← The Brief
Nº 45 ·
What I'm telling portfolio CTOs this month
A short memo on AI roadmapping, pulled from three conversations this week.
Stop planning for GPT‑5
You cannot roadmap to a model that does not exist yet. Roadmap to the product experience you want in Q3, then take the shortest path to it with the model that exists today. When the next model ships, you will get a tailwind, not a re‑plan.
The three‑tier stack
Your AI stack should have three tiers, not one. A frontier model for the 5% of hard calls, a medium model for the 80% of routine calls, and a heuristic or small model for the 15% you should never have been paying an LLM for. Most teams I talk to are paying frontier prices for the third tier.
Hire the eval person first
If you can only make one AI hire this quarter, make it the person who owns evaluation. A good eval lead will unblock every other AI hire you make. A bad one will make every hire you make look confused.
One thing I'm building
Hiveclaw's approval gates landed in production this week. Seven agents, four tiers of human review, one Stripe account that has never accidentally been charged. It is the most boring system I have ever been proud of.
Reading list
Three papers, one essay, one old O'Reilly book on distributed systems. Reply if you want the list.
Get the next one
One email a week. No noise, easy unsubscribe.