Personal website of Christian Henning, machine learning researcher and engineer. Home to a blog on ML theory, engineering and research leadership — and on what it takes to move frontier ML from research to real-world production.
AI timelines never converge, a life-long belief about human intelligence keeps shaping how new evidence gets read, and an imprecise definition of AGI means the two sides were never comparing the same thing. The first is a well-documented mechanism from the psychology of belief polarization; the second is simply definitional.
A shipped feature reveals almost nothing about how well it was built, because from the outside the quality of engineering work is a credence good. Trust accumulated across many projects is the only statistically valid way to price it.
The fastest-moving products are not the ones with the cleverest features, but the ones that can change what is already deployed. For self-hosted software that means building distribution and self-upgrade first, because without them there is no iterate step, and every cut made for the MVP quietly becomes permanent.
Every “cut it for the MVP” is one of two moves. You either defer a feature or build a throwaway stand-in. Whether either becomes debt depends not on the cut but on the boundary you leave beneath it, and on a vision clear enough to place it.
From testing and reliability engineering to team culture, the hardest problems are prevented, not fixed. Yet we chronically undervalue prevention, and let visible process stand in for the invisible outcomes that actually matter.
A provocative thought experiment. An LLM that knew only tokens would mistake their order for the structure of time, and we could break that belief without it ever noticing. The unsettling question is whether something could do the same to us.
Adding long-term memory to AI wouldn't just give it a stable self over time. It would force a design choice the discourse rarely names: one entity that consolidates many conversations, or many that diverge into a population of personalized selves.
Research is not slow engineering. Why companies that want breakthroughs need to cherish two cultures, not collapse them into one.
PFNs are often described as Bayesian predictors, but their training objective and inference mechanism suggest a more nuanced interpretation.
A critique of why modern ML uncertainty estimates lack clear semantics, reliable evaluation, and meaningful use cases.