←
Tom Sander (Ph.D.)131 posts
Tom Sander (Ph.D.) avatar

Tom Sander (Ph.D.)

@RednasTom

AI Research Scientist at Meta Superintelligence Labs (FAIR). prev: Ph.D. @ FAIR & Polytechnique, engineering @ Polytechnique (X18), MVA @ ENS.

◷ Joined February 2022▧ 49 Photos and videos
Tom Sander (Ph.D.) avatar
Tom Sander (Ph.D.)@RednasTom·Nov 2

Introducing TextSeal2: distortion-free watermark with copula mixing and memoryless races TLDR: • Beats textGrain and SynthID-Text under speculative decoding • No extra latency, same acceptance rate • A single parameter fully controls the diversity/detectability tradeoff

1:17
Mark Zuckerberg avatar
⟳ Tom Sander (Ph.D.) retweeted
Mark Zuckerberg@finkd·Nov 9

Releasing Muse Ember today -- our smallest model that delivers higher intelligence per dollar than others in its class. Built for token-heavy work that agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Available today on the Meta Model API and in Muse Code.

Tom Sander (Ph.D.) avatar
Tom Sander (Ph.D.)@RednasTom·Nov 2

Paper: arxiv.org/abs/2610.40613 Code: github.com/facebookresearch/… Joint work with @pierrefdz and others at Meta Superintelligence Labs @AIatMeta. Feedback welcome.

Tom Sander (Ph.D.) avatar
Tom Sander (Ph.D.)@RednasTom·Nov 2

Beyond speculative decoding: +8 to 22 points over TextSeal on 100-token answers across three models, at equal diversity. It is level with textGrain without per-token optimization, and a single strength ρ sets the diversity–detection trade-off.

Post attachment
Tom Sander (Ph.D.) avatar
Tom Sander (Ph.D.)@RednasTom·Nov 2

Speculative decoding is the hard part: tokens that replace rejected drafts usually lose the watermark or lower acceptance. TextSeal2 runs one more race on the leftovers of the draft races. By memorylessness, they are fresh, so the output stays exactly on distribution.

Post attachment
Tom Sander (Ph.D.) avatar
Tom Sander (Ph.D.)@RednasTom·Nov 2

TextSeal2 fixes it with copula mixing: merge the two keys into one virtual key, then blend it with fresh noise through a Gaussian copula. Every token carries the full signal, and more keys, including a public one, can be added without diluting detection.

Post attachment
Tom Sander (Ph.D.) avatar
Tom Sander (Ph.D.)@RednasTom·Nov 2

Why a sequel? TextSeal routed each token to one of two keys by a coin flip, but the detector did not know which key was used. Averaging the two scores halved the signal. The bottleneck was that coin flip, not optimal transport, which is what OpenAI's textGrain uses.

Tom Sander (Ph.D.) avatar
Tom Sander (Ph.D.)@RednasTom·Nov 2

On 200-token answers (equal diversity, FPR 1e-3), the top schemes are close without speculative decoding: 91% for TextSeal2, 89% for textGrain and SynthID-Text. With it, TextSeal2 still detects 89%, while OpenAI's textGrain drops to 67% and Google's SynthID-Text to 50%.

Post attachment
Alex Sablayrolles avatar
⟳ Tom Sander (Ph.D.) retweeted
Alex Sablayrolles@alexsablay·Oct 6

We are back! Excited for this release and for the many things we have in our pipeline. ML4 is a 1T model with 49B active. Amazing performance across coding, agentic workflows, and multimodal understanding. Competitive with the strongest OSS models, weights drop end of month

Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

0:17
Cosmo Du avatar
⟳ Tom Sander (Ph.D.) retweeted
Cosmo Du@Answeror·Sep 15

muse code subs update contributor tier limits are now way way larger we will reset usage today for all subs, everyone starts fresh

Post attachment
Alexandr Wang avatar
⟳ Tom Sander (Ph.D.) retweeted
Alexandr Wang@alexandr_wang·Sep 13

i am so proud of the @Muse team. if you knew them you wouldn't be surprised muse came out the way it did. they are soulful and thoughtful people, hard workers, brilliant, fighters - the kind who take the hard problem, stay with it through every dead end, and don't let up until it's solved. i’m lucky to work with them. every detail of the product and the model was turned over in somebody's hands, worried over, chosen on purpose by a thoughtful soul. that’s the whole secret. good people, good work.

Nat Friedman avatar
⟳ Tom Sander (Ph.D.) retweeted
Nat Friedman@natfriedman·Sep 10

I hope you like it!

Started work at Meta this week. My job is to make amazing AI products that billions of people love to use. It won't happen overnight, but a few days in, I'm feeling confident that great things are ahead.

Alexandr Wang avatar
⟳ Tom Sander (Ph.D.) retweeted
Alexandr Wang@alexandr_wang·Sep 10

we just reset everyone’s muse token usage. enjoy!!!

0:06
Mark Zuckerberg avatar
⟳ Tom Sander (Ph.D.) retweeted
Mark Zuckerberg@finkd·Sep 8

Introducing Muse, the personal agent that understands your goals and works 24/7 to get things done for you.

Alexandr Wang avatar
⟳ Tom Sander (Ph.D.) retweeted
Alexandr Wang@alexandr_wang·Sep 8

1/ today we're rolling out Muse, our new personal ai assistant. Muse is always-on, wicked fast, can use a browser, connect to your apps, and is designed to be secure. try it now: muse.ai.

2:00
Mark Zuckerberg avatar
⟳ Tom Sander (Ph.D.) retweeted
Mark Zuckerberg@finkd·Jul 9

(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.