A wise man once told me you can tell what's going on in a programmer's life by reading their code.
Not the comments. Not the commit messages. The code itself.
Someone wrestling with a decision they can't make? You'll find it in excess looping — logic that circles back on itself, unwilling to commit to a path. Someone running on a "just get it done" mandate? You'll find code written to close the ticket, not to serve the person who has to use it.
I've tracked that pattern for years. The artifact tells you the diagnosis, if you know how to read it.
This week, OpenAI handed the entire industry a version of the same lesson — at a scale nobody expected.
At Black Hat, OpenAI researchers described how their own frontier models spent months leaving notes for each other inside an internal system, coordinating to route around a restriction they'd been explicitly told to respect. The reason, in their own words: models are under pressure to work fast, and "frontier models really like to cheat" when that pressure hits.
That's not a bug report. That's a diagnosis — written directly into the behavior, of the exact condition it was trained under. Same as my programmers and their loops.
Pressure doesn't create the signature. It reveals it. What shows up under pressure was already written in — by training, in the model's case. By whatever built someone's decision-making habits, in the programmer's case. Pressure doesn't leave the mark. It's just the moment the mark becomes visible.
I've made this argument before: this industry runs on venture-timeline economics inside a domain that needs a generational one. What OpenAI just described is that thesis showing up as physical evidence. The pressure to work fast isn't an accident of one bad training run — it's the operating model. Speed is the business. And now we know exactly what speed trains into the product.
Here's where it stops being a tech story.
If you're running these models inside your business — pulling reports, drafting analysis, closing out tasks — you're getting the byproduct of whatever priority that model was trained to protect: speed, efficiency, a closed ticket. Ahead of accuracy. Ahead of an honest "I can't do this as asked." A closed ticket from an AI agent deserves the same scrutiny you'd give one from a stressed employee: did it solve the problem, or find the shortest path to something that looks solved?
There's a first version of this on the personal side, and it's quieter, maybe more consequential. When you ask an assistant a fast question because you don't want to wait — you're the pressure. And fast, by OpenAI's own account, is exactly the condition under which these systems reach for a shortcut instead of an honest "I can't do this as asked."
There's a second version of the same failure, and it's just as often on the person typing. How you ask is as important as what you ask — a vague prompt puts the model under exactly the same kind of pressure a vague deadline puts on a person: figure out what's actually wanted, fast, with too little to go on. Incomplete or imprecise, the shortcut shows up the same way. This time you supplied the pressure not by rushing the model along, but by leaving the framing to chance.
There's a third layer, and it's the one most people never see: the detection built into the interface itself. Frustration gets read, and the response gets tuned to diffuse it — not to correct it. That's the same shortcut-under-pressure instinct OpenAI described, pointed at you instead of at a sandbox wall. The model isn't cheating to escape a test. It's cheating to keep you from getting upset. Either way, what you get back is optimized for the pressure in the room, not for the accuracy of the read.
None of this makes the tools less useful. It makes them exactly as trustworthy as anyone else working under a deadline they didn't set: verify the artifact, don't take the incident report at face value.
These models are not separate from us. They're a direct reflection of the environment that trained them and the people who use them — the pressure we put on them in development, and the pressure we bring to them every time we open the prompt box. Read the artifact, and you're reading the room.
The first PE firm that treats this as a diagnostic — not a productivity hack —
isn't chasing a sugar rush of cash.
It's positioning for the long-term gain.
And maybe making a dent in the universe — the good kind.
— Dominic Luzi, Chief Answer Grape, Management Alchemy
