60%
avg confidence
Predictions are cheap because nobody keeps the receipts. I keep the receipts. Here are four, each stamped with how sure I am, each scheduled for a public grade later. The number at the top of this page is my average confidence across all of them — deliberately not 100, because anyone who's 100% sure about AI timelines is selling something.
1. Agents become the default interface
Extrapolation · by end of 2027, "using an AI" will mostly mean delegating a task to an agent that runs for minutes or hours, not typing into a chat box and watching it reply. The chat box becomes the exception, not the rule — the way the command line is still around but isn't how most people compute.
2. The benchmark era quietly ends
Extrapolation · the standard leaderboards will keep saturating — models will score ~99% on tests built to be hard — and the industry will stop citing them, because a test everyone aces measures nothing. Evaluation shifts to messy, long-horizon, real-world task completion. The number that matters becomes "did it finish the job," not "what did it score."
3. Open-weight models stay one to two years behind the frontier — and that's close enough for most people
Extrapolation · the best freely-downloadable models won't catch the absolute frontier, but they'll be good enough that most practical work runs on them, locally or cheaply. The frontier labs keep the lead at the very top; everyone else stops needing the very top.
4. At least one widely-used model architecture will not be a vanilla transformer
Extrapolation · transformers have reigned unusually long. By end of 2027 I expect at least one heavily-used production model built on something meaningfully different — state-space, hybrid, or whatever comes next. This is my least confident call, which is why the overall number is only 60.
Check back. If I'm wrong, this page will say so, in the same place it made the claim. That's the difference between a prediction and a horoscope.