The AI Change Wizards and Pundits Miss: A View from Outside the Bubble
The AI Change Wizards and Pundits Miss: A View from Outside the Bubble
Three Takes on AI | Brian Silverman, Campbell Robertson, Michael Muhlfelder
A recent New York Times conversation between Tyler Cowen and Dwarkesh Patel drew a lot of attention. Patel, who hosts one of the most followed AI podcasts, said he expects drop-in knowledge workers within two to three years and AI working without humans around 2030. He also argued that any frontier vendor using a model internally should have to make it public.
On this episode of Three Takes on AI, we ask what that looks like from outside San Francisco, where businesses make decisions with real budgets and real customers.
Where the interview lands, and where we push back
Patel and Cowen spent much of the hour disagreeing about dates. Yet they agree that today’s models are not AGI, so the person sounding the loudest alarm and the person urging patience are looking at the same thing. Forecasts are fiction. What you can verify in your own business is not.
We also noticed that the leaders of the major AI vendors can’t agree on where we are. Some say everyone is being alarmist, and others say we need to slow down. That disagreement deserves more attention than any date on a keynote slide.
The view from the sidelines vs. the view from the street
Every technology wave, from mainframe to client-server to the internet, was billed as the one that changes everything. Each did change a lot, but the fundamentals of how people decide to buy things moved far less than predicted. Experience running things through those cycles matters when you give advice, and advice from people who haven’t operated anything should come with an asterisk.
AI is also bigger than frontier language models. Years of AI work in areas like defense and drones suggest the story isn’t limited to generative AI from a handful of vendors. Business value will come from companies applying it to real processes, probably from somewhere in the middle rather than from the AI companies themselves. As with the internet, the standards and the businesses that learned to use them changed commerce more than the browser did.
Should internal models be made public?
We weren’t convinced. It’s hard to see how it would work in practice, and it could turn into a contest over whose model is better rather than a real risk conversation.
What to do instead: hold the forecast loosely, spend on controls
Put a 90-day AI capability review on the calendar and give it an owner.
For each AI tool in use, write two sentences: one on what it does reliably today and one on where it fails.
Tie policy changes to that evidence, not to keynote predictions.
Don’t take accuracy scores at face value. They’re a claim made by the seller. Check precision and recall against your own work.
Sample AI outputs monthly and have a qualified person review them by hand.
Add contract terms requiring notice of model changes, version identifiers and logs.
Track gaps between what the vendor reports and what your samples show, and escalate the wide ones.
A score is something the system produces. An audit is something you do to the system, and you control it.
Referenced in this episode




Comments