Part of me is weirdly happy every time an LLM makes a mistake.
A year ago, I would have been frustrated but now I see it differently. Over the last 18 months, I’ve watched these tools change how we work faster than any software shift I’ve seen in my career. Even looking back six months, it is hard to believe how much our workflows at Codex have changed. For a non-technical product person, it has felt a little like being handed wizard powers.
Not perfect powers. Not safe to trust blindly. But real leverage. My role is not to sit back and admire the machine. My role is to set the vision, provide context, keep it on track, and make sure the result is actually right. Never trust, always verify still feels like the right mantra.
So yes, I’m happy to see a mistake now. I literally still have a job because these things make mistakes. And at the same time, when you spend enough time with them and really learn their strengths and limitations, it is hard not to feel some awe. It really does feel a bit like Asimov in real life.
It’s especially great to see open models keep improving. A lot of my own curiosity now lives there, evaluating models at home, figuring out which ones fit different jobs, and trying to find my way to sovereign intelligence.
Once intelligence is deeply embedded inside corporations and governments, the people who control the models will control far more than software. They will have their hands on the intelligence layer itself. And power at that layer does not always announce itself with some dramatic act. Sometimes it arrives quietly through weight changes. A tweak to the model. A shifted preference. A subtle update that ripples outward through systems more and more people depend on.
That is the part that feels dangerous.
The deeper AI gets wired into society, the less comfortable I am with that kind of power living inside a handful of corporations. This is where crypto starts to feel less like a separate obsession and more like a necessary counterweight. If intelligence is going to become this important, we need credible ways to distribute control, verify behavior, and resist capture.
The more I work with AI, the more convinced I become that we are going to need a cryptographic answer to centralized intelligence.
Maybe that is part of the strange comfort in today’s imperfect models. Their flaws are still buying us time. Time to build better counterweights. Time to sharpen my own judgment. Time to decide whether the operating system of the future belongs to everyone, or just the few people allowed to change the weights.