Ep 184: Only 10% of Users Touch the Model Picker - Nobody Cares Which AI You Use

On this episode I’m talking to Olivier Legris, co-founder of Alter, about building an AI product on the bet that the model underneath it doesn’t matter to anyone but the four people building it.

Olivier found his co-founder the way most good partnerships apparently start now: scrolling Reddit. He tried a beta product called Outer, tracked down the small French team behind it, got on a call meant to last half an hour and stayed on for two. By the end they’d agreed to work together. Alter is what came out of it, a Mac-native cowork alternative that plugs into more than 70 models, including the ones you’d expect and plenty you wouldn’t.

The assumption they started with turned out to be wrong, or at least beside the point. Alter built a router called Best, designed to pick the optimal model for each task on the user’s behalf. The expectation was that people would want visibility and control over that choice. What they found was the opposite: only 10 percent of users ever touch the model picker. Nobody wants to know what’s running underneath. They want the task done, done well, and done for a sensible price. Olivier’s comparison lands well: nobody asks what chip is in their laptop either.

We get into the pace problem this creates for a small team. Alter is four people, fully bootstrapped, shipping every Friday. The hardest part of the job isn’t technical, it’s deciding whether a problem is worth solving now or will simply be solved for you by the next model release. Spend too long on the wrong thing and a frontier lab makes your work redundant in a month. Olivier’s answer is a fixed long-term vision paired with total agility on the week-to-week roadmap, which is a more honest description of how most AI-native teams are actually operating than anyone tends to admit publicly.

Then the frontier labs conversation, where Olivier draws his own clear line. He doesn’t think most people care who wins between Anthropic, OpenAI or anyone else, in the same way most people never cared about Intel versus AMD. What they’ll increasingly care about is consistency, cost, and whether the output can be trusted. His prediction on IPO timing for the big labs is worth hearing in full, but the short version is that he thinks it will be a call made on market momentum rather than fundamentals, and that the leaks will start the moment the paperwork does.

If you’re trying to work out how much attention your own team should be paying to model choice versus the actual problem you’re solving, this conversation is worth your time.