yeah the eclipse is amazing but have you stopped to consider how wild it is that the sun exists at all

Fine tuning popping up on my timeline again. I’ve said it before but I was wasting a lot of time ft open source models only to get marginal returns for products that did not have pmf or strong growth. There is a lot of overhead to ft, it’s not one and done - data changes, you…

Returning from an experimental ~2 week detox from the internet. Main takeaway is that I didn't realize how unsettled the mind can get when over-stimulating on problems/information (like a stirred liquid), and ~2 weeks is enough to settle into a lot more zen state.

I'm struck by…

really like seeing both of these together on my TL

If this were a science paper, you would expect a country that picks its science workforce at random as a “weak baseline” and a leading nation like the US to actively experiment towards state-of-the-art, or at least beat the baseline.

Not providing a guaranteed path for…

H1B lottery ❌

It was less than a 1 in 3 chance, but sucks anyway!

Can we localize the weights and mechanisms used by a language model to recite entire paragraphs of its training data?📄➡️🤖➡️📄
arxiv.org/pdf/2403.19851…

To find out, have a look at my Google AI intern project advised by Owen Lewis, Mitchell Gordon and Chiyuan Zhang.

Thread ⬇️

New Anthropic research paper: Many-shot jailbreaking.

We study a long-context jailbreaking technique that is effective on most large language models, including those developed by Anthropic and many of our peers.

Read our blog post and the paper here: anthropic.com/research/many-…

Localizing Paragraph Memorization in Language Models

Can we localize the weights and mechanisms used by a language model to memorize and recite entire paragraphs of its training data? In this paper, we show that while memorization is spread across multiple layers and model

Introducing `deep-seek` - an open source research agent designed as an internet scale retrieval engine.

It's a new approach to the current wave of answer engines. Instead of giving you one answer, deep-seek will retrieve an extremely comprehensive list of enriched results.

Move to NYC! We have bagels and culture and public transportation and Sasha Rush and bagels!

me irl

the world used to be glued together by code

now it'll be glued together by code and LLMs

what a bright future we have

I think we accidentally discovered a very effective benchmark for intelligence:

Ask your preferred LLM if it can draw a (theoretical) syntactic tree analysis of a sentence.

Very few (if not only one) models can do this. But those that can, do it with a high degree of…

Chess-GPT is a 50M parameter LLM playing at 1500 Elo. When it starts on a random board, its win rate drops from 70% to 17%. Does that mean it can't generalize?

No! In fact, we can restore much of its performance with one trick. We can also edit its internal board state.

🧵

There appears to be a mismatch between publishing criteria in AI conferences and 'what actually works'. It is easy to publish new mathematical constructs (e.g. new models, new layers, new modules, new losses), but as Apple's MM1 paper concludes:

1. Encoder Lesson: Image…