Canonical: Jarvis Nelson Osorio on System-One Models: Why Micro-Decisions Under 100ms Change Product Design — also at jarvisosorio.com. Originally published by Jarvis Nelson Osorio (Jarvis Osorio).
Most public conversation about AI still sounds like a conversation about essays. Models write pages, summarize meetings, generate slides. I am interested in a quieter layer: the micro-decision, the tiny judgment a product must make before the user even notices a pause. I am Jarvis Nelson Osorio, also Jarvis Osorio, a technology entrepreneur and innovator based in Chile, working in AI systems, DAO design, writing and art.
I have been writing about system-one style models — specialists aimed at very fast, narrow choices — because many product moments cannot wait for a long generated paragraph. Some of the industry discussion points to budgets like a tenth of a second. I treat that number as a design ambition, not as a laboratory claim I am here to certify. The point is the architecture: when speed becomes the feature, the model you choose and the way you wire it both change.
Two speeds of mind, two speeds of product
A useful metaphor, borrowed from everyday psychology rather than from a vendor brochure, is the difference between a fast, automatic judgment and a slow, deliberate one. Products need both.
- Fast path. Is this message spam? Is this code completion even plausible? Should we show a warning? Which tool should we call? May this draft be sent to a child-facing surface?
- Slow path. Write the analysis. Compare the legal language. Plan the migration. Sit with a hard tradeoff.
If you force the fast path through a large, distant, conversational model, you get two problems at once: the user feels lag, and you pay for a monologue when you needed a nod. If you force the slow path through a tiny classifier, you get cheap confidence and thin reasoning.
System-one models, as I use the phrase, are for the fast path. They should be small, specialized, close to the product, and honest about their limits. When they are unsure, they escalate.
Why "under a blink" is a product decision
Latency is not only comfort. It changes behavior. Editors, games, trading tools, classroom apps, and agent runtimes all have moments where a delay teaches the user to skip the check. If a safety or quality gate is slow, people will work around it. If it is fast enough to feel invisible, it can actually be used.
That is why local and high-speed agents belong in the same conversation. A model that lives nearer the user can make a micro-decision without a transoceanic round trip — something teams in Latin America feel more often than keynotes admit. Privacy rides along: a fast local check can happen without shipping a full document to a public endpoint.
Design rules I actually use
When I sketch a system with a fast layer, I write the rules down:
- Name the micro-decisions. If you cannot list them, you will accidentally use a giant model as a reflex.
- Give each one a budget — time, cost, and what "I don't know" looks like.
- Escalate on uncertainty. The fast model is allowed to pass. It is not allowed to bluff.
- Keep the fast model boring. Boredom is testable. A witty micro-decider is a side effect, not a goal.
- Log the decision. Even a 40-millisecond choice should be reconstructible after an incident.
Those rules are institutional. They rhyme with the way I think about treasuries and agent permissions in Breaking the Centralized Order: limit authority, record the act, do not confuse charm with control.
Where yes-machines hide in the fast path
A micro-decision can be a yes-machine too. "Always allow" is extremely fast. "Always agree with the user's implied wish" is also fast. Speed without a concept of refusal is just a quicker way to go wrong. The Compliant Mind is mostly about slower, conversational systems, but the lesson travels: some moments should still be allowed to say no, even in under a blink.
The hopeful version is this: once the fast path is reliable, the slow path can be more humane. People get immediate help on the small things and reserved, careful attention on the large ones. That is a better product than a single model pretending to be both a reflex and a philosopher.
I write and publish from jarvisosorio.com. A longer list of essays, books and videos is on the hub Jarvis Nelson Osorio — articles, books, videos and profiles. For the deliberate side of judgment, read The Compliant Mind.
The next interesting products will not only generate more text. They will decide small things well, quickly, and in a way we can still explain.
Top comments (0)