Building while AI keeps changing

Choosing the right model for each task was something we were stuck on. We started using Jev to choose among the options we supply and changed a lot of our implementation around it. What stands out in our use is how quickly Jev makes these small choices. A capability outside our own roadmap changed what we needed to build.

That speed makes me reconsider work that needs many decisions in sequence. An agent routing a task or choosing its next action on a computer can spend time waiting at every step. If those choices become fast enough, a team could attempt a larger job than it would have considered with slower decisions.

I wonder how much of a roadmap exists because of limitations we had to work around. Some choices protect the objective. Others preserve an implementation that made sense when we planned it. And some of our ambitions never reached the roadmap because they required capabilities or people we didn't have.

Scott Wu describes a value at Cognition as "when faced with a tradeoff, pick the ambition-maximizing direction". I like that. Reconsidering implementation should also make us reconsider how much we could attempt.

Keep the objective, test the constraint

The Jev change gives me a useful starting point for the next decision. I'd take the task behind a planned workaround and try the new capability on it before continuing that work. For model selection, I'd check its choices on representative tasks and inspect the results from the selected models. A fast choice is useful only if the model it selects does the job.

I'd want to know which choices need correction, what happens when the supplied options are inadequate, and whether dealing with those cases leaves us better off. They would tell me whether to expand its responsibility or keep a narrower use.

I'd still commit to delivering something and set a date to review what we learned from people using it. A new capability can change how we deliver that work without restarting the project. Evidence that the chosen approach no longer serves the customer could justify changing it earlier.

Reopen the larger idea

Meta offers a larger example of learning while pursuing an existing ambition. Alexandr Wang shared a document he said he wrote for Meta's board in September 2025, describing personal AI working toward someone's goals in the background. Nat Friedman later said using OpenClaw in January inspired Muse as a product, although Muse was built from scratch. An existing direction left room to learn from someone else's implementation.

In a 2025 interview, Kevin Weil described quarterly roadmaps and a year-long strategy at OpenAI, with new learning changing what shipped even within the quarter. Planning still helped coordinate dependencies and review what worked.

The difficult case is when the current project is useful and a larger idea starts to look feasible too. I'd test the constraint that originally ruled out the larger idea, then compare what customers could accomplish with each. I'd include the work we'd delay and how close we are to putting either option in someone's hands. A promising demo would leave too much of that choice unanswered.

I want to go back to the plans we called unrealistic six months ago. Some depended on capabilities we didn't have, or required more people than we could afford. Before we spend another quarter improving what we already know how to build, I'd like to see which of those constraints still exist.