AGI: The Path Forward – Jason Warner & Eiso Kant, Poolside

Poolside: Bridging Models and Human Intelligence - A Detailed Summary

Key Concepts:

• Poolside: A company focused on building AI models from scratch, specifically combining next-token prediction with reinforcement learning to create more capable agents.
• Malibu Agent: Poolside’s second-generation AI model, demonstrated in the session.
• ADA (Ada Lovelace): A programming language used for critical infrastructure, serving as a test case for the agent’s code conversion capabilities.
• Rust: A modern systems programming language used as a target for code conversion from ADA.
• Reinforcement Learning (RL): A machine learning technique used to train agents to make sequences of decisions.
• GB300s: NVIDIA’s latest generation of GPUs, crucial for scaling model training.
• Vertical Integration: Poolside’s strategy of controlling the entire stack, from data centers to models and interfaces.

1. Core Mission & Technological Foundation

Poolside’s primary goal is to “close the gap between models and human intelligence.” Founded 2.5 years ago, the company’s core belief is that while next-token prediction (the foundation of many large language models) is a significant breakthrough, it requires integration with reinforcement learning to achieve truly intelligent behavior. They are building their models from scratch, currently on their second generation, the Malibu Agent. This approach differentiates them from relying on existing models like those from OpenAI or Anthropic.

2. Demonstration: ADA to Rust Code Conversion

The core of the presentation was a live demonstration of the Malibu Agent operating within Visual Studio Code. The agent was tasked with converting a codebase written in ADA, a programming language commonly used in critical infrastructure (satellite systems were specifically mentioned), to Rust.

• Process: The agent initially analyzed the ADA codebase, providing a summary of its function. Upon request, it began converting the code to Rust, displaying a live diff view of the changes.
• Technical Details: The agent autonomously wrote approximately 1152 lines of Rust code. It then generated and executed tests (bell commands) to verify the conversion’s functionality. The inference speed was highlighted as being very fast.
• Challenges & Iteration: The initial conversion revealed a bug (an “unwrap” issue), which the agent was then prompted to fix. It identified a necessary package ("rusty line") and integrated it, successfully resolving the issue and passing tests. The agent also added a feature to allow command history cycling using the up arrow key, demonstrating its ability to understand and implement user requests.
• Security Considerations: A key point emphasized was the need for stringent permissions and control when deploying agents in high-consequence environments like defense and government sectors. Direct access to data sources is restricted, requiring careful configuration.

3. Beyond Coding: Emotional Intelligence & Creative Capabilities

Poolside’s models are not limited to coding tasks. They possess broader knowledge and are described as “emotionally intelligent” and capable of creative tasks like writing poetry and stories. The speaker shared a personal anecdote about using the agent to write love letters to his wife.

4. Future Roadmap & Scalability

Poolside is preparing to release its next-generation model publicly early next year. This model will be accessible through:

• Poolside’s own API.
• Amazon Bedrock API.
• Integration with existing engineering assistant tools (Cursor, Windsurf, Cognition, Replit).
• Integration with creative applications (Harvey, writing tools).

A significant investment is being made in compute infrastructure, with over 40,000 GB300 GPUs coming online. This increased compute power will enable further scaling and improvement of the models.

5. Vertical Integration & Partnership Opportunities

Poolside adopts a “full vertical” approach, controlling all aspects of the technology stack, from data centers in West Texas to model development and user interfaces. They are actively seeking partnerships with developers and researchers who are interested in building on top of their models, particularly those already working with reinforcement learning and fine-tuning open-source models like Quen, Gemma, and Minimax. They are open to collaborations ranging from early checkpoint access to utilizing the current models.

6. Historical Context & Founding Principles

The founders, Jason and ISO, have a history dating back to a failed acquisition attempt of ISO’s company by GitHub in 2017. Jason, as CTO of GitHub, recognized ISO’s early work on code completion using LSTMs (before the widespread adoption of transformers). Their shared vision for the future of AI, particularly the importance of reinforcement learning, ultimately led to the founding of Poolside in 2022. ISO specifically noted that the initial focus on reinforcement learning was a contrarian view at the time but is now gaining wider acceptance.

7. Long-Term Vision & The Future of AI Agents

ISO envisions a future where AI agents can handle increasingly complex and long-duration tasks, evolving from asynchronous operations (hours) to tasks spanning days or even years. He emphasized the importance of focusing on the fundamentals of intelligence and scalability, while acknowledging that the interface to these agents will continue to evolve. He believes the current era represents an “awkward teenage years” for AGI, bridging the gap between raw intelligence and economically valuable applications.

Notable Quotes:

• Jason: “Poolside exists to close the gap between models and human intelligence. That’s literally it.”
• ISO: “The world till date was built by intelligence. The world in the future will be built on top of intelligence.”
• ISO (regarding the initial approach to reinforcement learning): “Obsessing and focusing on reinforcement learning combined with LLMs felt like one of the most contrarian opinions in the world, but I think today it's absolutely not.”
• ISO (when asked about founding the company): “Oh, god damn no.” (His initial reaction to the idea).

8. Synthesis & Conclusion

Poolside is a vertically integrated AI company focused on building highly capable agents by combining next-token prediction with reinforcement learning. Their Malibu Agent demonstrates impressive code conversion abilities, particularly in challenging environments like those found in the defense sector. With significant investments in compute infrastructure and a commitment to open partnerships, Poolside aims to play a key role in advancing the state-of-the-art in AI and enabling the development of valuable, real-world applications. Their emphasis on fundamental intelligence and scalability positions them as a significant player in the evolving landscape of AI agents.