The AI Frontier: from open weights to open research — Eiso Kant, Poolside AI

He described a shift from keeping heads-down to revisiting openness after concluding that a future dominated by only a handful of model builders would be dystopian; he said Poolside’s contribution is releasing both weights and research even while key questions remain (business model, misuse, government response).

He framed weights as “a binary” that doesn’t let others recreate the process, and argued that sharing lessons from “tens of thousands of experiments” is a more meaningful accelerant for other teams than weights alone, despite practical limits around datasets.

He said the hard part isn’t a single training run but building an end-to-end “industrialized process” that minimizes time from research idea to trusted result, enabling frequent model cycles (he cited five-week and eight-week pretrain-to-launch timelines).

He credited streaming data into training (rather than repackaging and copying datasets) and treating the data layer as immutable, paired with fully versioned code, as key enablers of fast iteration and the ability to reproduce runs “down to the single token.”

He described agents increasingly handling parts of the model factory—writing code, launching jobs, evaluating results—while humans remain in the driver’s seat, and said this shows early “twinklings” of recursive self-improvement.