Cartesia is building a muscle of developing and applying innovations orthogonal to model and data scale. We are excited to apply these ideas to an expanded set of domains in the coming months.
English
Arjun Patrawala
23 posts
In January, @_albertgu got deeply involved in our TTS research program. Over the following 5 months, we developed ~3 innovations (no changes to data or model scale), each targeting a fundamental issue plaguing our prior models. The results:
Karan Goel@krandiash
We released Sonic-3.5 and Ink-2, the #1 streaming models for text to speech and speech to text you can use in your voice agents today. New architectures enable new frontiers for speed and quality. We're now the only provider to have #1 models for both speaking and listening.
English
@krandiash The demos versus elevenlabs are pretty insane
Español
We've raised $100M from Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA.
Today we're introducing Sonic-3 - the state-of-the-art model for realtime conversation.
What makes Sonic-3 great:
- Breakthrough naturalness - laughter and full emotional range
- Lightning fast -
English
@alexwei_ When doing verifiable math problems, it's +EV for the model to guess an answer, even when unsure. I'd bet this "breakthrough" is fixing this fundamental flaw. E.g., update the reward to:
0 for wrong answer
a for saying "idk"
b for correct answer
for some values, 0 < a < b
English
On IMO P6 (without going into too much detail about our setup), the model "knew" it didn't have a correct solution. The model knowing when it didn't know was one of the early signs of life that made us excited about the underlying research direction!
Daniel Litt@littmath
One piece of info that seems important to me in terms of forecasting usefulness of new AI models for mathematics: did the gold-medal-winning models, which did not solve IMO problem 6, submit incorrect answers for it? 🧵
English
what i’ve been working on for the past year 🥹! excited for users to finally try it
Ivan Zhao@ivanhzhao
So for 5 years, “offline” has been the #1 request. Today, thanks to the perseverance of our engineering team, @NotionHQ finally works offline. Your ideas don’t need Wi‑Fi to exist! For Notion community: thank you for your patience while we built this right. This is a journey, I want to share what we had to invent to make this real... 1/n
English
Programming languages with pedantic, correctness-oriented compilers (Rust, OCaml) seem good in the AI era. Hard to trust vibe-coded mallocs and frees lol
English
@alexwei_ Choosing good values of a and b allows you to select the "confidence" the model must have before presenting an answer.
If a is close to b, the model must be quite confident because the risk of outputting the wrong answer is high and the reward is small.
English
What if we RL trained the model to use chain of thought to do better at the pretraining next-token-prediction objective? Rather than a 0/1 verifiable domain?
English
Vision transformers have high-norm outliers that hurt performance and distort attention. While prior work removed them by retraining with “register” tokens, we find the mechanism behind outliers and make registers at ✨test-time✨—giving clean features and better performance! 🧵
English
@sea_snell It’s most interesting that we reached for the over-engineered solutions first (PRMs, human-labeled CoT, etc.) before trying the simple idea
English
R1-zero is such a striking example of a discovery that’s blatantly obvious in retrospect, yet eluded so many for such a long time
English
Spotted at the door of berkeley artificial intelligence research today
English
I hate it when people do the thing where they use their teeth to take the food off the fork/spoon instead of their lips. Like brosef. That’s why Charles Darwin gave us lips.
English
If I was a talented artist I would draw fake stills that look like they’re from old Disney movies to confuse people
English