Poolside's Path to AGI
How reinforcement learning, defense sector focus, and vertical integration position Poolside in the race to build advanced AI coding agents.
Executive Summary
In this revealing talk, Jason Warner (former GitHub CTO) and Eiso Kant (co-founder of Poolside) pull back the curtain on their strategy to build AGI through AI coding agents. Their approach combines three core differentiators: (1) Reinforcement learning + LLMs (a contrarian bet made 2.5 years ago, now validated by the market), (2) Defense sector focus (working in high-consequence code environments with strict security requirements), and (3) Full vertical integration (from multi-gigawatt data centers to proprietary models).
The conversation covers everything from technical architecture (why RL is essential for reasoning) to business strategy (why defense is the right market) to infrastructure (40,000+ GB300 GPUs). Whether you're an AI engineer, founder, or investor, this talk provides a window into how one well-funded startup is competing with giants like OpenAI and Anthropic—not by outspending them, but by outmaneuvering them.
Key Insights
The Contrarian Bet That Became Industry Consensus
Two and a half years ago, Poolside bet everything on reinforcement learning combined with language models. At the time, this was one of the most contrarian positions in AI. Today, it's becoming the industry standard—validated by OpenAI's o1, DeepSeek's R1, and Anthropic's approaches.
Notable Quotes:
"We're building our own models from scratch to do this. We're based on the idea 2 and a half years ago that we thought next token prediction was an amazing technological breakthrough, but it need to be paired with reinforcement learning really to make that leap."
"In the first 18 months of this company, you know, obsessing and focusing on reinforcement learning combined with LMS felt like one of the most contrarian opinions in the world, but I think today it's absolutely not."
Defense Sector: High-Consequence Code as Differentiation
Instead of competing head-to-head with OpenAI in general coding, Poolside pursued a deliberately different strategy: focusing on defense and government sectors where 'high-consequence code' requirements create natural barriers to entry. This isn't just about revenue—it's about proving capabilities in the most demanding environments.
Notable Quotes:
"We're working in high consequence code environments for the last year inside the the government and the the defense sector."
"One of the tricky parts about working on inside the defense sector and things like that is you can't have an agent that's just going to run around and do stuff. I mean like I can't walk into half of these buildings. You can't give an agent access to these data source and just say, 'Hey, go nuts.' You need to have the right permissions. You got to actually really ratchet these things down to do things inside those environments that you know they feel comfortable with."
From Hours to Days: The Long-Horizon Task Frontier
Poolside claims their agents can already run tasks for hours—demonstrating state persistence, error recovery, and autonomous planning. But the real frontier is days-long autonomous tasks, which would represent an industry-first capability. This isn't just about longer tasks—it's about fundamentally different agent architecture.
Notable Quotes:
"We have agents running that are doing tasks for for hours, and I think in the near future, we can see a world where they're able to start doing tasks in days in the coming years."
Vertical Integration: Multi-Gigawatt Campuses and 40,000 GPUs
Poolside is pursuing full-stack vertical integration similar to xAI's Memphis cluster and Microsoft/OpenAI's Stargate. The claim of 'over 40,000 GB300s' and a 'multi-gigawatt campus in West Texas' signals massive infrastructure investment. This isn't just about scale—it's about creating long-term economic moats through infrastructure control.
Notable Quotes:
"And now that it's unlocked for us and and with a large number of over 40,000 GB300s coming online, we see how we can start scaling up some of those models to get even further in their level of capabilities and software development and other types of long horizon knowledge work."
"It's why we go full vertical. It's why we go from our multi gigawatt campus in West Texas where we're building out data centers building out models."
The 'Awkward Teenage Years' Ahead of AGI
Eiso Kant coins a powerful metaphor for the current phase of AI development: we're in the 'awkward teenage years ahead of AGI'—capable but immature. This frames the current moment not as AI's toddler stage (as Sam Altman suggests) but as a transitional period where builders are creating the companies and applications that will bridge the gap to human-level intelligence.
Notable Quotes:
"We're entering these kind of awkward teenage years ahead of AGI where everybody in this room is building out incredible companies and applications is bridging this gap of what it really takes to make intelligence that in its raw form actually be valuable"
Practical Applications for AI Engineers
Reinforcement Learning for Reasoning
Poolside's RL+LLM approach is now the proven path for advanced AI agents. Pre-train on next-token prediction, then fine-tune with reinforcement learning for reasoning capabilities.
Security-First Agent Design
Defense sector requirements apply broadly: RBAC for agent permissions, audit trails for all actions, human-in-the-loop approval for sensitive operations, air-gapped environment compatibility.
State Persistence for Long Tasks
Hours-long autonomous tasks require robust state management, error recovery mechanisms, memory optimization for long contexts, and task planning/decomposition.
Infrastructure as Competitive Moat
Vertical integration from data centers to models reduces long-term costs and creates barriers to entry. 40,000+ GPU clusters are becoming necessary for frontier model training.
Competitive Landscape
Poolside isn't trying to be a "better OpenAI"—they're pursuing a deliberately different strategy through defense specialization and vertical integration. Here's how they compare to other major players:
| Company | Product | Approach | Differentiation | Status |
|---|---|---|---|---|
| Poolside | Malibu Agent | RL+LLM, from-scratch | Defense focus, vertical integration | Private beta |
| OpenAI | GPT-4o + o1 | LLM + reasoning | General purpose, largest user base | Public API |
| Anthropic | Claude 3.5 + Code | Claude + tools | Safety, enterprise focus | Public beta |
| Cognition | Devin | Autonomous agent | First to demo hours-long tasks | Public beta |
| Cursor | Composer | Multi-model IDE | Best developer experience | Public |
Key Insight: Poolside's defense focus and RL head start create differentiated market position, not technical superiority across all dimensions.
Key Concepts & Definitions
Reinforcement Learning (RL)
Training AI through rewards/punishments rather than just pattern matching. Critical for reasoning and planning capabilities.
GB300/H200 GPU
NVIDIA's flagship data center GPU with 141GB HBM3e memory, 4.8TB/s bandwidth. ~$35k each. Essential for frontier model training.
Long-Horizon Tasks
Autonomous AI tasks that run for hours or days. Requires state persistence, error recovery, and memory management beyond typical chat interactions.
Vertical Integration
Owning the entire stack from power → data centers → GPUs → models → applications. Creates economic moats but requires massive capital.
High-Consequence Code
Software where failures have serious impacts (defense, aerospace, medical). Requires rigorous testing, security, and audit trails.
ADA Programming Language
Legacy defense/satellite programming language known for safety. Poolside demoed converting ADA to Rust autonomously.
About the Speakers
Jason Warner
Co-founder, Poolside (ex-GitHub CTO)
Former CTO of GitHub where he oversaw technical strategy during Microsoft acquisition. Now building Poolside to close the gap between models and human intelligence through reinforcement learning and vertical integration.
Eiso Kant
Co-founder, Poolside (ex-GitLab)
Entrepreneur and technical leader with background in developer tools. Co-founded Poolside with Jason Warner after recognizing the potential of reinforcement learning combined with language models for AGI development.
Funding: $500M+ from Bain Capital, DST Global, StepStone