Old wiring → new wiring

Old wiring: Pick a side. New wiring: Keep both. Each one is telling you something the other can’t.

The story

In December 2022, I was at a Buca di Beppo in Chicagoland, at a big round table with my family and several hubcap-sized plates of pasta.

ChatGPT had been out for about a month. I’d spent most of that month inside it, and I decided dinner was the right moment to explain what it meant for everyone’s careers and for the college funds they were saving into.

It did not go well. The reaction was a mix of “who cares?”, “huh?” and “what’s he going on about now?” To be fair, early adopters are annoying, and I was competing with spaghetti.

I think about that table a lot, because four years later the same conversation goes very differently. Nobody says “who cares” anymore. Everybody has an opinion about AI now. Most of us have two, and they don’t agree with each other.

Under the hood: the jagged frontier

In 2023, researchers from Harvard, Wharton and MIT ran an experiment on 758 consultants at Boston Consulting Group. Ethan Mollick, whose writing on AI I’d hand to anyone, was one of them. Some consultants got GPT-4 and some didn’t, and then everybody did real consulting work.

On eighteen tasks the researchers judged to be within the AI’s abilities, the consultants using it finished 12.2% more of them, did it 25.1% faster, and turned in better work.

Then the researchers slipped in a task that looked about as hard as the others but sat just outside what the model could do. On that one, the consultants with AI were 19 percentage points less likely to get the right answer than the ones working alone.

These were the same people using the same tool, and it made them better and worse depending on which side of an invisible line the task fell.

One fair caveat: that was GPT-4, in 2023. Today’s models are far better and the edge has moved a long way out. It is still an edge, and it still zigzags.

The rewire

The old wiring says pick a side, and most of what you read about AI is written to help you do it.

This week: draw your own frontier

Fifteen minutes today, then two short lists for a week.

  1. List ten things you did last week, at work or at home. Use real tasks, like “wrote the board update” or “planned the trip.”
  2. Sort them into three columns. Inside: I gave it to AI, checked the result, and it held up. Outside: I tried it, checked, and it failed. Never checked: I either avoid AI for this or use it and don’t verify.
  3. Look at the third column. It’s usually the longest. It’s also where the contempt and the dependence both live, because neither one has been tested.
  4. Test three of them properly this week. When you get an answer you like, paste this: Give me the strongest case against what I’m hoping to hear. If it can’t make one, you’re talking to a mirror.
  5. Keep two lists as you go. On the first, write where AI felt like real support. On the second, write where it felt slightly off, a little too smooth or a little untrue to who you are. The second list is the one I’d keep. It shows you what you were about to hand over without noticing.

Sources

EVERY SUNDAY

The Sunday essay.

One named idea a week on rewiring how you think, work and raise your kids now that the machines got smart.

Get the Sunday essay

EVERY SUNDAY

Keep evolving.

One named idea a week. Free, in your inbox.

Signup powered by Substack. Open the form if the embed is blocked.