In project management, there is a famous concept called the “Iron Triangle.” You can have a project done Fast, Cheap, or Good—but you can only pick two.
As we build InfiniteUp, we have discovered that the exact same rule applies to Artificial Intelligence.
We are living in a gold rush where users expect magic. They want an AI that is Instant (under 200ms), Free (or extremely cheap), and Genius (terrific reasoning).
But as builders in the application layer, we are learning the hard truth: getting all three is a logistical nightmare.
The Trap of “Nano” Speed
Recently, we experimented with switching our task generation to a “Nano” model (a lightweight, super-fast LLM). On paper, it was a dream.
Speed: It slashed our processing time from 35 seconds down to 5 seconds.
Cost: It was a fraction of the price of the larger models.
It felt like we had cracked the code. We had Fast and Cheap. But then we looked at the Quality.
When we fed it the task “Plan my wedding,” the Nano model confidently tagged it as “Low Priority.”
To a human (or a smarter, slower AI), a wedding is obviously a high-stakes, high-priority life event. To a hyper-optimized Nano model trying to save milliseconds, it was just another item on a list. It lacked the nuance—the “terrific reasoning”—that users rely on.
This is the danger of optimizing for speed: you risk building a Ferrari that drives itself into a ditch.
The “Background” Illusion
So, how do we solve this? If we use a “Genius” model (like GPT-4o or Claude 3.5 Sonnet), we get perfect reasoning—it knows a wedding is a big deal. But it takes 30+ seconds to “think,” and the cost per token creates a burn rate that kills the business model.
The solution isn’t better AI. It’s better Psychology.
We are shifting our approach to Background Processing.
The Illusion of Instant: When you create a task in InfiniteUp, the app feels instant. You hit “Save,” and the task appears. You are back in your flow immediately.
The Invisible Work: Behind the scenes, we have “all the time in the world.” While you are scrolling or moving to your next task, our server is waking up a smarter, slightly slower model. It takes its time to analyze “Plan my wedding,” correctly identifies it as High Priority, breaks it down into subtasks, and quietly updates the record.
By decoupling the Input (Fast) from the Intelligence (Smart), we can afford to pay for the reasoning (Cost) without breaking the user’s flow.
The “Good Enough” Future
The future of productivity apps isn’t just about who has the smartest model. It’s about who creates the best magic trick.
We don’t need “Nano” to be a genius; we just need it to be fast enough to catch the ball. We don’t need the “Genius” model to be instant; we just need it to be invisible.
The goal is to stop making the user wait for the machine to think, and start designing systems where the thinking happens in the walls. That is how we break the Iron Triangle.