“You need to be Jev-maxxing.”
That was the one-line verdict on a podcast about Jev , a model released the week before. It is a lazy sentence, and it is worth taking apart, because the idea buried under it is why every engineer at Unipaas now has a gateway to models beyond their own seat.
Start with the suffix. It came out of incel forums, where “looksmaxxing” meant improving your looks to raise what those forums called your sexual market value, and it reached the mainstream as irony in the 2020s. Most people who say it now have no idea, which is its own small failure of attention. Then look at how the sentence is built. Take whatever shipped this week, bolt the suffix on, and you have a sentence that costs the speaker nothing. It makes no claim that could turn out wrong, carries no price if it does, and runs entirely on the fear of falling behind. Next week it will be a different launch wearing the same suffix. Advice with nothing at stake is noise, however confident it sounds.
Strip the fad away and what was said a moment later, on the same podcast, makes sense. The claim went like this: the people who got the most out of LLMs, the ones now running companies worth tens of billions, were the ones who said “no one really knows what this is for, but I’m going to try it in every single scenario that I can think of and see what it’s good at.” That is a real claim, and it is about method rather than any one model.
Comfort is the risk now. The technology is genuinely impressive, and that is exactly what makes it easy to settle: one model, one harness, one way of prompting that works, repeated until it is simply what we do. Comfort with a tool this good passes for productivity, because output is easy to count and the count keeps climbing. Dijkstra had that ledger right decades ago: code is “lines spent”, not “lines produced” .
The answer is to dare more: the model you would not have picked, the harness you have not opened, the approach that probably will not work. Kent Beck, after finally trying ChatGPT in 2023, reckoned most of his skills had lost their value while the rest had gained enormously, and admitted he could not tell which were which: “The only way to find out is to try a little bit of a lot of ideas.” The tooling to support that is there, and within reach. What it does not supply is the noticing.
Louis Pasteur put it in 1854, opening the new science faculty at Lille: “in the fields of observation chance favours only the prepared mind.” Alexander Fleming nearly threw away the contaminated plate that gave us penicillin; the clear ring around the mould was like one he had seen before, in his earlier work on lysozyme . Henri Becquerel’s experiment needed sunlight, Paris turned overcast, and he shut his uranium salts in a drawer with the photographic plates. He developed them anyway , expecting a faint image, and found radioactivity. Neither was luck. The accident arrived in front of someone ready to read it, who then did the work to prove it. Trying more is how you produce the accident. Design is what you do with it.
That is the bet we made. We opened a gateway to every engineer in R&D for exactly this: try more
models, see how each behaves inside the harness you already use (or a different one), weigh what it costs,
and learn which model fits which job. Today that gateway is OpenRouter . It could as easily be Ollama
Cloud , or a model running on your own laptop, and our own tooling already treats them alike: harn, an
internal command that runs a coding harness, Claude Code or Codex for example, against a seat, a
gateway or a local model, with a clean environment on every switch. Swapping the gateway changes no
command. That follows from how we choose to build: stay as agnostic as we can about harness, gateway
and model, because a choice we can swap is a choice we can test. It suits experimenting, and it leaves
us ready for the next Jev, which is already arriving. Within two weeks of Jev shipping, OpenAI
announced a decision API of its own .
Because we move real money, the guardrails came first: zero-data-retention models only, secrets blocked before a prompt reaches a model provider, and no customer, merchant or production data, ever.
The stake is ours now. We spent the time to make experimenting approachable and safe on purpose; if the access sits idle, or turns up nothing we did not already know, that is our call that missed. If we are maximising anything, it is the engineer rather than the model of the week: the range of what each of us has tried, and the judgement that range builds. Access is the part we can give. The prepared mind is the part we are investing in.