AGI: The Path Forward – Jason Warner & Eiso Kant, Poolside
In a live demo, Poolside's CEOs showcase their second-generation model, the Malibu agent, by migrating a complex codebase from ADA to Rust, including automated testing and iterative feature development. They outline their vision for achieving AGI through a full-stack approach combining proprietary models, reinforcement learning, and massive-scale compute, with plans for a public model release in early 2025.
Poolside’s mission is to bridge the gap between AI models and human intelligence. They build proprietary models from scratch, founded on the principle that next-token prediction must be paired with reinforcement learning (RL) to achieve significant advancements. Their current second-generation model is called the Malibu agent.
Live Demo: Malibu Agent
The presentation featured a live demo of the Malibu agent integrated into Visual Studio Code. The demo showcased the agent’s advanced capabilities in a high-consequence coding environment.
- Code Migration: The agent was tasked with migrating a codebase from ADA, a language used in critical government and defense systems, to Rust.
- It analyzed the existing ADA code to understand its functionality.
- It generated approximately 1,152 lines of Rust code, creating the necessary files and structure.
- The agent then automatically wrote and executed tests to verify the correctness of the newly generated Rust application. This highlights Poolside’s focus on building safe and verifiable agents for sensitive environments.
- Iterative Feature Development: Following the successful migration, the agent was asked to add a new feature: enabling command history (up-arrow functionality) in the Rust application’s command-line interface.
- The agent identified an appropriate external library (
rusty-line), modified the source code to integrate it, and rebuilt the application. - It then wrote and executed a shell script to confirm that the new feature worked correctly.
- The agent identified an appropriate external library (
Vision and Roadmap
Poolside is taking a full-stack, vertically integrated approach to scaling intelligence. This includes building their own multi-gigawatt data center campus in West Texas to support massive compute needs, with over 40,000 NVIDIA GB300s coming online.
- Beyond Coding: While the demo focused on code, the models are designed for all forms of knowledge work and are also described as “emotionally intelligent” and creative.
- Public Release: The next-generation model is planned for a public release in early 2025. It will be accessible via Poolside’s own API and through Amazon Bedrock, making it a new foundational model available to developers and enterprises.
- Future of Agents: The long-term vision is to evolve agents from performing tasks that take hours to handling complex work that spans days, moving progressively towards more autonomous systems and AGI-level capabilities.
Call to Action
Poolside expressed a strong interest in partnering with the developer community. They plan to provide access to their models at various checkpoints, enabling others to build on top of them and perform further reinforcement learning and fine-tuning for specific applications.