← back
AGI: The Path Forward – Jason Warner & Eiso Kant, Poolside
Takeaway
Poolside is betting on RL-augmented next-token-prediction foundation models trained from scratch for high-consequence coding and long-horizon knowledge work.
Summary
- Poolside trains its own foundation models pairing next-token prediction with RL; second-gen 'Malibu Agent' demoed via VS Code
- Demo: convert a 1,100-line Ada codebase to Rust including REPL with rusty-line history, all inside their custom IDE with live diff view
- Works inside defense/government with high-consequence-code constraints — sandboxing and permissions get real ratcheting
- Bringing on 40,000+ GB300s for next-gen model launch early next year, will be available via own API and AWS Bedrock for tools like cursors/windsurfs/Harveys
foundation-modelscoding-agentsrl
Original description
In Poolside's first ever public conference demo, Poolside's CEOs present their vision and roadmap towards achieving AGI-level capabilities for knowledge work.