AI Progress Is Really a Factory Line

William Palmer

Everyone everyone wants to believe that- that AI progress is this romantic story of a lone, brilliant researcher sitting in a dark room, scribbling some beautiful new neural net architecture on a whiteboard. But, um, but the reality? It’s actually closer to a high-speed factory floor. It’s- it’s dirty, it’s loud, and it’s 90% pure industrial engineering. You look at what Eiso Kant and the team at Poolside just did with their new Laguna S 2.1 model—a massive 118-billion-parameter model, though only 8 billion are active at any one time because it’s a sparse mixture-of-experts—and they took that thing from the absolute start of training to a full public launch in under nine weeks. Under nine weeks! And they did it with a team of fewer than 70 researchers. That is not just a fast research cycle. That is a highly optimized, fully automated factory line. It’s- it’s like the SpaceX approach, but for weights and tokens.

William Palmer

As a guy who spends his weekends wrenching on race cars, this- this immediately clicks for me. If you’re running an elite racing team, you don’t win the championship by just hoping your driver is a genius on Sunday. You win because your telemetry is so fast, so incredibly precise, that your pit crew can analyze tire wear, fuel flow, and suspension geometry in real time and tweak the setup in seconds. Shaving off those fractions of a second is the entire game. And at Poolside, the single KPI they obsess over isn’t some abstract academic benchmark—it’s the speed of an idea. How fast can a researcher take a concept from their head, spin up an experiment, and get back a trusted, verifiable result they can actually build on? While the traditional, massive frontier labs are still stuck on these incredibly slow, bureaucratic six-month model training cycles, Poolside’s factory is pushing out between 10,000 and 20,000 experiments every single month. It’s- it’s relentless.

William Palmer

And the plumbing under the hood to make that happen is just beautiful engineering. Think about how most companies train these things. They spend weeks prepaying for, cleaning, tokenizing, and packaging up these massive static datasets, and then they copy them over to the training cluster before they can even press start. If they find a bug or want to tweak the data mix, they have to stop everything, rebuild the whole set, and re-distribute it. Poolside threw that entire playbook out. They built a system called Blender that streams raw data just-in-time directly into the active training jobs. It treats the entire data layer as completely immutable, meaning every single token, every cursor position, and every line of code is perfectly versioned. If a run from two years ago glitched out, a researcher can go back and recreate that exact training run down to the precise token. And because they treated this as a rigorous distributed systems problem from day one, they had zero midnight on-call incidents during the entire run of Laguna S. No pagers going off, no researchers waking up at 3:00 AM to restart a crashed cluster. It just hummed.

William Palmer

But the engineering philosophy doesn’t stop at training. It actually changes how these models behave when you set them free as agents. Right now, the entire developer community is- is going completely crazy over Anthropic’s Model Context Protocol, or MCP, and this trend of stuffing system prompts with 30 or 40 rigid JSON schemas so the model can call external tools. Eiso Kant’s take on this? He says it’s stupid. Like, flat-out inefficient. When you’re trying to solve a complex, long-horizon task, forcing a model to select from a massive, bloated menu of predefined API endpoints just wastes precious context windows and creates fragile integration points. It’s an anti-pattern.

William Palmer

Instead, Poolside designed a ridiculously minimal agent harness with just six basic tools: shell, kill, shell wait, write, fetch web, and bash. That’s it. You don’t give the model a custom tool for every little database query or file manipulation. You give it a secure, isolated container with standard binaries, and you let the model write its own executable scripts on the fly. If it needs to loop through a directory, filter some files, and verify the output, it doesn’t make ten separate tool calls—it writes a python script with loops and conditionals, runs it in the shell, and looks at the stdout. The model is free to solve the problem its own way, like a real developer.

William Palmer

And this is where the real magic of Laguna S 2.1 comes in. Because it has this incredibly lightweight harness, it can allocate its compute where it actually matters: thinking. They built this feature called "max thinking mode" which lets the model dynamically scale its test-time compute budget, sometimes consuming up to 249,000 completion tokens for a single complex task. Instead of just guessing the next token and hoping for the best, the model uses that budget to write code, run it, see the error, backtrack, fix its assumptions, and persist until the job is actually done. It- it swaps raw, bloated parameter size for behavioral persistence. And to prove they aren’t just reward-hacking the standard benchmarks, they did something incredibly brave. They launched trajectories.poolside.ai, where they publish the unedited, raw execution traces of their final evaluation runs. You can go online right now and see every single shell command, every python script, and every messy reasoning step their model took. That is the kind of weaponized transparency this industry desperately needs, and honestly? It’s a masterclass in how to build the future without the hype.