Inside the Model Factory — Eiso Kant, Poolside AI

Full analysis

Your draft

Poolside co-founder Eiso Kant walks through how a lab with fewer than 70 researchers and 10,000 H200 GPUs shipped Laguna S 2.1 in eight weeks, open-weighted, claiming it beats models 2-3x its size on coding. The model has 118 billion parameters but only activates 8 billion at a time (that's the "sparse mixture-of-experts" trick: you keep a big pool of specialist sub-networks but only wake a few per token). The real payload for builders isn't the model. It's the claim that a small team can now industrialize frontier-adjacent training, and that most of the gains come from behavior (persistence, verification, backtracking) rather than raw scale.

This is briefing mode. Nobody is being asked to buy anything. The question is: does the "Model Factory + behavioral gains + open weights" thesis change how you should plan your model stack over the next year?

Reversibility: Mostly Type 2 for the reader. Trying Laguna S on a workload is a weekend, not a marriage. The Type 1 bet buried inside is architectural: if you build your agent stack around a specific tool-calling paradigm that Kant says is dead in 12 months, that's expensive to unwind.

Forcing function: None hard. But NVFP4 (NVIDIA's new 4-bit number format for Blackwell chips) and the Blackwell migration wave are real 2026 events that make some of these claims checkable soon.

The Skeptic. "Beat Thinking Machines' trillion-parameter model on several evals at launch day" is exactly the sentence that ages badly. Launch-day, self-selected evals from the lab that built the model are marketing, not measurement. Coding benchmarks are the most contaminated corner of the eval world. Training data leakage inflates them routinely. This only works if "behavioral" gains (persistence, verification) transfer off the benchmark to your messy production traffic, and nothing in the episode gives you a reason to assume they do. That's an aesthetically appealing story. It flatters everyone who suspected scale was hitting diminishing returns, which is precisely why I distrust how easily it's being repeated. For the PM: a lab says its small model out-punches giants; treat that like a startup quoting its own numbers until someone independent checks.

The Researcher. The genuinely interesting claim isn't the leaderboard. It's Peng Ming's decomposition: that Laguna S wins through verification and not-declaring-victory-early rather than "more intelligence." If true, that's an RL post-training story, and it's consistent with the DeepSeek-Zero direction Kant cites (RL signal migrating upstream into pre-training). The saved Domyn-Small paper (a 10B open reasoning model) and the MoE-routing-as-Huffman-code paper both point the same way: the field is squeezing structure and behavior, not parameters. What I'd want: the eval deltas isolated by behavior, not aggregate scores. "Persistence beats intelligence" is a testable hypothesis, and nobody in this episode tested it in public.

The Open-Source Advocate. This is the part that matters beyond Poolside. A sub-70-person lab shipping open weights that fit on a single DGX Spark at 30-40 tokens/sec is a real datapoint against the 3-5 company oligopoly. Kant explicitly calls on researchers to leave and start competitors, and calls out model terms-of-service that ban using outputs to build foundation models, a clause OpenAI, Anthropic, and others actually enforce. Pair Laguna S with Domyn-Small (MIT license, European) and you see a pattern: capable open reasoning models are arriving from outside the Big Five monthly now. For the PM: the menu of "good enough" models you can run yourself is growing fast, which is leverage on price.

The Compute Pragmatist. The least-hyped insight is the RL batch-size wall. Pre-training scales by adding GPUs. RL doesn't. It's batch-size constrained, so throwing more H200s at it buys you nothing past a point. Kant's proposed unlock is applying prefill/decode disaggregation (splitting the "read the prompt" phase from the "write the answer" phase across different hardware) to RL training, not just inference. That's a genuinely novel infra bet, and he says no one's doing it. That reframes silicon evaluation: if you're looking at Groq, Etched, or Blackwell for an RL-heavy roadmap, benchmark training throughput under batch constraints, not inference latency. FP8 today, NVFP4 next. The precision floor keeps dropping, and that's where the cost curve actually bends.

Where they split:

1. Skeptic vs. Researcher on the behavioral thesis. Both find it interesting; they disagree on whether it's real or a narrative. The Researcher wants the isolated per-behavior deltas; the Skeptic bets those deltas shrink the moment an independent party runs a held-out agentic eval on non-coding tasks. The whole "you're over-buying dense models" recommendation hinges on this one unverified decomposition.

2. Open-Source Advocate vs. Compute Pragmatist on the moat. The Advocate sees the Model Factory as proof the barrier to entry is falling. The Pragmatist points at 10,000 H200s and the RL batch-size wall and says: the barrier moved, it didn't fall. You now need elite training-systems engineering and a nine-figure cluster, just fewer bodies. "70 researchers" is the headline; "10K H200s" is the fine print.

3. The tool-calling prediction. Kant says stuffed-tool system prompts vanish in 12 months, replaced by models writing code in minimal shell containers. The Builder likes the direction (code-writing agents are more flexible), but MCP has enormous enterprise install-base momentum. Betting your architecture on that timeline is the Type 1 risk hiding in a Type 2 episode.

What it hinges on: Two facts, both checkable. (1) Do Laguna S's coding wins survive an independent, contamination-controlled eval, and do they transfer to non-coding agentic work? (2) Does the behavioral-gains story hold, i.e., can a small MoE match dense models 2-3x its size on your task, not the leaderboard's? The council leans genuinely intrigued but unconvinced: the compute and open-weights observations are solid and actionable now; the "small beats big" thesis is a hypothesis that benchmarks haven't actually proven. Before shifting any deployment plan, run Laguna S and one comparable open model (Domyn-Small, a Qwen or Mistral variant) through your own agentic eval harness on your own tasks. Cheap to do; it's open weights.

The prediction. The behavioral claim is soft, but there's a hard one worth making: the tool-calling shift Kant predicts.

Prediction: By the release of Laguna's next major model (or ~mid-2027, whichever comes first), stuffed multi-tool system prompts will NOT have disappeared. MCP-style predefined tool lists will still be the dominant integration pattern across major model providers and enterprise agent frameworks.

Confidence: Medium. Install-base momentum beats architectural elegance on 12-month timelines.

Why: Kant's own logic (models "want to write code, not call tools") is technically appealing, but MCP shipped into Anthropic, OpenAI, and a wide enterprise tooling ecosystem through 2025-2026, and enterprise buyers standardize on auditable, permissioned tool interfaces precisely because free-form code execution is a security and governance nightmare. The mechanism that kills stuffed tool-prompts is models writing arbitrary code in minimal containers, which is exactly what compliance teams block. Twelve months is far too short to unwind a standard that regulated buyers just finished adopting; architectural end-states usually arrive years after the lab that spots them declares them inevitable.

Revisit by 2027-07-01: We're right if MCP/predefined-tool lists remain the mainstream agent-integration pattern at major providers. We're wrong if the top labs have visibly moved default agent behavior to free-form code-in-container execution and tool-stuffed system prompts are deprecated.

The compute and open-weights takeaways don't need a prediction. They're already true and already actionable. Go run the eval.