Descriptions:
Eiso Kant, co-founder and CEO of Poolside AI, joins the Latent Space podcast to discuss the launch of Laguna S, Poolside’s latest coding-optimized model, and the company’s foundational beliefs about what it takes to build AI capable of handling long-horizon software engineering tasks. Central to the conversation is Kant’s provocative claim that MCP and tool-calling frameworks are fundamentally inadequate for complex agentic work — arguing that frontier models are increasingly moving toward operating directly inside virtual machine environments, writing code to interact with systems natively rather than through abstraction layers.
Kant traces his path into AI to Andrej Karpathy’s 2015 “Unreasonable Effectiveness of Recurrent Neural Networks” essay, which prompted him to pivot his startup Sourced into building language models on code — years before transformers or scaling laws were widely understood. He reflects on the lesson that the field kept missing: just keep scaling. On training methodology, Kant shares that post-training RL delivered the biggest gains in Laguna S, but questions whether pure environment-based RL scaling is the correct path to AGI. He argues the web remains massively underutilized as a source of structured reasoning signal beyond simple next-token prediction, and characterizes “mid-training” as little more than a two-stage curriculum workaround for compute constraints.
The episode also covers Poolside’s open-source model releases, the distinction between distillation-driven and environment-driven improvement, and Kant’s broader view that the race to AGI may require rethinking how models extract knowledge from pre-training data.
📺 Source: Latent Space · Published July 22, 2026
🏷️ Format: Podcast
