The Exponential Advantage
Why co-intelligence compounds — and tool mode does not.
Talk 2 of 2. The first talk was the architecture; this one is the math. Why a tool gives you linear returns while a partner compounds, what changes between session 1 and session 90, and the three requirements for intelligence that accumulates instead of resetting.
One percent better a day is 37x in a year. The interesting part is that it applies to BOTH partners at once — the human gets better at prompting while the system gets better at understanding.
Every session starts from zero. You re-explain the context, re-establish the standards, re-teach the same lesson. That is tool mode, and it is why most AI use produces linear returns.
A tool executes what you asked. A partner accumulates what it learned. The difference is not model quality — it is whether anything survives the end of the session.
Three requirements for intelligence that accumulates: something must be written down, it must be retrieved into the next session, and it must change behaviour. Miss any one and you are back to episodic.
The gap is already opening between people running episodic AI and people running compounding AI. It is not visible in week one and it is very visible by month three.
Read the deep-dive on Substack
The full essay behind this talk — the why and the how, in long form.