The Threshold

GPT-5 Pro and Opus 4.1 are both out right now, and something crossed a line.

Not incrementally crossed. Not "meaningfully improved the developer experience" crossed. The other kind — where you describe a system, walk away, and come back to something that required genuine architectural judgment. Where the model is wrong in the way a competent person is wrong, not in the way a token sampler is wrong.

Software development is different. The actual thing, not the "generate some boilerplate and paste it into a file I already had open" thing.

The strange part — the part that keeps sitting with me — is that the intelligence was never the bottleneck we needed to clear next. We already cleared it. What's missing is speed, and speed is an engineering problem. Not a research one. We're not waiting for a new idea. We're waiting for the inference cluster to catch up to what these models already are.

We built the thing. It works. It's just slow.

Which means at some point in the not-particularly-distant future, all of this gets faster, and nothing else changes — no new capability, no architectural leap, just the same model but faster — and we enter a genuinely different world. The intelligence is already there, sitting in a data center somewhere, waiting for the hardware to stop being the excuse.