Eiso Kant

Eiso Kant is a cofounder of Poolside, where he has helped shape research around reinforcement learning for software development and the engineering infrastructure used to build models. He was Poolside’s CTO in December 2024 and was described as co-CEO in his July 2026 Latent Space interview. His consequential decisions include bringing distributed systems engineers into research from the outset and, with cofounder Jason Warner, choosing to release model weights and research.[3][1]

Kant describes model building as “90% engineering”: data ingestion, filtering, distributed training, reliability, and subsequent training stages form an industrial process. Poolside integrated distributed systems engineers into research from the beginning rather than adding infrastructure later. Its Model Factory optimizes the interval from a researcher’s idea to a trustworthy experimental result and its inclusion in the next model training.[1]

One important choice was to stream data directly into training. In Kant’s explanation, this avoids repeatedly packaging, tokenizing, copying, and distributing a complete dataset before a run can begin. A service called Blender configures source proportions, repetition, and shuffling; training can start while remaining data is still materializing. Immutable data and versioned code make experiments traceable down to individual tokens. Kant credits Nikolai, a head of applied research who joined from Yandex, with helping turn that infrastructure toward more rigorous experiments and ablations.[1]

In July 2026, Kant estimated that Poolside ran 10,000–20,000 experiments a month, while noting that he had not checked the latest count. He reported five weeks from training start to release for Laguna XS.2 and eight weeks for Laguna S. These intervals describe training and release, rather than the preceding research and infrastructure work. He also described agents writing code, launching jobs, evaluating results, and modifying pipelines, with researchers continuing to supply ideas and debug.[1][5]

Kant and his colleagues deliberately sought researchers outside the Bay Area, building a remote, global team in anticipation of competition for talent; he acknowledged later Silicon Valley hiring. He recalls an early team of about five spending three weeks stabilizing training in a codebase written from scratch. Solving the optimizer problem strengthened their confidence in working things out themselves, although Kant could not recall the exact bug.[1]

In their October 2024 joint announcement, Kant and Warner argued that software-development models need deterministic feedback to improve reasoning and coding. Poolside’s Reinforcement Learning from Code Execution Feedback gives models execution feedback on attempted solutions, allowing them to explore alternatives and learn from outcomes. The founders compared this with developers writing code, executing it, and revising a solution when it fails.[2]

Kant later relayed applied-research head Peng Ming’s interpretation that Laguna S’s gains came partly from verification and persistence: taking less for granted, backtracking, and avoiding premature declarations of success. Kant said these improvements came entirely from post-training. They made him more optimistic about how much useful work smaller models could perform, while he continued to support scaling larger models and acknowledged that Laguna S needed more efficient reasoning. He explicitly credited Peng Ming, Nikolai, and the broader team rather than claiming to have built the model himself.[1]

Kant regards coding and long-horizon software tasks as a route toward AGI because they require sustained reasoning, external tools, and data. In July 2026 he predicted that reinforcement learning would move earlier into training and argued that next-token prediction alone extracts too little from web data. Training smaller models from scratch served Poolside’s research iteration better than distilling its releases at that time: frequent runs helped distinguish which changes produced improvements. He acknowledged that the company had conducted distillation experiments.[1]

For complex agent tasks, Kant favors a minimal harness and a container in which a model can write scripts, use loops and conditionals, and interact directly with available systems. His criticism of MCP and predefined tool calls accompanies continued support for both at Poolside. He describes multi-harness training as “polishing”—helping behaviors transfer to other harnesses and improving the user experience, rather than supplying the main capability gains. His July priorities also reflected deliberate focus: vision understanding was a next step and absent from the models discussed, while audio was deferred.[1]

Kant said he and Warner reopened the open-source discussion at the beginning of 2026, while Poolside remained below the frontier and changing direction was easier. His stated preference was for 100 foundation-model companies rather than five, even if Poolside were among the five. He distinguished usable weights from research that helps others recreate the work: “giving someone the weights does not allow them ultimately to recreate what you’re doing, right?” The decision left questions about business models and future misuse risks unresolved.[1]

Kant and Warner’s April 28, 2026 announcement released Laguna XS.2 as open weights under Apache 2.0, explicitly crediting approximately 60 people in the Applied Research organization. A May 26 update announced Laguna M.1’s open-weight availability. For Kant, sharing research also meant giving back: he acknowledged that Poolside had benefited from other labs’ published work, including Chinese labs.[5][1]

His management views emphasize delivery and autonomy within constraints. Kant proposes measuring engineering productivity by the time from identifying something valuable for a user to shipping it, together with the number of people involved. He describes leadership of high-agency teams as establishing a common goal and explicit boundaries—such as a focus on reinforcement learning—while giving people freedom to pursue solutions within them.[1]

In August 2026, Kant and Warner said they had missed a financing window and lost a planned compute cluster: “We didn’t close it in time, and we lost the cluster.” Their statement, quoted in an August 21 Latent Space report, accompanied reporting of a nonexclusive NVIDIA licensing agreement and investment in Poolside. The founders were “not ready to share the updated vision.”[6]

Before Poolside, Kant had already spent years pursuing neural models for code. He recalls redirecting his startup after reading Andrej Karpathy’s 2015 article, “The Unreasonable Effectiveness of Recurrent Neural Nets.” At source{d}, he and his colleagues worked with RNNs, LSTMs, convolutional models, attention, and later Transformers, continuing until the end of 2019. Looking back, Kant said they had missed the importance of continuing to scale; he described spending $12 million with a team of about 40 as the biggest failure of his career. ChatGPT later felt to him like vindication.[1]

Between source{d} and Poolside, Kant founded and served as CEO of Athenian, a data-enabled engineering platform. That earlier work provides context for his later discussion of engineering productivity: alongside research on code models, his career included building a company concerned with understanding engineering work.[4][1]