Jason Warner: A bold vision for automating software on the path to AGI
How poolside is scaling the LLM “wall” with reinforcement learning and synthetic data
You’re running a software tools company. Why the heavy emphasis on AGI?
We believe that the most valuable companies in the world are going to be the ones that push toward AGI. And we believe that we can get to AGI as quickly or quicker than everybody else because of our technical approach: reinforcement learning via code execution feedback (RLCEF). This is exactly what DeepMind did for AlphaGo, the AI which has mastered the game of Go, the complex, strategic board game in which two players vie for control of territory. We are applying the exact same idea to software.
DeepMind first trained an AI Go player on all human-played games available on the internet — all known Go games. It became a very good Go player but by no means world-class. Then they took that and put it in a self-play environment so it could explore all possible moves and learn what worked and what didn’t. That generated a ton of synthetic data because AlphaGo played and learned from so many games. That’s when it became a superhuman Go player. To apply this model to software, we start with a massive seed dataset: all the open-source code in the world. And at poolside we can do the exact same thing: First, build something that approximates a human developer. Then we play it in a software simulator environment, creating a flywheel effect that allows us to turn it into a super developer.
How do you give companies the benefits of this AI without asking for their most valuable asset, their data?
I don't believe it's appropriate for AI vendors to give customers intelligence in return for their data. It’s the devil's trade. So we’ve done the opposite. We're going to give customers our models in a way that they can run them wherever they want, and we will have no access to them or their data.
If we install poolside in an enterprise, it fine-tunes, indexes, and takes all of their internal data and puts it right next to the AI model. It exists inside their enterprise, and they can use it to build other tools. It's their system. I believe this concept will pick up in the next couple of years. I want poolside to be this software intelligence layer for the world.
.png)
.png)
