HNHacker News
TopNewBestAskShowJobs

gliptic

1,103 karma · joined March 22, 2009

submissionscomments
gliptic··on How to Run DeepSeek R1 671B Locally on a $2000 EPYC Server
Your best bet for 33B is already having a computer and buying a used RTX 3090 for <$1k. I don't think there's currently any cheap options for 70B that would give you >5. High memory bandwidth is just too expensive. Strix Halo might give you >5 once it comes out, but will probably be significantly more than $1k for 64 GB RAM.
gliptic··on OpenAI Furious DeepSeek Might Have Stolen All the Data OpenAI Stole from Us
So yes, it's a limitation of their own API at the moment, not a model limitation.
gliptic··on OpenAI Furious DeepSeek Might Have Stolen All the Data OpenAI Stole from Us
R1 is trained for a context length of 128K. Where are you getting 8K/32K? The model doesn't distinguish "thinking" tokens and "output" tokens, so this must be some specific API limitations.
gliptic··on SiFive's P550 Microarchitecture
> while it's also fabricated in an older node

Is this 7 nm node really older than raspberry pi 5's 16 nm node?

gliptic··on DeepSeek-R1
Where are you seeing Hyperbolic offering DeepSeek V3 API? I'm only seeing DeepSeek V2.5.
gliptic··on Nanoimprint Lithography Aims to Take on EUV
Apparently you don't believe in reading titles either.
gliptic··on Short Message Compression Using LLMs
Huffman still has a performance edge for static distributions. ANS bridges some of the performance gap between arithmetic coding and huffman.
gliptic··on Short Message Compression Using LLMs
Yes, the only exception is dictionaries used in preprocessing, but I think that's mostly a tradeoff to reduce the runtime.
gliptic··on Short Message Compression Using LLMs
This is especially true for instance when you have two or more tokens that are about equally likely, or one token that is virtually certain, which ranking would obscure.
gliptic··on Short Message Compression Using LLMs
This could be circumvented by _training_ the LLM on the fly on the previously observed file data. This is what Bellard's other NN compressor, nncp, does [1], which is currently #1 on Mahoney's benchmark [2]. Unfortunately this is too slow, especially running on the CPU as Hutter's challenge stipulates IIRC.

[1] https://bellard.org/nncp/

[2] http://mattmahoney.net/dc/text.html

gliptic··on Short Message Compression Using LLMs
The model gives you a probability distribution over the tokens. You could use that directly with arithmetic coding, but there are ways to convert that to a distribution over e.g. the next byte instead which would improve efficiency further by removing the redundancy in alternative token encodings. ts_zip does this, and README says this works similar to ts_zip.

EDIT: Hm, or maybe ts_zip uses just the token probabilities directly. I thought it was slightly more efficient about it.

"The language model predicts the probabilities of the next token. An arithmetic coder then encodes the next token according to the probabilities."

gliptic··on Short Message Compression Using LLMs
I'm pretty sure it doesn't use ranking. That leaves a lot of performance on the table. Instead you would use the actual predicted token probabilities and arithmetic coding.
gliptic··on OpenAI O3 breakthrough high score on ARC-AGI-PUB
"You can tell the RL is done properly when the models cease to speak English in their chain of thought" -- Karpathy
gliptic··on A liar who always lies says "All my hats are green."
Not D), because it says "The liar has at least one green hat." which isn't implied.
gliptic··on ChibiHash: Small, Fast 64 bit hash function
Maybe, but from what I've seen you run out of what theory can say about your non-crypto hash function pretty quickly.
gliptic··on ChibiHash: Small, Fast 64 bit hash function
Hash tests have always been targets. What else are you supposed to do for non-cryptographic hashes?
gliptic··on James Webb Space Telescope finds evidence for alternate theory of gravity
The weirdest thing about TeVeS IMO is that it adds additional fields that warp spacetime, so how is it not a dark matter theory?
gliptic··on Something weird is happening with LLMs and chess
Both an LLM and Stockfish would fail that test.
gliptic··on Memories are not only in the brain, human cell study finds
This is paranormal reincarnation mumbo-jumbo, not inheritance.
gliptic··on Diffusion for World Modeling
> I _think_ I understand the basic premise behind stable diffusion, i.e., reverse the denoising process to generate realistic images but, as far as I know, this is always done at the pixel level.

It's typically not done at the pixel level, but at the "latent space" level of e.g. a VAE. The image generation is done in this space, which has fewer outputs than the pixels of the final image, and then converted to the pixels using the VAE.

gliptic··on Large Text Compression Benchmark
The training cost is included because these NN compressors are running the training while they are compressing. They are general compressors, although the specific variants submitted here may be tuned a little to Wikipedia (e.g. the pre-processors).
gliptic··on Aro – Zig's new C compiler
Zig doesn't (and never will) have operator overloading, so using any user-defined numeric types is awkward. It also already has vector types.
gliptic··on Initial details about why CrowdStrike's CSAgent.sys crashed
No, it wasn't. Crowdstrike denied it had to do with zeros in the files.
gliptic··on Multiple airlines disrupted due to Microsoft Azure outage
Azure having problems is not a black swan event.
gliptic··on Google presents method to circumvent automatic blocking of tag manager
> Choose any path you want for setting up first-party mode. Examples of paths you might want to use include: /metrics, /securemetric, /analytics, or preferably a random alphanumeric string that you don't use on your website already.

Ah, yes, _preferably_ some random string that is hard to block rather than the descriptive ones.

gliptic··on A (hopefully) new compression algorithm that uses binomials
I mean without storing the freq table, which the last example you linked doesn't. Having to store a frequency table limits the encoder to do no better than the 0-th order entropy. Arithmetic coding can do arbitrarily complex prediction per symbol, see e.g. http://mattmahoney.net/dc/text.html#1072 which reaches the world record by training a neural network to predict the next symbol while it's encoding a file.

Not to say a 0-th order entropy coder could not be useful for certain applications if it was e.g. faster than, say, a range coder while maintaining similar ratios. But I doubt that is possible with this method.

gliptic··on A (hopefully) new compression algorithm that uses binomials
Sorry, I (and others) have explained several times what an equivalent arithmetic coding scheme would have to do to make the same assumptions (and have the same limitations) as your scheme.

It would have to assume there's _exactly_ N ones and N zeros in the bit string. Start with those counts:

remaining ones = N;

remaining zeros = N;

It would encode bit for bit. After every zero, it decrements the remaining number of zeros. After every one, it decrements the remaining number of ones. The probability of the next bit being zero or one is derived from the known remaining number of each, P(one) = (remaining ones / (remaining ones + remaining zeros)). When one of the counts reaches zero, the probability of the other becomes 100% and no more information need to be encoded.

Does the linked code do that? No.

I don't have a ready-made example because this is not an assumption about the input that is useful in the majority of cases. The linked example can encode _any sequence_, not just ones that have an equal number of ones and zeros.

If you spent time implementing this scheme you would learn a lot about this field.

gliptic··on A (hopefully) new compression algorithm that uses binomials
That isn't updating the counts after every symbol. Why don't you write one that does so you can compare them fairly?
gliptic··on A (hopefully) new compression algorithm that uses binomials
> Whenever you're compressing a file on your computer the frequencies are known beforehand.

Not at all. The best compressors learn to predict the frequencies as they go, they do not compute them up front and store them in the output.

gliptic··on A (hopefully) new compression algorithm that uses binomials
It's easy to see this octogonal wheel [rolling downhill] is faster than this round wheel [on a flat plain]!
← PreviousPage 2 of 12Next →