Gzip and KNN Outperforms Transformers on Text Classification (twitter.com)

Screenshot of twitter.com: Gzip and KNN Outperforms Transformers on Text Classification

First screen, captured 3 years ago.

Looks similar

Screenshot of twitter.com: Neural networks in the 1990s

Neural networks in the 1990s (twitter.com)

122 points by jrott 3 years ago | 75 comments

Screenshot of twitter.com: 50% on HumanEval with just 1.3B model

50% on HumanEval with just 1.3B model (twitter.com)

58 points by sytelus 3 years ago | 5 comments

A Baby GPT (twitter.com)

396 points by tosh 3 years ago | 89 comments

Tweet by Andrej Karpathy showing a tiny GPT with two tokens and context length 3 viewed as a finite state Markov chain, explaining learned transition probabilities.

gpt transformers markov-chains andrej-karpathy

Screenshot of twitter.com: SynJax: Jax library for efficient probabilistic modeling of structured objects

SynJax: Jax library for efficient probabilistic modeling of structured objects (twitter.com)

49 points by pizza 3 years ago | 10 comments

Screenshot of twitter.com: XB, or eXtreme Bullshitting: a less misleading name for LLMs

XB, or eXtreme Bullshitting: a less misleading name for LLMs (twitter.com)

83 points by gmac 3 years ago | 60 comments

Screenshot of twitter.com: John Carmack on the Similarity of Human Learning and LLMs Training

John Carmack on the Similarity of Human Learning and LLMs Training (twitter.com)

38 points by amrrs 3 years ago | 39 comments

Looks similar

Neural networks in the 1990s(twitter.com)

50% on HumanEval with just 1.3B model(twitter.com)

A Baby GPT(twitter.com)

SynJax: Jax library for efficient probabilistic modeling of structured objects(twitter.com)

XB, or eXtreme Bullshitting: a less misleading name for LLMs(twitter.com)

John Carmack on the Similarity of Human Learning and LLMs Training(twitter.com)