gliptic

14 ms·

61.
by gliptic 2y ago
I don't see why arithmetic coding wouldn't match this if the probabilities are suitably updated after every symbol. If you encode 8 bits, 4:4 ratio, after the first bit is encoded, either 1 or 0 has a remaining count of 3, etc. An
62.
by gliptic 2y ago
500 Internal Server Error. HN hug?
63.
by gliptic 2y ago
Punctuated equilibrium is not about higher mutation rates or even faster evolution, but about long periods of _stasis_ interrupted by "normal rate" evolution due to an upset in the equilibrium such as a sudden geographic isolation
64.
by gliptic 2y ago
Decoherence is what GP means with information leakage. The quantum state of the object becomes correlated with the environment. The article is saying this can be avoided in their system for a _microsecond_, still long enough to be potential
65.
by gliptic 2y ago
NFAs are decidable. As you said, a NFA can be converted to a DFA which is decidable. Maybe you mean they are non-deterministic (that's in the name)?
66.
by gliptic 2y ago
mIRC having a security hole with DCC SEND is exactly why some routers decided to filter it. EDIT: Perhaps I'm misremembering this. It seems it might have been due to some bugged port remapping code in the routers.
67.
by gliptic 2y ago
> I decided to open-source the version I built in September 2024. I think this is the most impressive part!
68.
by gliptic 2y ago
Zig doesn't have anything like the safety guarantees of Rust.
69.
by gliptic 2y ago
IRC was what was initially called multiplayer notepad.
70.
by gliptic 2y ago
llama.cpp isn't a model though.
71.
by gliptic 2y ago
Incorrect. The output of the decoder LLM is logits that are then divided by the temperature and passed through softmax to give the probabilities. You can't actually set temperature to 0 (division by zero), but in the limit where temper
72.
by gliptic 2y ago
Yes.
73.
by gliptic 2y ago
https://en.wikipedia.org/wiki/Z-order_curve
74.
by gliptic 3y ago
And where does it say that's the instruct model?
75.
by gliptic 3y ago
But context modelling isn't dictionary compression. Parent is correct that dictionary compression itself (e.g. LZ4) isn't very useful for floating point.
76.
by gliptic 3y ago
But it's using larger weights during training, not just the trits.
77.
by gliptic 3y ago
Lots of things running under Xwayland are unusable until the implicit/explicit sync issue between nvidia/wayland is resolved.
78.
by gliptic 3y ago
They also trained 3B with 2 trillion tokens. > The number of training tokens is a crucial factor for LLMs. To test the scalability of BitNet b1.58 in terms of tokens, we trained a BitNet b1.58 model with 2T tokens following the data reci
79.
by gliptic 3y ago
> Also, a quick reminder that Perplexity 12 is really terrible. The 3B model had a perplexity of 9.91, less than LLaMa 1 in fp16.
80.
by gliptic 3y ago
After the discovery of the greenhouse effect in Earth's atmosphere of course, as per the xkcd.
81.
by gliptic 3y ago
The unknown part is drift caused by measurement error, which will stack up very quickly.
82.
by gliptic 3y ago
From the repo: > Run LLMs locally on Cloud Workstations. Uses: > Quantized models from [Huggingface] > llama-cpp-python's webserver But sure, the blog post doesn't mention it.
83.
by gliptic 3y ago
> Google also took llama.cpp and used it in one of their Github repos without giving credit. Again, not cool. Are you talking about gemma.cpp? Then no, they didn't.
84.
by gliptic 3y ago
It's not nearly that bad. You effectively did the sqrt twice by setting m = 2^80 when it should be 2^160. You need 5.4e19 artifacts before the chance is 1 in a billion.
85.
by gliptic 3y ago
It = JAXA.
86.
by gliptic 3y ago
It's called Cislunar space if you please.
87.
by gliptic 3y ago
Rod logic seems to fit the bill ( https://anthony-zhang.me/blog/rod-logic/ ).
88.
by gliptic 3y ago
It's not. That's an old picture from last year.
89.
by gliptic 3y ago
No, Liero is a real-time game somewhat inspired by Worms perhaps, but mostly by Molez. This has several versions, some of which are (reversed engineered) open source versions of the original game.
90.
by gliptic 3y ago
They seem to be comparing against SDXL 1.0 at 512x512 which makes no sense to me as SDXL 1.0 is horrible at 512x512. > Using four sampling steps, ADD-XL outperforms its teacher model SDXL-Base at a resolution of 512^2 px

More ›