Start with Dwarkesh (most digestible)
Why I don't think AGI is right around the corner. June 2025.
Patel interviews these people for a living. Long, technical, unhurried conversations with the researchers actually building this. So when he says he isn't convinced, it lands differently than a random sceptic saying it. (Would recommend listening to his podcast as well)
His whole argument is one idea. Models don't learn on the job (which is maybe less true now).
Someone you hire in January is better in June. They picked things up, sat in the meetings, got the context. A model doesn't do that. You can't hand it feedback that sticks, so you get roughly the same thing out of the box every time, forever.
Then Aschenbrenner
Situational Awareness. June 2024. Around 165 pages, so give it an afternoon (but definitely worth the read)
Aschenbrenner was on OpenAI's superalignment team until April 2024. He left over an alleged leak he disputes, and wrote this not long after. Columbia valedictorian at 19, which tells you something about the register before you open it.
The method is extrapolation. Count the orders of magnitude coming from compute, from algorithmic efficiency, and from what he calls unhobbling. Draw the line forward. You land on AGI around 2027.
Then it turns geopolitical quickly. Trillion-dollar clusters, a race with China, and eventually the national security state taking the whole thing over.
AI 2027
AI 2027. April 2025. Daniel Kokotajlo, who left OpenAI, Scott Alexander of Astral Codex Ten, and three others.
Month by month from mid-2025 to late 2027, following four generations of AI agents as each one builds the next. It ends twice on purpose. One branch where the race keeps running, one where it slows.
It paints a (caveated) picture of how AI may progress and the two paths we could turn down.
The Intelligence Curse
The Intelligence Curse. April 2025. Luke Drago and Rudolf Laine.
The title borrows from the resource curse, where countries sitting on oil somehow end up worse governed than countries without it. The money arrives whether or not anyone builds a decent school, so eventually nobody builds one.
Now swap oil for intelligence.
Every state and company currently needs capable people. It's what moves industries like education, healthcare - the whole ladder up.
Take the need away and none of it gets cut out of malice. It just stops mattering to anyone with the power to fund it.
They aren't fatalists about it, which I appreciated. Their answers all point one way: get the tools into individual hands, augment people now, align models to users instead of institutions.
The Last Invention is the odd one out
The Last Invention. May 2025. Alex Brogan. Twelve parts.
Actually written by an Australian. It goes into a nuanced dive over how AI may impact specific industries.
This one asks what you're doing about it while there's still room to do anything, which it calls the prep window. Work, then money, then meaning.
Start here
Start with the Dwarkesh piece, work your way into it and give yourself enough time to go through them all.
Built by Alex Sidhu, Whitehorse AI.