Thinking in public

The lesson of the Fable and Mythos restriction is not that one provider became politically complicated. It is that access to frontier intelligence is now contingent. Governments can intervene. Providers can change terms. Models can disappear from a workflow overnight. Any serious organization that builds its intelligence strategy on direct dependence on one model provider has confused capability with sovereignty.

The knowledge that makes your organization work is almost certainly not in your systems. It is not in the ERP, the CRM, the process manuals, or the onboarding documents. It exists in a form that has no digital representation: in the heads of the people who have been here long enough to understand how things actually work, as distinct from how the org chart says they work.

There are two distinct ways that organizations fail to deploy their own knowledge. They look very different on the surface. One is the enterprise that hands its strategic questions to consultants. The other is the precision operator whose competitive advantage is craft — accumulated expertise that cannot be outsourced but also cannot scale. Beneath the surface, they are the same problem.

Every organization that has operated in a high-stakes domain for more than a decade has built an algorithm. Not deliberately. Not in any software sense. The accumulated pattern recognition — what suppliers to trust, which signals precede client defection, how regulators actually behave versus how they say they behave — is a decision system. The problem is the substrate it runs on.

A generation of operational architects is approaching retirement across Swiss industry. The knowledge transfer problem is real, large, and barely acknowledged. The most consequential asset — the decision intelligence that makes these companies function — has no transfer mechanism. This is not a human resources problem. It is a computational one.

Renaissance Technologies didn't beat Wall Street by being better at what Wall Street does. They beat it by delegitimizing the entire epistemological basis of how Wall Street worked. Strategy consulting is next.

When the board asks "Did you just run an algorithm?" the answer that follows reveals everything about how a decision was actually made. One answer is genuine intellectual ownership. The other is expensive deference. The first answer has never been available to a CEO before ogram.

The companies that survive the transition to agentic intelligence will not be the ones with the best models. They will be the ones that understand a single metric: the number of successful decisions produced per dollar of system cost. Philippe Van Caenegem calls this intelligence yield, and it changes everything about how technology companies are valued.

On the infrastructure layer that emerges when AI becomes the default execution surface.

The thesis that the next trillion-dollar companies will deliver outcomes, not tools. Services as the new software.

A precise reframing: LLMs as databases of vector programs, and prompt engineering as a search through that space.

The foundational case for why scaling neural networks produces qualitative leaps, not just incremental gains.

The benchmark that keeps the field honest about what intelligence actually requires beyond pattern matching.

Frontier work on interpretability and alignment: understanding what models actually learn.

The engineering discipline behind deploying intelligence at scale.

Practitioner-grade thinking on structured outputs, retrieval, and making LLMs reliable in production.