The Bet

How good of a teacher are we to AI? Are we teaching the AI models with the same amount of care and curiosity that our favorite teacher had for us? Or are we rushing to replicate humanity’s most common workflows as quickly as possible.

Novel capabilities emerge from carefully crafted teaching. We’re building the engine for identifying model deficiencies and producing just the right artifacts to improve them. Maybe it’s a knowledge gap in pre-training, perhaps a good chunk of expert traces, or a need for lots and lots of trial and error. The product will not be data, it will be incremental model capabilities. And incremental model capabilities will unlock higher levels of intelligence.

There is clear demand for AI training data, but the quality is nowhere near where it can be. Slop data hurts the models more than it helps. The current creators of AI data are orchestrators not thinkers and when creating tasks, thinking can not be outsourced. People who make the models need a data research partner who deeply cares about making the model more intelligent as much as they themselves do, if not more. The moat comes from quality & genuine curiosity: real experts, well-thought tasks, scientific measurement and well-documented methodologies.