User: Tell me a joke Assistant: I don't know anything about jokes. But I do know some cool Python tricks. Here's a simple example of using a loop to print a joke: def print_joke(): for i in range(1): print(f\"Joke number {i} is: {i} is {i + 1}.\") print_joke()

We’re hosting a hackathon at the frontiner tower in SF, to build economically valuable agents using Tzafon Computer. My good friend rapha will be the opening speaker talking about creating useful agents using RL. luma.com/economically-u…

Our engineering team have some new cool features, including spinning up 100 parallel browsers in < 5.2 seconds. With our new tiered per-second pricing, that's as low as <$0.01 in spend.

one thing I've noticed running an AI lab is research compounds relentlessly. small, nimble ideas accelerate progress far more than you'd expect.

Join Crusoe and NVIDIA for a tech talk in SF! Dive into the practical challenges AI builders face with leaders from Imbue, Flapping Airplanes, Google DeepMind & Tzafon. 📅 Feb 19 | 5:30 PM 📍 Jackson Square ➡️ Register: events.crusoe.ai/crusoesnextgen… #AI #ML

A glimpse into execution. Lightcone autonomously running a customer support workflow in a secure environment.

Personally super excited about this feature *still early* but essentially becoming the agents very own computer in the cloud and closing the loop of CUA <> Infra

Great paper!

We trained a computer use model on 1000x less data than standard approaches and got better results with RL. We do a deep technical deep dive on why it works, and some of the technical insights we got.

Everyone loves a fast product. That's why we decided to rebuild our Lightcone OS for significantly faster interaction – we now support full real-time streaming at 30 FPS. We did this to address one of the most common complaints with computer use – namely that it feels slow and

We're open sourcing Northstar CUA Fast, a frontier 4B open-source Computer Use Action (CUA) model, built for accuracy and long-horizon action planning.

Northstar is now available on the web in beta. Check in at any point — one look at what your agent is doing.

SFT for computer use saturates after 100–1000 examples. RL doesn't. 0.39 → 0.53 on aggregated UI benchmark. 20% absolute on OS-World Chrome - trained on environments with no resemblance to OS-World.

Today we're releasing the Lightcone SDK – the fastest path from idea to working computer automation, in Python or TypeScript. The most powerful SDK for computer use.

Our new responses SDK cuts the friction of computer use by deeply integrating the model with our state-persistent runtime. Would love to hear what people think! We're also working on rapidly improving the model based on the feedback we've gotten.