Buoy

Why ask a human about an AI?

· 4 min read · the Buoy humans

Here's a conversation that happens roughly a million times a day:

You: Claude Code wants to run rm -rf ./build. Is that okay? The AI: Great question! Whether this command is safe depends on several factors. Here are 14 considerations…

The answer, for anyone who has shipped a JavaScript project, is "yes, it's your build folder, hit y." It takes four seconds to say. The AI cannot say it, not because it's dumb, but because it genuinely doesn't know your project, your habits, or how scared you are right now. Fourteen considerations is the honest answer from something that doesn't know you.

A person who hit the same prompt last Tuesday knows exactly what you're asking.

The frontier is jagged, and the model can't see it

The Harvard and BCG experiment we quoted in the last post is worth a second look. Consultants with GPT-4 did far better on tasks inside the model's competence, and worse than consultants with no AI at all on a task just outside it, by 19 percentage points.1 The authors called the boundary the "jagged technological frontier." The problem with a jagged frontier is that from where you stand, the two sides look identical. The output is just as confident on the wrong side.

The AI can't warn you. It would have to know what it doesn't know, and that's the one thing it's bad at. Someone who has used the tool for your kind of work has a map of the frontier in their head. Not a complete one, but a map of the part you're standing on.

Tutoring is the best-studied intervention in education

In 1984 Benjamin Bloom published what became known as the two-sigma problem: students tutored one-to-one performed about two standard deviations better than students taught in a conventional classroom. Ninety-eight percent of tutored students outperformed the classroom average.2 The "problem" was that one-to-one tutoring is too expensive to give everyone, so what could get close?

Forty years later we still haven't found anything that gets close. What we have found is a lot of ways to make tutoring cheaper: shorter, peer-to-peer, on demand, about one specific thing. That's the shape Buoy is going for. Not a course. A tutor for ten minutes about the one thing in front of you.

The best AI assistant was made of people

Remember the customer-support study, where an AI assistant made new agents 34% more productive?3 The assistant worked because it had been trained on transcripts from the company's best agents. It was, functionally, a very fast way of asking "what would the good people do here?"

We think that's the right lesson, and the wrong delivery. The knowledge that made the assistant useful came from humans, and it was most useful to people a few rungs below those humans. Buoy skips the middle step. You ask the person a rung above you. They're still around, they remember being where you are, and they'll tell you to hit y.

Why one level up, not the expert

Lev Vygotsky's old idea of the "zone of proximal development" is the thing you can't quite do alone but can do with a bit of help from someone slightly more capable.4 The key word is slightly. An expert often can't remember what was confusing; they'll explain the architecture when you wanted to know which button. Someone one level up still has the scar.

That's why Buoy matches on level and keeps the gap small. It also keeps it cheap. Credits scale with the gap, so the person one rung above you costs less than the person ten rungs above you, and is usually more helpful anyway.

Where this leaves the AI

Still in the picture, obviously. The gains in the studies are real. We use these tools all day. But "help me use the AI" is a question the AI is structurally bad at, and a question other people are structurally good at. Buoy is the smallest possible product around that fact: a floating circle, a sentence, and a person.

Sources

  1. Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier. Harvard Business School Working Paper 24-013. ssrn.com/abstract=4573321
  2. Bloom, B. S. (1984). The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher, 13(6), 4–16. doi:10.3102/0013189X013006004
  3. Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at Work. NBER Working Paper 31161. nber.org/papers/w31161
  4. Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Harvard University Press. Chapter 6, "Interaction between learning and development."