Apple's 3B LLM Outperforms GPT-4

> Apple has released ReALM (Reference Resolution As Language Modeling)

Interesting use of the word “released”. As far as I can tell, they published a paper (https://arxiv.org/abs/2403.20329), but didn’t release their model, nor their training set, nor their code.

All we have is a rough description of the approach and numbers measuring how well it works.

I wouldn’t know whether it’s easy or difficult to reproduce their results, but they don’t make that easy.

Edit: other articles use ‘reveal’, not ‘release’. That’s a bit better, but IMO still a bit too optimistic.

> Apple has released ReALM (Reference Resolution As Language Modeling), a new method for improving how AI understands references made during conversations and to items displayed on a screen or operating in the background. Imagine asking your phone to “call the top pharmacy on the list” without specifying which one – ReALM aims to figure out exactly what you mean.

Seems like they're talking about GPT-4 Vision. It's still impressive as it is, there is no need for clickbait...