Sarah Pan

student at @mit, prev at @answerdotai, @coatuemgmt

[Sarah Pan @spantacular]

@aaravwattal @Etched Congrats dude! This is huge

[Sarah Pan @spantacular]

@elainexliu So cool!

[Sarah Pan @spantacular]

@enthusednotebk /explain

[Sarah Pan @spantacular]

@aaravwattal Fire thread!

[Sarah Pan @spantacular]

Should test-time training receive a share of growing inference compute budgets? I’ve spent the past year pondering this question in depth, and have become convinced that it is one of the greatest ideas waiting to be explored at scale.

https://t.co/ocmPokw7gS

[Sarah Pan @spantacular]

@aidangch @attentioninc This is fire, congrats!

[Sarah Pan @spantacular]

@enthusednotebk 80/20 rule?!

[Sarah Pan @spantacular]

@AdamZweiger @thinkymachines @MIT Congrats!!

[Sarah Pan @spantacular]

Reminded me of another paper arguing transformers are topologically constrained to use the context window as working memory: https://t.co/0IrbdQrjEI

Super cool result that the J-Space can invoke concepts not present in context, a good case for increasing depth budget!

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[Sarah Pan @spantacular]

The next frontier is non-verifiable, and most of it is unlocked with whatever we mean by “taste.” Congrats to team!

[Soumith Chintala @soumithchintala]

Bridgewater, one of the worlds largest hedge funds, a Tinker customer talks through how they've carefully fine-tuned a model focused on what makes interesting financial news.
Their fine-tuned model is more effective and cheaper than any frontier model.

[Charlie O'Neill @oneill_c]

1/ You can shrink a language model's KV cache by 200×, in a single forward pass, and it still answers correctly.

At 256k context that's 36 GiB of cache down to ~360 MiB, with no change to the base model.

Here's how we did it 👇

[will savage @wavage_]

For the nuclear renaissance to deliver on its promise, we need radically better methods for developing advanced nuclear technology.

In my new paper, I use deep learning to rethink one of the most consequential problems in nuclear engineering: critical experiment design. (1/5)

[Austin Huang @austinvhuang]

Try https://t.co/4Fyz6npWFm

@spantacular and I worked on this to get people started with WebGPU for compute.

[Machine Learning Street Talk @MLStreetTalk]

A masterclass from @jeremyphoward on why AI coding tools can be a trap -- and what 45 years of programming taught him that most vibe coders will never learn.

- AI coding tools exploit gambling psychology
- The difference between typing code and software engineering
- Enterprise coding AND prompt-only vibe coding are "inhumane" i.e. disconnecting humans from understanding-building
- AI tools remove the "desirable difficulty" you need to build deep mental models.

Out on MLST now!

[Sarah Pan @spantacular]

Less operational overhead and more client outcomes.

Excited to help push forward the transformation of investment banking and financial workflows with Drake, Kent, and Esteban.

[Y Combinator @ycombinator]

.@MaywoodAI automates deal execution for investment banks, from decks to diligence. Dealmakers can focus on what they're best at: closing deals.

Congrats on the launch @DrakeAGoodman, @KentGoodman4, and Esteban!

https://t.co/x1Oyrp8NLT

[Nathan Cooper @ncooper57]

At @answerdotai, we integrate @stripe into lots of projects. Every time, I found myself doing the same dance: create product, create price, create checkout session. Then hunting docs for parameters for each. So we built FastStripe, a self-documenting Stripe SDK that's easy to use!

[Sarah Pan @spantacular]

@bhutanisanyam1 Congrats!!

[Andrej Karpathy @karpathy]

I often rant about how 99% of attention is about to be LLM attention instead of human attention. What does a research paper look like for an LLM instead of a human? It’s definitely not a pdf. There is huge space for an extremely valuable “research app” that figures this out.

[Sarah Pan @spantacular]

@AchyutaBot @cogconfluence @jacobandreas @ArthurConmy 🔥🔥

[Sarah Pan @spantacular]

@jeremyphoward 👀