0-_-0

4 ms·

31.
by 0-_-0 4mo ago
No, it means the motor is smaller and it can be put into the wheel
32.
by 0-_-0 4mo ago
And this is just fraud that was done with incompetence, so easily caught. How much is done competently?
33.
by 0-_-0 4mo ago
This can't be used to save VRAM in practice. To generate a new token with the primary model, you first need to decompress the cache, which involves regenerating the whole sequence from scratch. I.e. generate 1 million tokens with the s
34.
by 0-_-0 4mo ago
You can use the original model to compress the kv cache and get ∞x compression, since the prediction is perfect. The cost is time, and I don't see how this could be worth it.
35.
by 0-_-0 5mo ago
Agents don't buy stuff they see in an ad
36.
by 0-_-0 5mo ago
I thought it can do multiple tokens at a time
37.
by 0-_-0 5mo ago
3.6 already supports multi token generation AFAIK
38.
by 0-_-0 5mo ago
How about next year?
39.
by 0-_-0 6mo ago
What else would it be?
40.
by 0-_-0 6mo ago
I don't think he wants to buy one
41.
by 0-_-0 6mo ago
I wonder why that's not supported by Biscuit despite dozens of other functions being present. E.g. download and display an image once an hour. Can't wake from sleep regularly?
42.
by 0-_-0 6mo ago
You're being downvoted, but you're right. Determinism is a different concept and doesn't characterise LLMs well. You can have deterministic random number generators for example.
43.
by 0-_-0 6mo ago
https://en.wikipedia.org/?title=Emergent_behavior
44.
by 0-_-0 6mo ago
Some of my friends are beige
45.
by 0-_-0 6mo ago
Obscurity makes no sense on a world with patents.
46.
by 0-_-0 6mo ago
How about caveman+be brief?
47.
by 0-_-0 6mo ago
I have no idea how much that is so it doesn't terrify me, but a 47 cm dragonfly is terrifying!
48.
by 0-_-0 6mo ago
There was a recent HN post where they checked the oxygen hypothesis (insects grew this large due to higher atmospheric oxygen levels) and it didn't hold up
49.
by 0-_-0 6mo ago
What I hear is that getting your network to work on AMD is a huge pain.
50.
by 0-_-0 6mo ago
It's the cumulative weighting based on the softmax output? Is it per layer?
51.
by 0-_-0 6mo ago
Training algorithm is the programmer, weights are the code that you run in an interpreter
52.
by 0-_-0 6mo ago
Weights are the source, training data is the compiler
53.
by 0-_-0 6mo ago
Weights are the source, training data is the compiler.
54.
by 0-_-0 6mo ago
"We also haven't seen any holes that couldn't have been dug by an elite construction worker." Therefore excavators are nothingburger
55.
by 0-_-0 6mo ago
So then the question is: which pipeline is used less? Bit or add?
56.
by 0-_-0 6mo ago
Pun intended?
57.
by 0-_-0 6mo ago
1 mm thickness is a fine trade-off
58.
by 0-_-0 6mo ago
A 256 element lookup table is much bigger than a simple multiplier
59.
by 0-_-0 6mo ago
Exactly, that's why they are not building more capacity and that's why RAM prices will stay up for years.
60.
by 0-_-0 6mo ago
You want to make multiplication cheap, it's not just about compression

More ›