Clinton Mills

Opus 4.6 >>>>>>>> GPT 5.4

We've entered into an agreement to join OpenAI as part of the Codex team. I'm incredibly proud of the work we've done so far, incredibly grateful to everyone that's supported us, and incredibly excited to keep building tools that make programming feel different.

Geoffrey Hinton just dismantled the bureaucratic obsession with perfect algorithmic transparency. The enterprise world is paralyzed because it can’t read the algorithm’s mind. That paralysis is the competitive death sentence. Hinton: “In a big neural net, I don’t think it’s ever gonna be possible to prove things about what it will do. It’s not like lines of code where you can prove things. You’ve got lines of code for doing learning, but once it’s learned, it’s just a big set of weights.” The traditional system wants to treat a neural network like a standard software update. Demanding line-by-line proof of exactly what the machine will do before they’ll touch it. But when you transition from hard-coded software to a massive neural architecture, you surrender the ability to read the code. You’re no longer auditing a program. You’re interacting with an alien cognitive entity that learned its own logic from scratch. If you refuse to deploy until you can perfectly map its internal reasoning, you’ve already forfeited the board to adversaries who are perfectly comfortable operating without that map. Hinton: “If you ask, ‘Why do you get into a taxi? Why aren’t you scared getting into a taxi?’ The answer is it’s not because I understand how the taxi driver’s brain works, and it’s not ‘cause I have guarantees on what the taxi driver will do. It’s because I have a lot of statistical information that people have used taxis a lot and very few of them have died.” The regulatory class is demanding a 100 percent mathematical guarantee of safety before they’ll allow the compute engine to scale. Absolute guarantees don’t exist in the physical universe. There is only statistical confidence. We don’t demand a complete cognitive map of every biological operator we trust with our lives. We verify the statistics. We assess the incentives. And we move. That is the geopolitical reality of this moment. There is no mathematical guarantee that autonomous superintelligence won’t make a mistake. But if the United States halts deployment to search for an impossible proof of safety, adversarial regimes will accelerate their own black-box models and capture the century while we’re still auditing ours. You don’t win by demanding a guarantee. You win by running the most rigorous safety testing on the planet and deploying the system the microsecond the statistics tip in your favor. Hinton: “I think the best we can do in having safe AI is having good safety tests that give good statistics.” We are entering an economy where the most complex problems on Earth are solved by systems we fundamentally cannot reverse-engineer. Medical diagnostics. Global logistics. Drug discovery. All executed by massive sets of weights that are structurally inexplicable to the human mind. You don’t need to understand the physics of a taxi driver’s brain to get to your destination. You verify the outcome. And you get in the car. The ones who waste the next decade trying to unpack the black box will still be auditing when the ones who accepted uncertainty own the entire board.