Don’t Hire a Professor to Change the Font Size
AI models are increasingly less like pieces of software and more like workers with different levels of ability and different costs. Once several models and reasoning levels are available, always choosing the most powerful one makes about as much sense as hiring a professor to change the font size on a website.
A useful mental model is to treat model choice as the seniority of the person you hire, and reasoning effort as how much time you give that person to think.
| Model | Low | Medium | High |
|---|---|---|---|
| Luna | Intern | Smart Intern | Obsessive Intern |
| Terra | Junior Engineer | Grad Student | PhD Student |
| Sol | Senior Engineer | Professor | Professor in Research Mode |
| Astra | Expert Consultant | Nobel Laureate | Nobel Laureate With a Month to Think |
The two dimensions are different. Moving from Terra to Sol is like hiring someone more capable. Moving from Medium to High reasoning is more like telling the same person to spend the afternoon thinking instead of giving you an answer in twenty minutes.
That means Terra High and Sol Low aren’t simply “better” or “worse” than each other. A capable grad student given several hours may outperform a professor taking a quick look at some problems. On other problems, expertise matters more than additional thinking time.
For practical purposes, the entire matrix can be compressed into four employees:
| Persona | Configuration | Give them |
|---|---|---|
| The Intern | Luna Low | Mechanical, tightly specified changes |
| The Grad Student | Terra Medium | Normal implementation and everyday work |
| The Professor | Sol Medium | Architecture, difficult debugging, ambiguous problems |
| The Nobel Laureate | Astra Medium | Novel or unusually difficult problems |
My default is the Grad Student. Most work simply doesn’t require a professor.
If I already know what needs to be done and need someone to implement it, that’s grad-student work. If it’s completely mechanical, give it to the intern. If the grad student gets stuck because the problem itself requires better judgment, bring in the professor. The Nobel laureate should be sitting around doing nothing most of the time.
Reasoning level provides a second control:
| Reasoning | Interpretation |
|---|---|
| Low | Take a quick look and do it |
| Medium | Think about it carefully |
| High | Go into your office and don’t come back until you’ve figured it out |
This suggests a simple escalation strategy:
Intern -> Grad Student -> Professor -> Nobel Laureate
Start with the cheapest level of intelligence that can reliably perform the task. Escalate when the nature of the problem requires it, not simply because a more powerful model exists.
There is a broader idea hiding here. Once AI systems can choose among models, reasoning budgets, tools and other agents, allocating compute starts looking like organizational design and labor economics. Different kinds of cognitive labor have different requirements and different costs.
The interesting question stops being:
Which AI model is best?
and becomes:
What is the cheapest level of intelligence that can reliably perform this cognitive task?
That is a much more useful question, whether you’re choosing a model manually or designing an economy of autonomous agents.