Voice agents still don’t understand who’s speaking to them. That’s a huge gap compared with humans, hidden by all the “phone-call” demos. But that changes today! NVIDIA is open-sourcing Nemotron 3 Diarization: a model that can reliably track speakers in live conversations, under a commercial-friendly license! In my tests, the quality is really good with one-second speech chunks. So we can use it for voice agents! I tested it with Reachy Mini and speech-to-speech running on a DGX Spark. It’s super fun to see the robot notice a new voice, ask for a name, and remember it. The model has day-zero integration with Transformers! Kudos to the NVIDIA team for shipping useful tools for the whole community!

Really cool model! The pricing is really good for some applications as well, 3 usd per hour + extra for background thinking and tools. Not something to leave on all the time with reachy, but definitely something you can use for call centers.

I'm tired of LLMs solving math problems I don't understand. Please put them to work on something easy to understand, like an efficient exact solution to the traveling salesman problem. It would even have direct implications in day-to-day life.

wtf is this way to handle mathematicians work and scientific communication TLDR: Leven and Tristan worked over several months on one of the Millenium Prize Problems with various AIs to reach final interesting results. OpenAI apparently heard about it in the last days and prompted their latest models to work on the direction Leven and Tristan found fruitful. They then tried to push for controlling communication of the result and dropping Leven from authorship with some very bad taste social pressure. Hope this is not a glimpse of the future we’ll get in science research with these dominating players playing marketing games hurtful for the real scientific community.

Live action with real robots instead of animation: episode 3! This was the very first take. Pretty much everything went wrong and I almost stopped filming. Then I realized Microduck was giving an Oscar-winning performance. Enjoy!

Astra is a good model. It motivated me to spend my Saturday coding. It ate through 60% of my pro subscription using extra high and fast. It feels better than Sol in that it returns WAY faster. I was used to most tasks taking 20 mins, and now it's 4 mins. I also got it to spend more than one hour on a few. It also feels like it understands long context/hard tasks better. I was working with Sol on a big feature and had created 4 stacked PRs and was working them individually. Astra managed to look at the 4 PRs together and come back with cohesive thoughts and opinions on what should change. Overall happy, not feeling like AGI for coding, but a clear step-up over Sol. Given the clear difference in pricing, it will probably make me select models more. I used to only use Sol and get on the Pro subscription, but I don't think that will work out with Astra.