The Boring Future of AI: Why I Stopped Predicting AGI

Three years ago we kept an AGI countdown on a sticky note in the office. December 2027. We were confident enough to wager dinner over it. We're not anymore. The reason isn't that the models got worse.

The reason is that the most useful AI work we've seen in the past eighteen months has nothing to do with AGI. It's a logistics company replacing a 4-step Zapier flow with a single Claude call. It's a hospital using a small fine-tuned model to triage incoming clinician notes. It's a six-person SaaS company shipping a feature that would have needed two engineers a year ago.

The predictions we got wrong

Going back to our 2023 notebook, here's the scoreboard:

• "Most knowledge work will be automated by 2026." Wrong. Most knowledge work is still done by humans, sometimes faster, occasionally annoyed.
• "Multi-agent systems will eat single-prompt apps." Wrong. The single-prompt app, when the prompt is good, still beats elaborate orchestration in 80% of the use cases we see.
• "Enterprise will lead consumer." Half-right. Enterprise procurement is faster than we expected; enterprise deployment is slower.

What did we get right? Almost nothing about the technology. A few things about the workflow: that prompts would become a craft, that retrieval would matter more than parameter count, that latency would matter more than capability for actual users.

The boring view

The boring view is that AI is going to look like every other useful technology. Slow integration, weird path-dependencies, more value captured by the patient than the visionary. Nobody's going to wake up to AGI. They're going to wake up to a calendar that books their meetings, a CRM that drafts their replies, a legal tool that flags weird clauses. At some point, they'll realize they haven't done that work themselves in months.

That's not a disappointing future. It's a normal one. Electricity didn't have a launch day either.

What it means for people building today

If you're shipping AI products in 2026, our advice is small and unglamorous:

1. Pick narrow problems. The wider you go, the more competition you face from the next OpenAI release. The narrower you go, the more your domain knowledge compounds.
2. Build for the boring case. The 90th-percentile request is the one you should obsess over, not the demo.
3. Treat the model as the cheapest part of your product. Because it is. Distribution, trust, integration, support: that's where the work actually lives.

We're done predicting AGI. We've replaced the sticky note with a Notion doc titled "wrong predictions." It gets longer every quarter. So far that's been the most accurate thing we've ever shipped.