leaderboard https://huggingface.co/tiiuae/falcon-40b
51.
by kirill5pol 3y ago
Same here, live YouTube premium but hate shorts. I’ve even reached out to Google to have an option to disable (which I know won’t go anywhere but )
52.
by kirill5pol 3y ago
That’s nearest neighbour search which scales O(n^2) for the number of vectors in your DB, what these DBs (and libraries like FAISS) use is approximate nearest neighbour, which makes the search much, much faster.
53.
by kirill5pol 3y ago
From sounding out “XTC”, I’m assuming it means ecstasy.
54.
by kirill5pol 4y ago
Oops thanks for catching that!
55.
by kirill5pol 4y ago
Yeah, it's definitely not at the level where it can be fully trusted yet, but I found GPT (and this) quite helpful for learning about something new where I don't have any background in. That's also why I think that something
56.
Show HN: Simply explain 20k concepts using GPT(platoeducation.ai)
61 points by kirill5pol 4y ago|30 comments
57.
by kirill5pol 4y ago
This is something I’m working on. I think having a good understanding of the material for good explanations is doable in the near future, but the main problem is actually building a good model of what the learner knows to guide the LLM
58.
by kirill5pol 4y ago
This is what I’m actually working on! One major problem with LLMs is that they don’t have a long term way of figuring out what your “knowledge space” is so no matter how much good the LLM is at explaining, it won’t be able to give you custo
59.
by kirill5pol 5y ago
To be fair per capita is the only relevant comparison.
60.
by kirill5pol 5y ago
UBC sailbot?

More ›