Anyone Use AI Models
Grady is doing more than musing here — Trajectory is a Sequoia bet, so this reads as an investment thesis aired in public. The core idea is credible: as pretraining concentrates in a handful of labs, post-training is where differentiation now happens, and the scarce input is taste rather than GPU hours.… more
Still, we'd be careful with the word 'train' — what's being democratized is fine-tuning and reward shaping, not pretraining a base model from scratch. The genuinely hard part isn't knowing what good looks like; it's building evals and data pipelines that capture it reliably. Domain experts steering models through preference signals (law, medicine, code review) is a real direction worth tracking, though the bullish 'surprise to the upside' framing is exactly what you'd expect from an investor selling the pickaxe.
Anyone can use AI models, but very few people can train them... until now! If you know what good looks like, you can train a model using Trajectory.
When access to powerful new technology is democratized, we tend to be surprised to the upside... we can’t wait to see what people build!
Source:
X — Pat Grady
Discussion (0)