The world belongs to hackers

Most engineers and scientists I talk to are finally willing to say something that would have sounded crazy two years ago. Recursive self-improvement, with the models we already have, is enough for substantial growth from where we stand today. Not infinite growth, and not a free pass on everything. But enough that the interesting question is no longer whether the next frontier will arrive. The interesting question is which half of the stack you should spend your life on while it does.

I think there are two sectors. The first is models improving themselves, through reinforcement learning, fine-tuning, and even pre-training the next frontier from the last one. That work is real, and it is already happening. The second is what people call embodied or physical AI, though I mean something simpler (the surrounding constraints around the model, like compute, access to general machinery, and the ability to operate in the physical world through humanoids, lab equipment, or something like Anthropic’s Model Hardware Standard, which I find genuinely interesting, and we’ll talk more about later). If the model sector is the theoretical frontier, the second sector is the practical one. It is the engineering that turns a capable model into an outcome you can measure.

That latter half is where most of the world should be spending its time. We are lucky that great minds are on the model path. The rest of us live in the gap between a score on a leaderboard and something that happens in the world. The industry keeps inventing names for that gap (code-maxxing, token-maxxing, Meat-proxy, autonomous engineering). The name that sticks for me is outcome machines. An outcome machine tracks whether the agent took a reasonable action given the environment it was dropped into, not whether it sounded confident in a thread.

I founded Reason Machines to work on that. We want to improve outcomes, measure them, and get real value out of artificially generated code. But the first problem to tackle is the second sector I mentioned, the surrounding constraints, because that is where we are bounded today. We have always been bounded by bottlenecks. Here is the uncomfortable version. If the largest bound in scaling laws is compute, and you do not have the physical capacity to move compute, what are you left with?

You are left to hack.

I am starting to believe the world belongs to the hackers. Unless you can move the large things (clusters, models, fabs), the best move you have is to iterate by taking small things apart. Take a feature like cron and hand it to agents until every machine becomes a stateful worker. Take the accessibility tree on a MacBook (built for screen readers) and wire logic through it so computer-use agents can drive apps the way OpenAI’s Sky team did inside Codex, by reading a tree Apple already shipped for a different purpose instead of stealing the cursor. These are hacks in the precise sense. They use a product for something the creator did not hold first in mind. And they become extremely relevant the moment a bottleneck makes the intended path unavailable.

So what do you do when you are surrounded by bottlenecks? Pick the simplest quality-of-life improvement in front of you and go hard at it. Challenge the operating system, the platform, the network, in ways that have not been done before. Those moves bring new ideas, new culture, and often new products that later let everyone else do the same thing in a more standard way. What does it take to begin? Almost always a proof of concept that arrives as a hack.

The model will keep improving itself. That is not an excuse to sit and watch. The people who will matter in this next decade are the ones who treat every surrounding constraint as a surface they can open and rewire, one small scaling law at a time.