Poolside AI, founded by Jason Warner (former CTO of GitHub) and Eiso Kant, is building foundation models trained specifically for software engineering. Unlike general-purpose models that happen to code well, Poolside trains on code execution and uses reinforcement learning from actual code outcomes, creating models that fundamentally understand how software works.
Core Products
Poolside’s models (codenamed malibu) are trained with a unique approach: rather than just predicting the next token in code, they learn from actually executing code and observing outcomes. This reinforcement learning from code feedback approach aims to produce models that understand software engineering at a deeper level than text prediction alone can achieve.
Competitive Position
Poolside’s differentiation is methodological: training specifically on code execution outcomes rather than just code text. This approach, if successful, should produce models that are better at debugging, understanding runtime behavior, and writing correct code on the first attempt. The founder’s GitHub background provides unique insight into developer workflows and needs.
Recent Developments
Poolside raised a massive $500 million Series B in late 2024 at a $3 billion valuation, one of the largest rounds for a company with limited public product availability. The company has been building its training infrastructure and models, with initial results shared with select partners. The large funding reflects investor conviction in the code-specialized model thesis.
Outlook
Poolside represents a bet that purpose-built code models trained with execution feedback will outperform general-purpose models for software engineering. The company has substantial capital to pursue this thesis but must demonstrate results that justify its valuation against rapidly improving general models from OpenAI, Anthropic, and others. The AI coding market is large enough to support multiple winners.