The OpenAI usage debacle
Competing fixes for AI limits, our Decisions walkthrough, and what personal projects teach us.
The models keep getting more capable. Planning a day’s work around the usage meter is getting ridiculous.
Better tools. More time watching the meter. · AI-generated illustration
On Monday’s Runpoint call, one engineer had burned through two usage resets in a week. Several of us were hitting limits fast enough to disrupt work. We were debating which model should supervise which other model just to keep going.
That is a lousy tax on tools that are supposed to make us more productive.
One idea is to let an expensive model plan and a cheaper one do the work. But the advice points in different directions:
-
Put Astra in charge of Luna. Melvin Vivas uses the stronger model to direct cheaper workers.
-
Keep the work in separate conversations. Kevin Kern has Astra and Sol report back when they finish.
-
Skip the extra agents. Ahmed argues that Astra can handle most tasks itself and extra coordination can waste usage.
We’re experimenting too. My first attempt wasn’t a clean enough test to tell us much. Model routing is still an open problem. There’s no settled recipe I can recommend.
I want to compare useful work finished, supervision required, and the cost of the whole job. A cheaper answer means little if somebody spends the afternoon fixing it.
The human part of building software
Send a question to fifteen people in Slack and you can still end up with no answer. Everyone assumes someone else will respond.
Ryan Mish built Runpoint Decisions around that problem. Put an interactive mockup in front of the people who will use it, name the approvers, and ask for feedback within 24 to 48 hours.
Reviewers can pin comments to the screen, see each other’s questions, and approve, request changes, or decline. Sales, finance, and leadership get to react to the same thing while the work is still taking shape.
In the new video, Matthew and Ryan walk through the tool using a fictional CRM project. The aim is to make the human part of building software move with the rest of the work.
What we’re building on the side
We make time for personal projects on our Monday calls. They’re a great way to learn, and sharing them lets the rest of the team learn from the useful parts and the ridiculous ones.
The shopping app stays. The instructions change. · AI-generated illustration
A grocery app you can talk to. One engineer is connecting an AI assistant to his wife’s shopping app so she can add recipes and plan groceries while doing laundry. She keeps the app she already uses; the project changes how she gives it instructions.
Another look at my wardrobe app. I’d built an app that suggests outfits from photos of my clothes, found it useful, then stopped using it. The new models gave me a reason to revisit it. I’m wearing combinations I wouldn’t have picked myself.
It also suggested tight jeans with boots over them.
“I hated that one.” · AI-generated illustration
At least the test is clear: would I actually wear it?
What are you building or testing? Email Sam.