Thomas Dullien: Three Lessons That Changed How He Thinks

Thomas Dullien, a security researcher better known by the handle Halvar Flake, published a personal essay on 21 August 2026 titled “Three important steps in my maturation process”. Its starting point is two personal events: the recent death of his father, who was twice his age, and joining a company that skews younger than him. Dullien is now roughly the age his father was when he was born, and notes that he has become “the old generation” — no one is left in the generation above him.

Understanding your own incentive structure

The first lesson: do not believe everything you think, and examine your own incentive structures carefully. Oppenheimer was publicly guilt-ridden about the creation of the nuclear bomb; von Neumann quipped that some people profess guilt to claim credit for sin.

In his younger years, especially when he had 0day nobody else had, Dullien agonised about the responsibility that comes with it: should he fix them, or use them for good? Today he argues that while individuals matter, many ideas have a time at which they are ripe, and the actions of an individual matter less than that individual thinks in the moment; there is also almost no way to predict how one's work will affect the broader world. The same exploit could be used to apprehend a terrorist — good — or to arrest someone and then waterboard him 183 times — bad.

His conclusion: anxiety about the impact of your work is self-flattering and has to be kept in check, because history will largely route around you. The valuable skill is metacognition — observing your own thoughts in a detached manner and asking how you might be the villain in this story.

Monocausal determinism is an illusion

The second insight concerns programming culture. Dullien argues that the monocausal determinism young computer enthusiasts get used to is an illusion that generations of electrical and process engineers spent their lives perfecting and maintaining — which let computer scientists get away without probabilities or empirical grounding for a long time.

Computing machines are physical devices: they wear out, items differ in quality, and they behave only “probabilistically deterministically”. Push temperature, voltage, electromagnetic fields or rapid accesses to adjacent DRAM rows, and determinism goes out of the window. That makes him wonder about model alignment: even a perfectly aligned model is subject to random bit flips during inference, and guarantees are hard to imagine in the presence of flips to inopportune values at inopportune times.

Reason and emotion are not opposites

The third conclusion is cultural: the western belief that reason and emotion are two ends of a spectrum is a construct, as is the idea that higher-order reason must reign in basal emotions. In most non-western cultures, integrating rational deliberation with impulses and emotions is more common — and, he says, closer to biological reality.

The logical argument follows: if using more information improves decisions, deliberately eliminating a source of information makes them worse. That is not an argument for acting on impulse alone — only for integrating the full spectrum, emotions included.