rnosov

31 ms·

31.
by rnosov 4y ago
Someone uploaded leaked LLaMA model on hugginface already: https://huggingface.co/spaces/chansung/LLaMA-7B
32.
by rnosov 4y ago
Generally, you'll need multiply model size by two to get required amount of video RAM. There are 4 sizes, so you might get away with even smaller GPU for say 13B model.
33.
by rnosov 4y ago
HN user (with >3k) karma seems to confirm the leak. Take it for what it's worth.
34.
by rnosov 4y ago
it's a pull request from ChristopherKing42. He is unlikely to be associated with Meta.
35.
by rnosov 4y ago
The linked page is just a pull request, the actual repository readme doesn't mention torrent option at all.
36.
by rnosov 4y ago
Just to make it clear, does this torrent include model weights?
37.
by rnosov 4y ago
If I understood docs correctly to create a special token you'd need to supply it in a special array (which users presumably have no access to).
38.
by rnosov 4y ago
From reading the docs it looks like there are ( or will be soon ) two distinct ways for API endpoint to consume the prompt: 1. Old one when all inputs are just concatenated into one string (Vulnerable to prompt injection) 2. Inputs supplied
39.
by rnosov 4y ago
My understanding that the breakthrough was the attention mechanism where attention layers learn where important words are.
40.
by rnosov 4y ago
Not that anxious at the moment. Could be a problem in future. There are also about 13000 nuclear warheads, climate is changing in a dangerous way, etc. etc. It could be the case that only AI might be able to survive on our planet in a centu
41.
by rnosov 4y ago
I think the point they make is that emissions from oil and gas industry are as deadly as Covid. Medicine, infrastructure, GDP all existed before we started to exploit large scale fossil fuel deposits. It would make our civilisation a lot mo
42.
by rnosov 4y ago
I think the answer is a lack of large scale storage mechanism for renewables. Perhaps, synthesising ammonia from renewable hydrogen could be a potential solution.
43.
by rnosov 4y ago
Surely people will be using it as a doctor. It takes over week to see real doctor where I live so basically Google is frequently the only "doctor" I can readily access.
44.
by rnosov 4y ago
Can you place more than two objects in the sketch? I'm trying to sketch a house with a tree side-by-side, but it always drawing trees in front of the house.
45.
by rnosov 4y ago
Could you quantify it (Twice, three times, etc)? Another question, does it apply to only attention heads learning or the whole shebang?
46.
by rnosov 4y ago
Is there any chance you can answer in simple terms whether the "Hopf coherence" method would be any faster way to do a back propagation than current methods (gradient descent)? My math skills are bit rusty now so I can't tell
47.
by rnosov 4y ago
For example, great apes look somewhat like humans but can't pass a Turing test. Should we give them same rights as humans too? It works other way too.
48.
by rnosov 4y ago
Bing bot never passed Turing test with a COMPETENT judge. Some people even thought that Eliza was sentient so not a high bar.
49.
by rnosov 4y ago
Pass a Turing test with a competent judge. ChatGPT or any other bot won't be able to pass it.
50.
by rnosov 4y ago
Airplane makes some sense too as according to Wikipedia: "Aéroplane" originally referred just to the wing, as it is a plane moving through the air. In an example of synecdoche, the word for the wing came to refer to the entire air
51.
by rnosov 4y ago
How different is this demo from just having input text box with suggestion to enter your name in it? If a user is foolish enough to divulge their name to pirate accented bot perhaps you might bypass it and ask for it directly.
52.
by rnosov 4y ago
GPT-3 performed really well on synthetic benchmarks. It was later made palatable for general public consumption. You might say that a LLM needs to be good on synthetic benchmarks first before you can make public facing chatbots with it.
53.
by rnosov 4y ago
They list 7B, 13B, 33B, 65B architectures. Presumably, they compare 65b one to GPT-3 175B. Chinchilla model which is about 70B outperformed a much larger GPT-3 model. So not that fantastical. EDIT: I stand corrected. They do compare 13B mod
54.
by rnosov 4y ago
It is a different backend but it supposedly should be roughly comparable to ChatGPT. Also, looks like it's both open source and requires a lot less hardware to run and train.
55.
by rnosov 4y ago
Looks like they are making ChatGPT clone that would be possible to run a single GPU. HN dream come true!
56.
by rnosov 4y ago
I guess it's really a layoff season in tech.
57.
by rnosov 4y ago
I see. You might want to explain it somewhere on your landing page. Perhaps a comparison of a more generic ChatGPT responses to the ones your app generates. Otherwise, it's not clear ( for me ) as to why should I download your app when
58.
by rnosov 4y ago
Hmmm, ChatGPT seems to work even as date coach. It gave me sensible answer using a rephrased prompt from your website: "Propose question to discuss, an activity to do and a self-reflection question for a date" Would your app do b
59.
by rnosov 4y ago
Jekyll site hosted on Github. Will probably work for the rest of your lifetime.
60.
by rnosov 4y ago
Correct output will be desirable. If you feed nonsense either human or AI generated you might break it.

More ›