InOrbital

COMPOUND HUMAN INTELLIGENCE
We build systems that help human intelligence compound across individuals, institutions, and generations.

[InOrbital @inorbitallabs]

The same protocol can be evidence-based, and still be wrong for you.

Because evidence tells us what works on average. It doesn't tell us what will work for this brain, under these particular conditions.

We usually explain this gap with effort. One person was more disciplined, followed the protocol properly, tried harder. The other didn't. But there's a simpler explanation that has nothing to do with willpower: the protocol was never matched to the person in the first place.

A brain receiving an intervention isn't a blank, standardized machine. It's the product of a specific history, development, learning, habits, stress, prior experience, reward, failure, thousands of small adaptations accumulated over years. So when two people get the same instruction, they aren't starting from the same cognitive state, even if the instruction sounds identical.

Take "do two hours of deep work every morning." For someone with a clear objective and strong attentional control, that's exactly what they need. For someone with ten unresolved decisions competing for attention, the real problem isn't sustaining focus, it's deciding what deserves focus in the first place. For a third person, the plan is clear and the problem is initiation: they know exactly what to do and still don't start, which more calendar time won't fix. Same instruction, three different mechanisms underneath it.

The same thing happens in learning. Give two people the same curriculum at the same pace, and one with a strong existing mental model will benefit from more challenge, while one missing foundational representations needs the opposite: less load, more scaffolding, time to repair the representation before adding difficulty. This is close to what Robert Bjork calls desirable difficulty: the right amount of friction accelerates learning, but only when it's calibrated to what the learner already has to work with. Miscalibrate it either direction and the same curriculum stops teaching.

This is why individual differences matter so much in cognitive science and neuroscience. Population-level findings tell us what tends to work on average. They don't guarantee any single brain will respond the same way.

Behavioral science adds another layer: what people say they need and what actually changes their behavior often aren't the same thing. Someone says "I need more motivation," but their behavior points somewhere else entirely, an ambiguous first step, a reward that arrives too late, a plan sized well beyond their actual capacity this week. Peter Gollwitzer's research on implementation intentions found something similar: people who form specific if-then plans (if it's 8am, I open the document) follow through far more often than people who just intend to do something, because the bottleneck usually isn't desire, it's a missing mechanism for translating intent into action. The useful question stops being "how do we motivate this person" and becomes "what mechanism is actually blocking the behavior."

Now bring this into building a company. Two founders both say "I want to build an AI company," and that sentence alone tells you almost nothing about what protocol they should follow. One has a validated problem, a technical cofounder, capital, and full-time availability, so the right move is toward an MVP and iteration. Another has an early idea, ten hours a week, and real uncertainty about the problem itself, so the right move is closer to clarifying the problem and generating cheap evidence before committing to anything larger. A third founder already has a working company: revenue, product, team, but every important decision still routes through them personally. Their question isn't "how do I build" anymore. It's "how do I build a system that no longer depends on me." Same domain, same stated ambition, three completely different protocols.

This is where personalization gets more interesting than a better recommendation engine. A system built to adapt shouldn't open by asking what type of person you are. It should open by watching: what are you trying to do, what do you already know, what do you keep doing regardless of what you say, where do you get stuck, how do you respond when something doesn't go as expected. That last one matters more than it sounds. A prediction made before acting gives you something an outcome alone can't: a window into the person's actual model of the world right now. Prediction, action, outcome, prediction error, repeated enough times, and the system starts learning not what someone says about themselves, but how they actually create, decide, and adapt.

Neuroscience helps reason about the conditions under which attention and decision-making operate. Cognitive science explains how someone processes information and forms mental models. Psychology speaks to motivation, identity, and internal models of self. Behavioral science tells you what actually happens when intention meets reality. Real-world behavior is what keeps the whole loop honest.

None of this is about finding a permanent label for someone. It's building an increasingly accurate model of how this particular person creates, learns, and decides, and what conditions let them do their best work, knowing that model has to keep updating. People aren't static. The protocol that's right for you today may be wrong in six months, once you've moved from exploring to committing, from learning to executing, from needing structure to needing room to move.

So the future of personalized systems probably isn't about finding the perfect plan. It's about building something that keeps rediscovering which plan makes sense for this person, under these conditions, right now.

Cookie-cutter advice doesn't usually fail because the science behind it is wrong. It fails because a population-level average gets mistaken for an individual-level protocol. The science tells you what tends to work. Personalization asks the harder question: what should work for this person, here, now, and how will we know if we're wrong.