rfbr (@romfbr31) profile picture

MTS at @kyutai_labs, still the human in the loop

Paris
Joined December 2022
rfbr retweeted
i'm looking to join some designer communities, in Paris/Europe or anywhere online, where I could chat regularly with other creatives and swap ideas. any recommendations?
1
2
2
177
I turned TamaClaudchi into a full-on Claude Code mod 🥚 I can now monitor and control all my Claude sessions as little “tamas”, all ready to manage experiments for ICLR rebuttal 🫡 Switching between sessions, messaging across agents, tracking subagents, and more!
6
5
1
19
1,970
Neural Thickets (@yule_gan @phillip_isola): add random noise to an LLM's weights, keep the best copies, let them vote. Reasoning jumps, no training. Why does it work? It's sampling in disguise. Swap noise on billions of weights for one random vector on the logits: same gains 🧵
18
57
10
526
41,663
rfbr retweeted
This is not talked about enough but with @GradiumAI we released a chrome extension last week ! It can read, transcribe and translate any content you want. Pretty useful plus i had a lot of fun designing it!
3
8
344
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
261
826
288
7,383
1,047,405
- be Gradium - founded by the co-founders of Kyutai, one of the most cracked research labs in Europe. - kyutai created Moshi, the first real-time, full-duplex voice AI. Way before GPT voice and all current ones. - founders invented neural audio codecs, audio language models, and powered the first voice cloning, text-to-music, and speech-to-speech translation applications - pedigree hard to replace: ex DeepMind, Brain, FAIR, and Jane Street. - raised $100M seed backed by Nvidia, Eric Schmidt, Xavier Niel, FirstMark, Eurazeo, and DST. Yann LeCun and Guillermo Rauch as angels. - revenue within weeks of founding - super lean team. - shipped the fastest TTS out there - now hiring in Paris and the Bay Area. - another chance to make exceptional wealth. gradium.ai/about/careers
We're hiring! At Gradium, we’re building the best voice models out there, with people we enjoy spending time with, in and out of the office. Join us. gradium.link/careers
4
7
56
3,743
We're hiring! At Gradium, we’re building the best voice models out there, with people we enjoy spending time with, in and out of the office. Join us. gradium.link/careers
15
15
4
315
26,405
Gradium and @covaldev are hosting an evening in Paris for people building voice agents. Thursday, 5 Nov, 6 pm CET. We’ll start with a short fireside chat between our research teams, followed by drinks and conversation with both teams. RSVP: luma.com/wszf3b7v
8
4
20
2,370
Two papers from the lab will be presented at #EMNLP2026 in Budapest this week. Feel free to drop by Poster Session F on Tuesday to hear more about each work and talk with our researchers! @emnlpmeeting
3
7
37
3,005
We release several open artifacts: Our training and inference code, models, and a demo where you can drive the world models from your keyboard 👇 Paper: arxiv.org/abs/2610.30129 Blog, code & demo: kyutai.org/blog/2026-10-22-c… Authors: @RamaAdrien @YutongBAI1002 @nico_dufour @ptrkprz @david_picard
2
1
7
1,425
Many world models borrow their video autoencoder from text-to-video generators, and with it, a noticeable lag on every action. Today we release CADRE, open CAusal Diffusion REpresentation autoencoders that let real-time world models react fast, on a single GPU 🧵
6
32
4
160
9,252
A new frontier for on-device speech. We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date. Phonon-31m is available under Apache 2.0. gradium.ai/phonon🧵
167
284
162
2,193
368,702
Pocket TTS has been getting updates those last few days. While it was ~100M parameters, we release smaller versions, down to 1.9M: an 18.2M model with the quality of the original, a 6.9M model, and a 1.9M model that runs 103x faster than real time on one CPU core. All with voice cloning.
28
62
19
554
34,958
Introducing REKI, a 2.4B vision-language-action model pretrained on public data only. With a simple design, it rivals larger models trained on in-house data with complex multi-stage pipelines, setting a new Pareto frontier between performance and size. kyutai.org/reki
9
41
9
262
15,298
Many generative models work in an autoencoder's latent space; They train faster and generate better when this space is well organized. One way to organize it is distilling a pretrained model into it. But how can it work when the two spaces have different structures?
2
27
2
218
12,360
We release a technical blog on training PocketTTS, our 100M parameter on-device TTS, through drifting, a recent one-step generative objective from Deng et al. Less than 1% WER, high quality and voice cloning. To the best of our knowledge, it's the first speech model, and the first autoregressive model, trained this way.
13
80
7
537
24,729
Meta is making a tamagotchi for its AI? Cute, I've had a few on my desktop for a few weeks 🐣 Tamaclaudchi is a (cute) Claude session manager to see and monitor all your Claude sessions at once Made in collaboration with @romfbr31 and more features to come with @RamaAdrien
1
5
1
10
620
We're releasing Voice of Reason, a speech-native model that does math out loud. Give it a spoken problem without transcription nor text LLM in the loop, and it reasons and answers in speech. GSM8K goes from 27.3% for GLM-4-Voice to 77.1%. Link in 🧵
9
39
9
271
19,780
ECCV workshops scheduler webapp: kyutai.org/eccv_2026_worksho… Kyutai at ECCV 2026: 🏠 CASA (kyutai.org/casa) @royaleerieme @MoritzBoehle J. Marrie 👁‍🗨 OVIE (kyutai.org/blog/2026-04-14-o…) @RamaAdrien 🎲 The FID Lottery (kyutai.org/fid-lottery) @nico_dufour 🏎️ MIRA (kyutai.org/mira) @RamaAdrien @vvolhejn 🗣️ AURA (kyutai.org/aura, more details to come) N. Cohen
1
6
979