jack morris Profile picture
research @cornell @meta // language models, information theory, science of AI
Aug 13 14 tweets 5 min read
OpenAI hasn’t open-sourced a base model since GPT-2 in 2019. they recently released GPT-OSS, which is reasoning-only...

or is it?

turns out that underneath the surface, there is still a strong base model. so we extracted it.

introducing gpt-oss-20b-base 🧵Image
Image
if you're not familiar with base models: here are some samples comparing our new model to the original!

we basically reversed the alignment part of LLM training, so we have something that produces natural-looking text again.

the outputs can be pretty random 🤷‍♂️ Image
Aug 8 14 tweets 5 min read
curious about the training data of OpenAI's new gpt-oss models? i was too.

so i generated 10M examples from gpt-oss-20b, ran some analysis, and the results were... pretty bizarre

time for a deep dive 🧵 Image here's a map of the embedded generations

the model loves math and code. i prompt with nothing and yet it always reasons. it just talks about math and code, and mostly in English

math – probability, ML, PDEs, topology, diffeq
code – agentic software, competitive programming, data scienceImage
Image
Jun 24 6 tweets 3 min read
In the beginning, there was BERT.

Eventually BERT gave rise to RoBERTa. Then, DeBERTa. Later, ModernBERT.

And now, NeoBERT. The new state-of-the-art small-sized encoder: Image the key insight, i think, is using an optimal depth-to-width ratio for the transformer architecture. and training on good data. a lot of good data.

even though NeoBERT has slightly more parameters, it's still faster AND more effective than ModernBERT for long sequences: Image
Jun 20 9 tweets 4 min read
NEW RESEARCH: Approximating Language Model Training Data from Weights

ever wonder how much information is available in an open-weights model?

DeepSeek R1 weights are 1.2 TB...

what can we learn from all those bits?

our method reverses LLM finetuning to recover data: 🧵Image
Image
to do this, you need TWO sets of model weights: the initial model and a finetune

this is realistic. open-weights models often come with two checkpoints

instead of one-shot generating data from weights, we select data from the web with gradients that point along the model diff Image
Jun 3 10 tweets 5 min read
new paper from our work at Meta!

**GPT-style language models memorize 3.6 bits per param**

we compute capacity by measuring total bits memorized, using some theory from Shannon (1953)

shockingly, the memorization-datasize curves look like this:
___________
/
/

(🧵)Image
Image
this all started from a quest to come up with a proper measurement of model memorization

it's hard to compute *per-example* memorization, because models "share" info between datapoints

so we start with random uniform strings, where sharing isn't possible. and we get this: Image
May 21 8 tweets 3 min read
excited to finally share on arxiv what we've known for a while now:

All Embedding Models Learn The Same Thing

embeddings from different models are SO similar that we can map between them based on structure alone. without *any* paired data

feels like magic, but it's real:🧵 a lot of past research (relative representations, The Platonic Representation Hypothesis, comparison metrics like CCA, SVCCA, ...) has asserted that once they reach a certain scale, different models learn the same thing

this has been shown using various metrics of comparison
Jan 3 4 tweets 1 min read
no AI here, just the coolest paper i've seen in a while Image turns out the way paints mix (blue + red = purple) is much more complicated than how light mixes (blue + red = pink)

they have to use a little bit of nonlinear modeling to capture this, and "add" paints in this nonlinear latent color space Image
Oct 4, 2024 7 tweets 3 min read
We spent a year developing cde-small-v1, the best BERT-sized text embedding model in the world.

today, we're releasing the model on HuggingFace, along with the paper on ArXiv.

I think our release marks a paradigm shift for text retrieval. let me tell you why👇Image Typical text embedding models have two main problems
1. training them is complicated and requires many tricks: giant batches, distillation, hard negatives...
2. the embeddings don't "know" what corpus they will be used in; consequently, all text spans are encoded the same way
Apr 4, 2024 6 tweets 2 min read
New Research:

a lot of talk today about "what happens" inside a language model, since they spend the exact same amount of compute on each token, regardless of difficulty.

we touch on this question on our new theory paper, Do Language Models Plan for Future Tokens?Image I think our most crucial finding is that although humans think far ahead while speaking (especially while doing complex reasoning problems) it turns out that transformer language models.... don't seem to do that.

they just predict the next token. Image