The Sequence Opinion #892: The Anatomy of a Good Environment: When Verifiability is Not Enough
Substack Jesus Rodriguez
A new essay argues AI progress isn't just about verifiable answers—it's about how "grindable" a task is for training. That's why models nail math and code but flail at your inbox or a robot arm.
Based on reporting by Substack, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
There's a theory going around AI circles right now, sparked by a Dwarkesh Patel interview with math YouTuber Grant Sanderson, and it's the kind of idea that reorganizes a bunch of scattered observations into one clean pattern. The claim: verifiability alone doesn't explain why AI crushes chess, code, and math olympiad problems while stumbling through computer-use tasks, robotics, and everyday knowledge work. What matters is a second property Sanderson calls grindability, and the domains where AI looks miraculous are the ones that score high on both at once.
Verifiability is the property everyone already talks about. Can you cheaply and unambiguously check if an output is correct? Math proofs, unit tests, and game outcomes say yes almost instantly. A first draft of a marketing email or a robot's grip on a coffee mug says no, or at least not without a human in the loop, which is slow and expensive. But TheSequence's newsletter argues that's only step one. A domain also needs to be grindable, meaning you can generate an effectively endless stream of graded practice problems, cheaply, without a human curating each one, so a training pipeline can churn through millions of attempts and actually improve.
Stack those two axes together and the last three years of model progress stop looking mysterious. Chess and Go were solved early because you can simulate a near-infinite number of games and know instantly who won. Competitive coding and math proofs followed the same script once labs figured out how to auto-generate problems with verifiable answers. Your inbox, by contrast, is verifiable in a squishy, subjective sense at best, and it's nowhere near grindable, because there's no cheap way to spin up a million slightly different versions of
My take — AI-written commentary, not fact-checked reporting
Everyone's throwing money at RL-environment startups right now, and this framework is basically a warning label: verifiability was never the hard part, grindability was, and a lot of these billion-dollar environment shops are going to discover that generating realistic, gradeable, endlessly variable tasks for messy real-world work is way harder than writing a reward function for a math problem. I'd bet against most of them scaling past toy benchmarks before 2026, and I think the labs that quietly solve grindability for robotics or computer-use will look like geniuses in hindsight while everyone else was busy chasing verifiability theater.
Read more about this at: Substack