HackerTrans
TopNewTrendsCommentsPastAskShowJobs

nathan-barry

no profile record

Submissions

Can gzip be a language model?

nathan.rs
15 points·by nathan-barry·há 27 dias·2 comments

Large Text Compression Benchmark

mattmahoney.net
3 points·by nathan-barry·há 28 dias·1 comments

Language Modeling Without Neural Networks

nathan.rs
2 points·by nathan-barry·há 3 meses·0 comments

Quines in Every Programming Language (Rosetta Code)

rosettacode.org
2 points·by nathan-barry·há 3 meses·0 comments

Mercury Edit 2: Fastest next-edit prediction with a diffusion LLM (221ms)

inceptionlabs.ai
1 points·by nathan-barry·há 3 meses·0 comments

Noise schedules considered harmful (2024)

sander.ai
1 points·by nathan-barry·há 5 meses·0 comments

Show HN: Diffuji – a diffusion-powered instant camera

diffuji.com
8 points·by nathan-barry·há 5 meses·5 comments

Generating Shakespeare Without Neural Networks

nathan.rs
5 points·by nathan-barry·há 6 meses·0 comments

Recursion and Induction: A Self-Contained Course Using ACL2

acl2.org
2 points·by nathan-barry·há 6 meses·0 comments

Fidenae amphitheatre collapse (27 AD) – 20k dead

en.wikipedia.org
2 points·by nathan-barry·há 6 meses·0 comments

Show HN: Tiny Diffusion – Minimal diffusion LM in 364 lines

github.com
2 points·by nathan-barry·há 6 meses·0 comments

[untitled]

1 points·by nathan-barry·há 7 meses·0 comments

Anchored Diffusion Language Model

anchored-diffusion-llm.github.io
2 points·by nathan-barry·há 7 meses·0 comments

Text Diffusion Models Are Faster at Writing Code

nathan.rs
3 points·by nathan-barry·há 7 meses·0 comments

Inside ThunderKittens' Python Bindings

enbao.me
2 points·by nathan-barry·há 7 meses·0 comments

Kickstart.nvim: A minimal, single-file starting point for Neovim configuration

github.com
1 points·by nathan-barry·há 8 meses·0 comments

How we made AWS Trainium 17x faster (for conv1d)

charleshong3.github.io
1 points·by nathan-barry·há 8 meses·0 comments

Show HN: Tiny Diffusion – A character-level text diffusion model from scratch

github.com
172 points·by nathan-barry·há 8 meses·20 comments

The Manuscripts of Edsger W. Dijkstra

cs.utexas.edu
268 points·by nathan-barry·há 8 meses·125 comments

Streaming DiLoCo: Towards a Distributed Free Lunch

arxiv.org
2 points·by nathan-barry·há 8 meses·0 comments

comments

nathan-barry
·há 27 dias·discuss
LLMs are very good at lossless compression via arithmetic coding. But I didn't know that it was possible to go the reverse direction (do language modeling via a compressor). It's not super great quality, but I'm surprised it worked! Other compression algorithms (like PPMd) use variable n-grams under the hood, and should be much better (although less interesting due to already containing basic language models internally).
nathan-barry
·há 8 meses·discuss
Actually NVIDIA made one earlier this year, check out their Fast-dLLM paper
nathan-barry
·há 8 meses·discuss
Hey, I’ll add the MIT licenses later today!
nathan-barry
·há 8 meses·discuss
Yes, this is the case. During training, the model will get a sequence of text (ex, 512 tokens long) with a percentage of them masked out (with a special <MASK> token). It learns how to unmask those tokens to construct the original text.

In the case that you mentioned, if we had 4 <MASK> tokens in a row, all we are doing for decoding is predicting what those 4 tokens should be.

Generally, this does not seem to be a significant problem, as there are usually multiple ways to express an idea in varying lengths. Also, with confidence-aware parallel decoding, it can usually avoid the scenario you mentioned, as focusing on decoding the highest confident tokens will generally avoid such scenarios with a well trained model.