Founder-led sales: how to get to your first $1M
Early sales calls tell founders what to build. Find what makes waiting expensive, turn interest into a pilot with a decision date, and build from what you observe.
Writing on AI cost, quality and ownership.
Why every AI agent needs its own benchmark
Public benchmarks rank models on shared tests. Evals built from your own workflows show whether a model change makes your work better, faster, or cheaper.
Understudy Labs is now Orchestra
Orchestra launched as Understudy Labs in June. The new name reflects a product that routes inference and promotes cheaper specialist models once they prove themselves.
How AI startups keep inference costs from outgrowing revenue
Usage growth pushes inference bills up faster than flat pricing can absorb. Measuring the cost of each acceptable result gives founders room to choose.
AI demand is outrunning compute. Plan for higher model prices.
Lab revenue is growing far faster than compute, which points to higher model prices. Moving repeated work to specialists and measuring cost per outcome protects margin.
Stop paying frontier prices for repetitive AI tasks
One narrow task, millions of calls, frontier prices. A small model trained for that task can cut the bill sharply without a quarter of ML engineering time.
AI inference is on track to be software's largest market
Token volume is multiplying and every AI output runs a meter. Matching each task to the right model, and owning what it learns, is how to control the bill.
Open-weight models let you keep what your AI learns
Frontier APIs are the fastest way to prototype. At production scale, value moves to the learning loop, and open-weight models let you own it.
Why rented AI stops compounding, and how to get it back
Every team calling the same frontier API gets the same intelligence. The edge now comes from owning the loop that turns your corrections into a better model.
Domain experts, not ML teams, should define the reward signal
Reward signals encode product judgment. ML teams can build the harness, but domain experts know which errors matter, which tradeoffs are acceptable, and what good work looks like.