I originally posted this as a quick take, but I just discovered that I needed to make it a post in order to include it in a sequence.

This is a brief comment on an old version of the strategy course (August 2026). I created this note to share with students in the technical and governance courses as many students went through the strategy course first.

• Bluedot uses power concentration rather idiosyncratically by including rogue AI as part of this pathway (I expect they made this choice so as to avoid needing a fifth pathway; also they included a video where Richard Ngo makes an fascinating argument for analysing risks in this way, though as far as I can tell, this seems to be a pretty idiosyncratic position). The term “power concentration” generally isn’t used this way within the broader community, so I thought I’d mention this to reduce potential miscommunication. It’s worth noting that the people who are most worried about rogue AI and the people who are most worried about power concentration often propose opposite interventions: risks from rogue AI can arguably be reduced by maintaining high standards in terms of who is allowed to build rogue AI, whilst people who are worried about power concentration would generally prefer AGI to be developed as an open source project.
• Honestly, I’m not really a fan of using the term “power concentration” to describe a risk, even though I sometimes find myself using the term that way myself. I’m generally making a distinction between the risk and the intervention. I worry that using the term “power concentration” to describe the risk causes people to assume that the solution is automatically to distribute power. I think this is a reasonable proposal, but I want to see people to pursue solutions because they’ve consider the alternatives and decided that their proposed intervention is their best option, not because they were pushed in that direction by the problem framing.
• I am not really a fan of the defense-in-depth framing. For example, you can’t stop a tank by putting ten kids with slingshots in front of it. Some layers matter more than others and focusing on defending at all possible layers can obscure this. I also think it’s important to think about where we want to head: we need a combination of shorter-term measures to buy us time and longer-term measures that take us towards a stable solution. At the same time, I’m aware that defense-in-depth is the default strategy in the absence of co-ordination and I think it’s a useful frame to have worked through. Going further: Oscar Delany has analysed the strengths and weaknesses of defense-in-depth in-depth =P.
• Kill chains are useful, though it is worth being aware of, not only the strengths, but also the weaknesses and limitations of this tool so that you know when to use it (and when not to). In general, not knowing the weaknesses or limitations of a tool doesn’t mean that there aren’t any, it just means that you don’t know what they are. My personal view on strategy is that it isn’t so much about learning specific tools, but learning when to use the tools that you learn.
• One framing I find fruitful for describing AI risk succinctly is what I personally call the SUV Triad - speed, uncertainty and vulnerability. AI appears to be moving fast, there seems to be a lot of uncertainty about how it would develop and what needs to happen for things to go well and there are many different vulnerabilities (or threat models) we have to worry about. It is the combination of these three factors that worries me.

The latest version of the strategy course opens with a Personal Statement by Daniel Selsam. If you went through an older version and you wanted to read just one thing from the new course, this would be my top recommendation.

(Link to Substack version)