brandon wang

398 posts

brandon wang

@fluorane

... | prev undergrad @miteecs and @mitbiology, @cartesia @janestreetgroup | usa imo 2019

california Katılım Nisan 2021

292 Takip Edilen1.2K Takipçiler

Sabitlenmiş Tweet

brandon wang@fluorane·11 Tem

happy to announce that we've gotten rid of tokenizers! especially excited with what we've replaced them with: end-to-end trainable modules that not only learn to group characters into (sub)words, but can iterate to group words into phrases and further higher-order concepts see @sukjun_hwang's thread for more details 👇

Sukjun (June) Hwang@sukjun_hwang

Tokenization has been the final barrier to truly end-to-end language models. We developed the H-Net: a hierarchical network that replaces tokenization with a dynamic chunking process directly inside the model, automatically discovering and operating over meaningful units of data

English

760

77.7K

brandon wang@fluorane·4h

@_isomerism you might also be the first person to use chatgpt work mode

English

Oliver Ye@_isomerism·5h

am i the first person to run doom in chatgpt work mode

English

167

brandon wang@fluorane·6h

back then the usa ioi community was approximately the subset of the math olympiad community that was interested in coding/informatics (less true now, there is more specialization), e.g. basically everyone there went to mop, started out with mathcounts, etc. to your other point if you did imo you probably self-selected into a group that was not super into cs/more into pure math

English

1.4K

qt cache🪷@frontier_foid·6h

idk why people refer to their backgrounds as competitive math when most of the folks here were much better at ioi than imo (im actually not sure if any of them even went to the imo?). in fact, the best imo participants of this era haven't done much of note in ai!

Jesse Zhang@thejessezhang

x.com/i/article/2076…

English

123

21.1K

brandon wang@fluorane·8h

there's a lot of incentive to come up with RL algorithms that work on long-running tasks with complex outcomes (because autonomous ML research agents also require running long/expensive experiments etc.) so i feel like we'll get some pretty useful stuff soon will require doing smarter things than grpo family but seems not fundamentally limiting

English

177

Sanjay Srivatsan@SRsrivatsan·1d

While we're thinking about RL in bio -- biology won't have the immediate rewards that enabled coding agents. Even if you could instantaneously spin up an experiment, diffusion and cell division both place hard time bounds on speed. RL here will have through a massive number of simultaneous trajectories, with models forecasting time evolving states, and reward scored on achieving a desired end state.

English

6.2K

brandon wang@fluorane·1d

@allylyq 4.6 certainly has some very useful capabilities

English

224

alli 📍 back in SF@allylyq·1d

fable is smarter but opus 4.6 was wiser.

English

1.1K

brandon wang@fluorane·4d

@NepalAadim > Theory is all about building a model that is predictive. this is a very physics-y perspective, people have spent centuries doing math for the love of the game and not for its predictive value

English

600

Aadim @ ICML 2026 🇰🇷@NepalAadim·5d

I feel like theory has no place in ML. SVMs and diffusion theory is the only proper cute math in ML, everything else is just philosophy. Theory is all about building a model that is predictive. Unfortunately, most ML theory is incapable of making predictions.

English

11.8K

brandon wang@fluorane·5d

@elipughresearch is that a data flywheel i see

English

Eli@elipughresearch·5d

We've been evaluating ink-2 on real voice-agent calls and measuring turn-taking performance, hard entity recall, and accuracy for diverse accents. Ink-2 is built with these things in mind, and it's amazing to see the difference they make in call quality.

English

Eli@elipughresearch·5d

Ink-2 has been at the top of several external benchmarks over the last few weeks, but how does that transfer to great experiences in your voice-agents?

Cartesia@cartesia

For voice agents, STT has to nail three things - accuracy, turn detection, and latency. If any one falls short, the experience breaks down: the agent misunderstands, interrupts, or just feels slow. We built Ink-2 to lead on all three. Here’s how it stacks up against other providers: link in comments.

English

1.4K

brandon wang@fluorane·6d

i am once again offering a prize to whomever outdrinks albert, this time 20k korean won

Albert Gu@_albertgu

Cartesia is hosting an ICML party tomorrow (Thursday) night! it'll go late, but come early or i may have to bounce you again luma.com/gy1b4ryq

English

9.5K

brandon wang@fluorane·6d

@AnjneyMidha many are saying

brandon wang@fluorane

welcome back terra luna openai model naming clownery stays undefeated

English

432

brandon wang@fluorane·7 Tem

@zacharylipton well anthropics complaints read more realpolitiky (it is bad for china to distill) than moral (distillation is theft)

English

539

Zachary Lipton@zacharylipton·6 Tem

AI companies complaining about distillation is the single greatest act of hypocrisy in the history of humanity.

English

561

43.1K

brandon wang retweetledi

Victor Li@victor_ljz·6 Tem

We will be at ICML in Seoul to present dnaHNet in Oral + Poster sessions! Swing by and say hi if you wanna learn about the current SOTA in DNA foundation models :) Oral session: Hall D2 at 10 AM, Jul 7 Poster session: Hall A #900 at 2 PM, Jul 7

Arc Institute@arcinstitute

Most genomic AI models use fixed rules to process DNA into chunks, imposing arbitrary boundaries on a sequence with its own biological structure. @arnavshah0, @victor_ljz, and team developed dnaHNet, a tokenizer-free foundation model that learns its own segmentation from scratch, supervised by @_albertgu, @genophoria, and @BoWang87.

English

20K

brandon wang@fluorane·4 Tem

@arnavshah0 🫡

QME

Arnav Shah@arnavshah0·4 Tem

@fluorane Brandon u better not miss me

English

121

brandon wang@fluorane·4 Tem

i'm in korea for icml! come chat with me about doing cool new science with ai :)

English

2.6K

brandon wang@fluorane·1 Tem

@KabirGoel bait posting for views/ugc has been so terrible for the information ecosystem

English

118

kabir@KabirGoel·29 Haz

kinda annoying how they take headlines from 2 years ago and repackage them as "breaking news"

Polymarket Money@PolymarketMoney

BREAKING: Anthropic CEO says open-source AI is moving down a “very dangerous path.”

English

842

brandon wang@fluorane·30 Haz

@unixpickle i feel like chemistry is in this weird space where in principle everything is deterministic (and thus simulate/compute-able) but in practice intractable. and so a lot of the game is finding useful approximations (mo theory, arrow pushing, your favorite reaction mechanism etc)

English

729

Alex Nichol@unixpickle·29 Haz

I felt in high school chemistry like there were no useful or reusable abstractions, e.g. I couldn't actually predicting what molecules could exist / how they would behave from first principles, and it's all based on a ton of observations empirically -- is this actually true?

English

329

39.1K

brandon wang@fluorane·28 Haz

@juli_li_ have you tried chatgpt image gen you can just link ur paper and tell it which figures to use

English

594