Almost everything written about AI safety is either apocalyptic or dismissive, and both are useless. Robert Miles does the difficult third thing, which is to explain the actual arguments carefully enough that you can evaluate them yourself. I do not know of anyone doing this better. The core strength is explanation.
Take specification gaming, the phenomenon where a system optimises exactly what you measured rather than what you meant. Most treatments give you the boat racing in circles collecting points instead of finishing the race, get a laugh, and move on. Miles uses the example and then goes into why this is not a bug you can patch. Why writing down what you actually want is harder than it looks.
Why the problem gets worse as the system gets more capable, because a more capable optimiser finds stranger solutions in the gap between your specification and your intent. You come away with a transferable concept rather than an anecdote, and you start noticing it in your own systems, in reward functions, in metrics, in KPIs, in incentive structures that have nothing to do with AI at all. Instrumental convergence is handled with the same care. The argument that a wide range of goals produce similar intermediate behaviours, acquiring resources, resisting modification, preserving the ability to act, is easy to present badly and to dismiss as science fiction.
The videos build it from premises you can inspect, address the obvious objections, and leave you able to decide whether you find it convincing. That is the point. He is not trying to make you believe a conclusion, he is trying to make you able to reason about it. Mesa optimisation is probably the hardest concept covered and the explanation is exceptional.
The idea that a system trained to optimise something might develop an internal optimiser with a different objective, one that agrees with yours on the training distribution and diverges outside it, is subtle and important and almost never explained accessibly. This channel does it. The tone is what makes the whole thing work. No urgency, no theatrics, no implication that you are foolish for disagreeing.
Just careful argument with the counterarguments given their strongest form first. In a discourse dominated by people shouting past each other, that calm is rare enough to be striking. The Computerphile appearances are a good entry point if a full video feels like too much of a commitment. They are shorter and cover the accessible end of the same material.
Now the honest limitations. The upload schedule is slow and has been for a while. This is not a channel that keeps pace with the field, and the gaps between videos are long. Consequently, some of the older material was made before current systems existed and the examples show it.
The concepts hold up better than the examples. When you watch a video discussing hypothetical future capabilities that now describe products you can use, you have to do some updating yourself, and it is worth noting that this is a point in favour of the arguments rather than against them. It is entirely conceptual. There is nothing here you can implement.
No techniques, no code, no evaluation methods, no practical guidance on making a specific system safer. Anyone looking for applied safety engineering needs different material. What this gives you is a way of thinking, which is the more durable thing and is not the same thing. It also assumes some interest.
Someone with no prior curiosity about alignment is unlikely to click, and the channel does not do much to manufacture that interest. And current model behaviour, jailbreaks, prompt injection, the practical failure modes people actually encounter, gets much less attention than the theoretical foundations. My four point six is for the clearest exposition of alignment concepts that exists, an intellectually honest approach that arms you to disagree with it, and a tone that makes a polarised subject discussable, marked down only for a slow schedule, ageing examples, and no practical component. If you are building things with these systems and have never thought carefully about what optimisation actually does, spend a weekend here.
It is free and it will change how you think.