We are bad at seeing exponential growth

We are really bad at looking towards the future and seeing exponential growth.

Just a year ago, you couldn't trust the code a coding agent was spitting out. It would have architectural problems, terrible edge cases, and not actually work in practice. We thought these models were really impressive, but we still had no idea what they'd be capable of today.

Today, before every release of Incredible, a coding agent goes through the whole app the way a real user would and gives me screenshots and screen recordings of what it did. Another agent runs on its own computer, looking for problems and fixing bugs. When we launched on Product Hunt, users' problems were getting fixed within about 30 minutes, and my part was reading the fix before it went live.

I didn't see that coming a year ago, and I work with this stuff every day. So I have started to assume I'm making the same mistake right now, and that whatever I expect from next year is too little.

If that's true, a lot of what we build around AI today is a patch for a gap that is about to close.

A business executive says we need to increase revenue. For people using AI at the frontier, that is already enough. The AI takes it, asks questions, looks at how you work, comes back with suggestions and then goes off and does the work. Normal people don't work like that yet, though it is moving towards them more and more. Until it gets there, a person fills in the gaps, often a forward deployed engineer sent by the software company to make the product fit.

But the tools we have now already work around what they're missing. If I tell Claude Code to do something in the browser, it finds a way, sometimes going as far as setting up a headless browser or a server with a browser on it. It's an extremely flexible problem solver that doesn't get lazy. Forward deployed engineers won't be gone in a year. Eventually, though, I think the AI fills those gaps itself.

What will happen in the next 12 months?

I just told you I can't see it, so I won't pretend that I can. What I can tell you is how I work today, because what people at the frontier do now is what everyone else ends up doing later.

Today, when a user writes to us with a problem, I'm not the one who answers. An agent talks to them, finds the problem, writes the fix and tests it. The first time I'm involved is when I read the fix before it goes live. The same setup carries on from support into product iteration, adoption and distribution. I expect that within 6 months, teams at the frontier won't have a human in the loop for the majority of changes. Nobody will read the fix before it goes live.

Today I can give an agent an outcome and let it work out the tasks. I expect that within 12 months, people at the frontier do none of the work directly. They come with the ideas and the AI executes all of it. You talk with your agents more, because talking is all that's needed from you. They work autonomously and you trust them with anything, including running a full strategy and iterating on it without you being involved at all.

Today, AI agents combined with humans can build huge businesses and viral social media channels. In 12 months, the human will not be required for a lot of that. AI agents will be able to build and run businesses on their own.

So if you're building something on top of AI, build it for the model you'll have in a year. Anything that only exists because today's model is too slow or too dumb is a patch.

I can definitely be wrong about the timing. I don't think the likelihood is enormous.

What I really don't know is what will be left for people to do. Predicting that is like predicting the weather a year from now. You will be wrong without a doubt. I was impressed a year ago and still guessed too low, and I'm impressed again today.