The LLM is the wheel.

Undeniably foundational to the next era of human progress. Multifaceted in ways nobody has yet enumerated. Transport, grain crushing, the conversion of steam to electricity. But you cannot get to the workshop the day after inventing the wheel and expect to drive home in a Daytona 365 GTB/4. The wheel is sufficient. The infrastructure to make it useful is not.

Today's models are far too dumb to replace human experts in serious professional work. They are also far smarter than the threshold needed to make those experts' lives dramatically easier. As an industry we have spent the last three years chasing the former, at the cost of the latter. Some of those replacements have worked. Replacing a non-load-bearing member of a truss with papier-mache does not bring the building down. It also does not help you build a taller one.

Models have been sufficiently capable since at least 2024. They are still rarely being deployed in any way that gets the existing work done better, faster, or with less of the principal's time. This idea is heresy to a San Francisco AI disciple, and painfully obvious to anyone trying to run a serious specialist firm.

The Productivity Paradox. Every general-purpose technology in history has rewarded the same small group of people. The ones who recognised that the technology only delivers value when the system around it is redesigned. The electric motor sat in factories for thirty years, replacing the steam engine while the line-shaft architecture stayed in place, before anyone realised the motor's value was in eliminating the line shaft entirely. The personal computer raised productivity in six industries, in the hands of operators who rebuilt their work around it. The rest of the economy bought computers and kept having the same meetings. The pattern holds. The technology arrives. Most buyers bolt it onto what they were already doing. A small minority rebuild the work around the new capability. The minority compounds.

The Compiler. The dominant deployment pattern for AI in professional firms is the chat box. The principal types what they want, the model produces output, the principal reads it and revises and tries again. This is the leather belt of the AI era. A lossy, centralised interface that degrades intent at every pass. The output looks right at first glance. It is wrong in the ways only a practitioner of the work would notice. The first ninety per cent feels like magic. The last ten per cent collapses.

The alternative is to treat the model as a compiler rather than an oracle. Compilers do not interpret intent loosely. They require structure, conventions, source code that conforms to a syntax. They produce deterministic output from precise input. The principal's intent is the source code. The model is the compilation pass. The deliverable is the work. The interface between the three is where Tacita lives.

The Principal. The dominant ambition of the AI industry is to replace human cognitive work. Tacita does not share this ambition. The principal of a specialist firm is not a cost to be eliminated. They are the load-bearing point. The place where the firm's judgment, taste, network, and accountability actually live. The work Tacita does sits underneath them, not in front of them. The analyst pack, the diligence file, the model behind the call. Assembled to the standard the principal would have produced, in the form the firm already operates in. The call stays with the principal. It always will.

Tacita is the work behind the work.