I’ve learned to recognize a pattern. When a topic starts generating strong opinions everywhere, from family group chats to the LinkedIn feed, and people’s conviction grows much faster than their actual technical grasp of it, the real question has already moved somewhere else. Widespread certainty about a subject tends to be the clearest sign that what actually matters has slipped underneath the surface everyone is arguing about.
We saw this during the pandemic. Suddenly the whole world was debating vaccine platforms and immunity mechanisms with the confidence of people who’d spent twenty years in infectious disease research. Certainty fired in every direction, ignoring the gap between how strongly people felt and how much they actually understood. Wherever conviction turns epidemic, it’s worth getting suspicious. The surface only gets crowded once what really matters has already sunk deeper.
In AI, that symptom got a name: tokens. The conversation got swallowed by cost, by prompt tricks to save a few cents, by the whole vocabulary of tokenomics, repeated by people who’ve never opened a terminal to run a script or looked closely at how any of it works. Common wisdom said the advantage belonged to whoever had more: more data, more GPUs, more capital. It looked like a brute-force race measured in billions.
The detail everyone missed is that problems framed purely as “whoever spends the most wins” are rarely solved by whoever has the biggest budget. The real break came from whoever refused to play that game at all. The shift didn’t come from optimizing the expensive model further. It came from tearing down the architecture, activating only the parameters a task actually needed and cutting the dead weight. Engineering stopped asking how to make the process cheaper and started questioning the structural premise itself. While the feed was still dissecting API pricing, the real innovation was already happening one layer down, in the architecture nobody outside the field was discussing.
That became obvious once new entrants proved you could train frontier-level models for a fraction of the historical price. The cost barrier turned out to be just the shell. Underneath it waited a thornier question, one that shows up in a familiar scene for anyone in technical leadership: the moment of execution.
Picture a concrete case. Camila leads technology at a healthtech company. Her team built a predictive triage engine that runs on highly sensitive clinical data. A leak here would be a nightmare for any regulator. The data runs into the millions of records, at a volume where human review simply isn’t an option. She has to rely on AI. Now she has to choose which model.
On one side, a cheaper model with open weights, released by a lab of murky background. Murky, not wrong. Not ugly, not a scam either, just short on detail about how its decisions actually get made behind the curtain. On the other side, the traditional giants, expensive and everywhere. Her cursor hovers over the decision. “Just run it locally” doesn’t solve this. Controlling where the data flows doesn’t undo the opacity of how the model was trained. Owning the weights doesn’t mean knowing what went into them. Infrastructure choices just push the risk under a different rug.
The benchmarks already settled which model is faster or more accurate. Camila’s real anxiety is a different question: which of these black boxes is she willing to hand her company’s core business and her patients’ confidentiality to.
Handing patient records to a model with murky governance sounds irresponsible on its face. The tech industry’s history with data misuse and IP appropriation is long, and the instinct is to run from vendors with questionable incentives.
The bitter part is realizing that compass is broken too. Look closely at the usual suspects, the traditional giants included, and the picture isn’t necessarily cleaner. Their corporate interests rarely line up with yours, and decades of built-up credibility tend to get liquidated the moment it’s convenient. If corporate virtue and predictability are just theater funded by different marketing budgets, the whole idea of trust as a selection criterion falls apart. The compass is, in fact, broken. And Camila is still there, under pressure, with the system needing to go live Monday.
I’ve sat in positions where operating with a broken compass was simply the routine. In technical leadership and advisory work, we sometimes joke that the real options range from bad to worse. A former leader of mine put it exactly right once: to manage is to make irrevocable decisions with insufficient information.
From the outside, leadership looks like finding the correct alternative. Sitting in the seat, you find out it isn’t that simple, because you never have all the information, and time keeps pressing toward a decision anyway. Comfortable, reversible trade-offs rarely survive contact with the real world.
The pandemic shows the other side of that same coin. Public health leaders faced brutal bets: the guaranteed lethality of a virus left unchecked against the uncertain risks of vaccines developed under pressure. Choosing the irrevocable option in the dark means accepting the risk of an unlikely disaster to avoid a guaranteed one. The real risk of a vaccine causing problems years down the line was the smaller bet, once weighed against a virus left to run without any brakes at all. Where, in that moment, was the perfectly tested vaccine with zero long-term side effects? That’s exactly how executives make a good share of their decisions too. What justified those calls was never some untouchable purity. It was the ability to measure, correct, and challenge the course as reality came in. There was room to be proven wrong.
That room to be challenged is what separates a defensible technical decision from Russian roulette. With the old moral yardstick broken, Camila has to make peace with choosing between imperfect options, holding onto a single compass point: which of these mistakes can I actually defend myself against?
If the system goes down, she needs a way to audit, to sue, to migrate. Going up against a company that answers to courts and regulators leaves a way out. Betting on someone operating beyond any jurisdiction’s reach closes that door. One kind of mistake allows a legal fight. The other ends the game outright.
Of course, having someone to appeal to is nowhere near a guaranteed win. My father-in-law is my go-to on anything related to cars, and once, after we’d settled on a set of tires, I asked him the obvious question about the warranty. He looked at me and said it straight: life doesn’t come with a warranty.
That sums up the whole practical problem here. Proving a lab misused your data is an uphill legal battle. Our legal frameworks promise proportional remedy in theory, but courts are murky in practice. The right of recourse doesn’t work like accident insurance. It only guarantees you get to step into the ring and fight.
The illusion of zero risk needs to go. Working with technology as an adult means carrying the weight of irrevocable decisions without the comfort of having made “the perfect choice.” We’ve reached a point where no new open-weight release, no new tool, no benchmark jump is going to save us from this, for the simple reason that the problem stopped being technical a long time ago.
Once cost stops being the bottleneck, dilemmas like this one will keep showing up in everyday use of AI. Facing a decision, with no guarantee the companies involved don’t have hidden interests, knowing you’ll have insufficient information and no way to reverse the choice, which option would you pick? There’s no option that guarantees you won’t get it wrong, or that you’ll pick the best one. That’s the one illusion this is meant to take off the table. Everything else, which trade-off, by what criterion, at what price, is the decision nobody else can make for you, in your seat, with your finger on the button. That choice can’t be outsourced to AI. It’s yours.