Research at Kyutai by day, creative coding by night. On YouTube under my name and as part of Polylog
Paris, France
Joined February 2022
- Tweets163
- Following179
- Followers2K
- Likes668
Pinned Tweet
New side project: Samuel, a model I trained to mimic speech using Pink Trombone, a (pre-existing) silly vocal tract simulator. 🧵
Part of work I'm doing on a longer video for Samuel: samuel.vvolhejn.com/
The heatmap shows how far the sound is from the reference, as a function of the 2D tongue position. The sound you're hearing is before the gradient descent.
If it halts, all mathematics collapses.
Specifically: if it halts, simulating it is a proof. So it either doesn't halt OR everything we know about mathematics is wrong (=ZFC set theory is inconsistent). But we cannot prove it doesn't halt!
To clarify, it's not "undecidable" in the halting problem sense - it would've been more accurate to say "independent of ZFC".
Check out our video on Busy Beaver for more Turing Machine weirdness: youtube.com/watch?v=kMHCl2PG…
14
1,597
Takayuki Arai's channel has a bunch of these demonstrations, all of them done with a straight face:
youtube.com/watch?v=DyQ96oer…
If you don't know about the Busy Beaver rabbit hole, our new video is for you.
Released yesterday and it's at 100k views already!
Polylog is Richard Hladík, Václav Rozhoň, Tom Sláma, and yours truly. Video: youtube.com/watch?v=kMHCl2PG…
I tried out moving the virtual hands and it's so fun, it's funny to see how robotically they move in response to the keyboard controls
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
4
535
So apparently you can do text-to-speech with 2M-param models now?? Interesting that as you shrink the model, you lose expressiveness more than audio quality. Great job @Manu__Orsini
1
18
1,085
There's a new kind of coding I call "hand coding", where you actually write the code with your hands, git commit manually, and forget that coding agents even exist.
There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I'm too lazy to find it. I "Accept All" always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.
I used the recently released training code for @kyutai_labs Pocket TTS and combined it with Samuel, the model I trained earlier to reproduce speech using Pink Trombone. Essentially, Samuel acts as a neural audio codec.
2
22
5,785
We're collectively figuring out new etiquette around AI text, e.g. I get mad if someone directly replies to me with something from Claude, but "here is the AI-written details of the PR just in case" seems ok. I guess the former just feels like trying to pull a fast one on me.
1
7
803
After taking a wrong turn onto an unmapped path on a hike in Georgia 🇬🇪, I decided to fix it and make my first OpenStreetMap contribution. But it turns out the path is new and none of OSM's satellite images show it, and you're not allowed to trace Google's images! 🧵
1
7
489
I asked Reddit what to do but didn't get a useful map source. Then @atha_theo recommended I check @CopernicusEU and sure enough, the path is there! We can even determine it was made in October 2022.
1
5
224
Excited to hear that LLMs can finally automate having fun for me while I focus on more important things
i asked LLMs to have fun, write down what happened, then rate who had the most fun. the newer models seem to have less fun! many models played word games, but opus 4.6 won over and over by imagining brief worlds made up of contradictions living inside falling water droplets 🥺
19
605
Unlikely alliance possibility between self-hosting and AI: if you want to host your own calendar/email/cloud storage for data privacy, coding agents lower the technical barrier a lot. But many self-hosters are also anti-AI, creating a conflict of interest.
2
2
668
I can't stop thinking about the term "AI vegan" for people who don't want to use AI at all. Makes me wonder what other kinds of vegans are out there. Are self-hosters big tech vegans? Data vegans?
1
10
670
In many cases, Samuel learns to position its virtual mouth as a human would, even though its not explicitly trained to do so.