Karan Goel - Cartesia

Name: Karan Goel - Cartesia
Uploaded: 2024-06-12T16:45:48+00:00
Duration: 2 min 8 s
Channel: Karan Goel
Description: In this presentation Karan Goel outlines Cartesia’s mission to create private, personal AI assistants that can understand and generate any modality…

12 Jun 2024 · 2 min

Watch on YouTube

2:08

In this presentation Karan Goel outlines Cartesia’s mission to create private, personal AI assistants that can understand and generate any modality instantly on a phone. He explains that current transformer‑based models struggle with long context windows, limiting their efficiency on edge devices. Drawing on research from his Stanford PhD, Goel introduces state‑space models as a fundamentally different architecture that processes information continuously without fixed context windows, enabling efficient handling of very long sequences. Cartesia, a six‑month‑old startup backed by investors and angels, has just launched its first product—a real‑time voice‑generation model built on state‑space models—and is actively hiring engineers and a product lead to scale the offering and incorporate user feedback.

Chapters

Hey everyone, I want you to imagine a world where AI is unbelievably efficient, where you can understand and generate any modality instantly on your phone for any context window. In this kind of world, you would have a personal assistant or an embodied assistant that would be very useful, that would be personal to you and private to you, and that would help you with a lot of problems. Our belief is that this future is possible, but it requires fundamentally new architectures for intelligence. Karan Goel: Today, all the models that you're using are built on this architecture called the transformer. Now, transformers are amazing, but they have a fundamental limitation. They slow down as they see more context. This means that you can't process a trillion tokens of video on a phone-sized chip with a transformer, which is something that you do every year as a human. My co-founder Albert and I were doing our PhDs at Stanford where we built a new paradigm for foundation models called state-space models. Now, state-space models are a very different way of building foundation models. They're incredibly efficient compared to Transformers, and the reason they're so efficient is because they have no context windows. They simply take information and process it in and then gist it as it comes in. This means that you can actually model very long context windows very efficiently. At Cartesia, we're taking this technology to build real-time foundation models that will power long-lived intelligence on any device. We've been around 6 months. We're continuing to build out our team, and we have a great set of investors and angels that are supporting us. We actually just launched a product today, and we are, which is a first model built on state-space models that's incredibly efficient and used for real-time voice generation. And, um, we're hiring, so we have a lot of roles open in engineering. We're also looking for our first product hire because we have a lot of real users that are coming on board, including today, and we're getting a lot of feedback, and we'd love someone to help us sift through that and actually build a great product and models. Thank you.