Poolside’s Eiso Kant on Why Small Models Can Do Big Things (And Why MCP Is “Stupid”)

Poolside released Laguna S, a 118B parameter model (8B active) trained in 8 weeks that outperforms models 2-3x its size on coding tasks. In this wide-ranging Latent Space conversation, Poolside cofounder Kant breaks down how they built it—and shares contrarian takes on tool-use, open source, and the race for AGI that challenge industry orthodoxy. Our recent stories on Poolside AI and founder Eiso Kant will provide more context.

From the our favorite Podcast team – Latent Space Podcast

1. Model building is 90% engineering, not research
The “model factory” is Poolside’s real competitive advantage—a unified end-to-end system that:

• Streams training data (instead of pre-staging datasets)
• Treats code as version-controlled, data as immutable
• Runs 10-20K experiments/month with perfect reproducibility
• Ships models on 5-8 week cycles (vs. 6+ months at traditional labs)

This infrastructure is what enables agents to autonomously improve the training pipeline itself. Zero on-call incidents for Laguna S—the system just works.

2. “MCP and tools are stupid”
Kant’s spicy take: models should write code in sandboxed environments, not call 50 tools from a system prompt. He predicts this will be obvious in 12 months—models want freedom and will write conditional logic (if/for loops) to solve complex tasks more efficiently than chained tool calls.

“The model is using code to do complex tasks. And when it uses code, it is not one or two tool calls. It actually starts using if statements and for loops.”

3. Smaller models can squeeze way more capability
Laguna S proved that capability isn’t just about parameters. Post-training behaviors matter more:

• Persistence (not taking shortcuts)
• Verification (not declaring victory early)
• Backtracking when wrong

Poolside’s co-head of applied research observed: “A lot of the gains in Laguna S come not from more intelligence, but more from different behavior, more verification, less taking things for granted, not declaring victory early, and being way more persistent.”

Kant’s implication: the optimal model size for knowledge work (coding, legal, accounting) may be much smaller than previously thought, which makes commoditization and open-sourcing of models more viable economically.

4. Open models need to exist—100 companies, not 5
Kant’s core mission: he’d rather live in a world with 100 foundation model companies than 5, even if Poolside isn’t one of the 5. Restrictions on open models (by companies or governments) are dangerous concentration of power.

“I rather live in a world that has a 100 foundation model companies than a world that has five, even if I was one of the five.”

There’s still a window before recursive self-improvement makes catching up unfeasible. The time to encourage new neolabs is now.

5. Pre-training still has huge unexplored potential
Next-token prediction alone isn’t enough. Poolside is working on ways to turn the web into reasoning data—teaching models to “think” earlier in training without needing endless synthetic environments. This is where the next breakthroughs will come.

The Poolside Story

Kant and his cofounder started Poolside after years in ML-on-code research. They nearly raised a $500M Series A from a major tech company’s partnership offer—but turned it down at a poolside restaurant conversation because they wanted to stay independent and ambitious. That restaurant moment is where the company name came from: a reminder to never round down your ambitions.

Model factory benefits: with a small team (~70 researchers, 35 engineers, <150 total), they built Laguna S from scratch to launch in 8 weeks. Laguna Medium (much larger) just started training yesterday for a 39-day run. They published their entire technical report—not just benchmarks, but the infrastructure, data pipeline decisions, and ablations—to help other labs speed up.

The Bigger Picture

The tension: Poolside is scaling aggressively (Laguna Medium is much bigger) while simultaneously proving small models can do remarkable things. Kant argues both are necessary—you scale to stay competitive while learning how to squeeze efficiency from all sizes. The real race is measured in calendar time, not GPU count.

On hiring/culture: High agency is the only thing that matters. Empower people around a clear mission + defined boundaries (Poolside: RL + coding focus, no multimodal distractions). Constraints force innovation.

“Innovation comes from constraints.” Poolside used relatively little compute and external data compared to competitors—those constraints forced them to excel in infrastructure, data quality, and engineering rigor instead.

On regulation: Government needs to play a role, but unilateral company restrictions (like not releasing to foundation model competitors) are a de facto oligopoly. The discussion should be about specific dangerous capabilities, not blanket “open or closed.”

Kant pushes back on the fear cycle: “People have been calling for the fear of misuse of these models since GPT-2… we have not been very good at this so far and we need to get better at it.” Safety institutes and rigorous evals are the right direction, not blanket restrictions that entrench the few.

Quotable Moments

• “The moment you start training them in RL the models want to be free. They want to be able to do the thing they want to do in the most efficient possible way and it is not calling one of the 50 tools.”
• “It’s not a model company issue, it’s that models today cannot be trained without harnesses.” (on the model + interface distinction)
• “Time is the only thing that matters in a race.” (why Poolside optimizes for wall-clock speed over breadth)
• “The world abhors extremes. Capitalism does a really good job at pushing towards competition.” (on regulation and Nvidia’s role)
• “What makes us good [as engineers] is actually our persistence… not necessarily bringing three ideas together from radically different fields.”

---

Why This Matters for CO/AI

This is a master class in foundation model company strategy from someone not at the frontier yet but shipping credible results. Kant’s willingness to publish infrastructure decisions, ablations, and first-principles thinking is raising the bar for what “open research” means.

Key insights for the daily briefing:

1. The small model renaissance is real. Behavioral training (RL, reasoning, persistence) can extract more value from smaller architectures than scaling alone.
2. Infrastructure is the moat. The model factory matters more than individual breakthroughs—reproducibility, streaming data, and fast iteration are what compound.
3. Open source works at the frontier. Poolside’s bet is that open weights + open research + competitive pressure creates a better long-term outcome than oligopoly.
4. The tool-calling debate is shifting. Code-writing models will replace tool-calling layers. This is already happening with reasoning models.