Fluxive is where the Influence Design in the AI Era series lives, along with the research underneath it. Both are about the same question from two directions: how work should be organized once it is shared with machines, and what the machines themselves are able to account for.
The research half starts from a quiet problem. The processors that run modern AI systems have no notion of what those systems are doing. Nothing is broken and no alarm sounds, and the industry's answer so far has been to put more software on top of it — which seems worth understanding properly before much more gets built on the assumption that it is handled.
It is unhurried work. The output is writing: books first, essays and notes as the thinking settles. Everything finished enough to read is on this site.
A series in three volumes. The first is out, the second is being written, the third is planned. Each takes the argument a step further out from the machine and closer to the organization around it.
How organizations should design influence, decision boundaries and operator accountability once the work is shared with machines. Taught against real business cases.
Why governing machines turns out to have rules that behave like physics — and what follows for the way we work, build and lead. Nine chapters, from the shift to ecosystems through to the operator's decade.
The third volume, where the argument reaches the surfaces themselves — what it means for behavior to be something a system can be built to express rather than something bolted on afterwards.
Four observations, in order. None of them requires anything to have gone wrong.
A general-purpose processor executes instructions. It does not interpret them. The chip cannot distinguish a document summary from a database exfiltration, and that indifference is by design — it is the substrate on which every modern AI system runs.
Silicon has accumulated defenses for four decades: privilege rings, memory management, secure boot, confidential-compute enclaves. Each targets a specific adversary class — memory corruption, privilege escalation, isolation breach, key theft. None of them evaluates whether a workload's next action falls inside a behavioral envelope.
Authorization is not governance. A controller with permission to drive the motors can drive the robot past the finish line. Nothing on the chip was wrong. The chip asked whether the code was allowed to run; it never asked whether the action was supposed to happen.
The gap has a name: the ungoverned-compute problem. The base layer on which AI systems run has no notion of behavioral policy, so every guardrail deployed today sits above it, in software, sharing the same substrate as the workloads it constrains. Closing that gap is the research direction.
Fluxive is a small research office. Its published output is books, essays and the working notes behind them. It is not a product company and it is not trying to become one — the size is a choice about how the work gets done, not a stage on the way somewhere.
What that buys is the ability to follow a question for as long as it takes, including the stretches where it does not look like progress.
More about the workLetters about the books or the research are welcome. Students and people writing about this are welcome too — ask, and the office will usually send whatever is useful.
Get in touch