As we move toward fully autonomous coding, the cost of every single line of code is dropping to zero. Consequently, the value of exploration is far greater than it used to be. It used to be the case that we needed to plan carefully and think deeply before implementing a new feature because the cost of getting it wrong was high; but now, we can throw away code, implement five versions of the same feature, deploy them all to staging, and test them. The value of coding is moving more and more toward evaluation rather than generation.

This creates an opportunity for what I call "evolutionary platforms," where your code evolves in generations. You have multiple versions of the same feature—or even multiple versions of the same application—competing and merging with each other to find the best solution to a problem. We should see coding as an optimization problem rather than a step-by-step construction, viewing the design space of your product as a landscape to explore in parallel with multiple agents, rather than something that must be thought through very thoroughly.

In this new world, the most important bottleneck we have right now is that models are very uncreative. Their distributions are highly collapsed. If you ask them the same question, or even slightly similar questions several times in different contexts, they will give you more or less the same answers. Improving this creativity is one of the next most important research lines.