The AI Frontier: from open weights to open research — Eiso Kant, Poolside AI
Summary
Poolside argues its durable asset is the model factory, not any individual checkpoint. With fewer than 70 researchers plus roughly 35 engineers, it says it runs 10,000–20,000 experiments a month, moved one model from pre-training to delivery in five weeks and Laguna S in eight, and had no meaningful on-call event this year beyond early-run configuration fixes. The goal is to make each checkpoint “just another launch” and the model merely “an artifact of someone’s process.”
The open-source turn is a pre-frontier governance choice: Eiso Kant would rather live among 100 foundation-model companies than be one of only five. Weights alone are “a binary” that cannot reproduce a lab; publishing the experiments and engineering lessons is the more consequential contribution. Poolside does not pretend the economics or safety policy is solved—it still lacks a complete open-foundation-model business model and expects misuse thresholds to require reevaluation eventually.
Laguna S makes a concrete small-model claim: behavior can compensate for a surprising amount of parameter count. The sparse model has 118B total parameters and 8B active, runs at roughly 30–40 tokens per second on a DGX Spark, reportedly solved Erdős 397, and was working through a request to make a macOS Wi-Fi scanner without external libraries or internet access. The Poolside speaker attributes its edge to “more verification, less taking things for granted, not declaring victory early, and being way more persistent,” while explicitly saying Fable and 5.6 remain better overall.
If persistence matters more than raw intelligence for most knowledge work, the economically optimal model size may arrive much sooner than Poolside previously assumed. Poolside frames knowledge work as roughly 25% of the global economy, or $25 trillion, and wonders whether its ROI peak sits around 1 trillion, 5 trillion, or 10 trillion parameters—not another two or three orders of magnitude away. That would support model commoditization and open weights, but Poolside rejects becoming “king of open source small models” and says it must still scale with frontier competitors.
Poolside’s deeper research bet is that pre-training must learn to reason, rather than leaving behavior formation almost entirely to post-training environments. The Poolside speaker expects reinforcement learning to move earlier than today’s “mid-training,” calls distillation and environments useful “drugs,” and argues that next-token prediction is “barely squeezing” the knowledge encoded across the web. The uncertainty is important: the speaker is “not sure” that simply multiplying task environments is the path to AGI.
The harness is a training surface, but Poolside does not want it to become the product boundary. It uses a minimal six-tool coding harness and a small amount of multi-harness “polishing,” while Swyx pushes Poolside to optimize directly for OpenCode or Hermes if it truly calls itself a model company. Eiso’s stronger prediction is that MCP-style tool menus will yield to models writing conditional code inside virtual machines: “in 12 months,” he expects system prompts stuffed with 20–40 tools to largely disappear.
Calendar time—not headline GPU count—is the binding variable in Poolside’s race to AGI. Its current footprint is described as a 10K H200 cluster; the larger Laguna M entered a 39-day pre-training run, while reinforcement learning remains the wall-clock bottleneck because limited task supply constrains batch size. Eiso therefore highlights lower-precision training, prefill/decode hardware specialization, and mixed-chip RL as more consequential than the nominal cost of a hero run.
Poolside opposes restricting today’s open models, but Eiso rejects the absolutist claim that every future capability must remain open forever. He prefers democratically accountable, capability-specific rules over unilateral company decisions, while a co-host notes that one jurisdiction cannot “err on the side of safety” for the world. The cautionary analogy is cigarette-advertising regulation: a good safety rule can still entrench an oligopoly—and freezing model competition in 2026 would resemble “chapter 14 of the most dystopian sci-fi novel.”
Deep dive
Not yet available upstream; scheduled sync will retry.