My philosophy is that if you have a belief about something that matters, you should test that belief to see if it holds up to reality.
Below is a list of beliefs that I have. Each is associated with a probability (in increments of 10%; I do this to make it easier to assess my calibration at the end of each year.) For some, these probabilities come from whims or light research, and are meant to test how accurate my instincts are around world events. These are “loosely held” beliefs, which can be dislodged fairly easily. Even a 90% probability here really means “my model of reality gives 90%”; but these are areas where I can be convinced that my model of reality is wrong fairly easily, in which case the number can jump.
For other beliefs – “strongly held” beliefs – they come from a foundation of core principles and battle-tested understanding of the world. They can still absolutely be wrong; but whether my probability is 50% or 90%, they’ll be harder to move than my “loosely held” beliefs.
But I don’t want to be hard to move if I’m wrong! So for all my strongly held beliefs, I also share my thoughts on one or more cruxes – foundational sub-beliefs which, if proven to be inaccurate, would change my thinking significantly. Breaking beliefs into cruxes is extremely helpful for staying open-minded, because it helps move from interminable debates (“people are basically good!” “No, they’re basically bad!”) to tractable debates (“most people would choose to do a selfish thing under these conditions”) to either resolutions or the realization that you’re not actually debating the thing you think you are.
I’ll write more about cruxes and probabilistic bets sometime and link those posts here when I do. But for now, here are some beliefs of each type.
Strongly Held Beliefs
Recursive Self-Improvement Won’t Create God
Chaos theory is the branch of mathematics which describes how many things – famously including weather among many others – are chaotic…which specifically means they can behave in vastly different ways than you’d expect given tiny inaccuracies in your initial measurements.
I believe that most of the things that matter are chaotic – or worse, anti-inductive. I believe that broadly includes human nature. And as a consequence, I believe that even if recursive self-improvement generates a “super-intelligence”, that super-intelligence will be far less able to effect drastic change in the world than most people seem to think.
That’s not to say it won’t matter. Superintelligence will significantly change the way we code, the way we read and write, and the way we discover knowledge and educate – it already has. It may change, over time, the way we drive, the way we work, and even the way we date and find love. What I specifically am doubting is that a sufficiently smart superintelligence will be able to craft strategies to achieve any arbitrary goal. I do not believe that the first people to “own” a superintelligence “just win”; I believe they get an advantage similar to the comparative gap between the modern (pre-AI) age and the mid 1900’s. That’s a big advantage, but it’s not god-level transformative, and leaves plenty of room for smart ‘outsiders’ to still engage.
Specifically, my cruxes here are:
- Most important things are hard to measure, and therefore cannot be fed as accurate inputs even to a superintelligent AI. This includes human beings’ mental states and predilections as well as natural phenomena like earthquakes, meteors, and oil reserves, and especially applies at large scale.
- To persuade me otherwise, I’d need an actual theory of how AIs would model people in particular. The “scary accuracy” of recommenders like Facebook and the like is not sufficient proof for me; this is drastically different from the material influence that would be required to e.g. rewrite world governments, be a maximally effective investor, or disseminate life-changing technology at truly rapid pace.
- Even if we could measure things, even the slightest inaccuracies will actually change the outcomes.
- To persuade me otherwise, I’d need to see evidence that human beings actually behave repeatedly in similar-but-slightly-different circumstances. Things like twin studies seem to provide counterevidence, as does the inconsistency of most AI agents (not that AI agents are great approximations of people.)
- Specific Predictions:
- <10% of any AI directly governing any territory currently (as of Oct 2026) belonging to a sovereign nation; INCLUDING through a figurehead human leader who ‘merely’ defers all meaningful considerations to an AI.
- Using an AI prominently does not refute this! This only applies if the human explicitly will not or cannot override the AI even if they decided they disagreed.
- 80% that there are still at least 4 separate, highly competitive firms at the peak of AI research as of 2030, even if they are different firms than the ones in the lead today.
- 90% that for any “material breakthrough” in AI chip design made by software companies or their agents independently, at least three other material breakthroughs” are made by actual chip/hardware companies.
- 70% that NVIDIA and Apple (or otherwise the two most valuable companies operating in the hardware space) have market caps larger than OpenAI and Anthropic (or otherwise the two most valuable companies operating in the software space).
- <10% of any AI directly governing any territory currently (as of Oct 2026) belonging to a sovereign nation; INCLUDING through a figurehead human leader who ‘merely’ defers all meaningful considerations to an AI.
Loosely Held Beliefs
On a random smattering of world-relevant events intended to help me take a quick pulse check on my intuition.
At the end of 2026…
- Will the EU have required mandatory age verification on social media or AI?
- 10%
- Will there be a bilateral ceasefire in the Russo-Ukrainian conflict?
- 10%
- Will Benjamin Netanyahu still be Prime Minister of Israel?
- 80%
- Will any AI system exceed 97.5 on ARC-AGI-2?
- 20%
- Will an AI-created song chart in the top 20 of the Billboard Hot 100?
- 20%
As Of EOQ1 2027…
- Will the best bot in the ACX 2026 Predictions Contest on Metaculus exceed the the best human score?
- 10%
- Will Democrats win at least 218 seats in the US House in Nov 2026 elections, based on AP race calls?
- 60%
- Will Republicans hold at least 51 Senate seats when the 120th Congress first convenes?
- 60%
- Will a lapse in federal appropriations (a government shutdown affecting at least one cabinet department) be in effect on any day between November 1, 2026 and March 31, 2027?
- 30%
- Will Donald Trump’s job approval in the RealClearPolitics average be higher on March 31, 2027 than 38.8% (his rating as of October 15, 2026)?
- 20%
As of EOQ2 2027…
- Will a sitting U.S. Supreme Court justice die, resign, or publicly announce retirement between October 15, 2026 and June 30, 2027?
- 20%
- Will the President formally invoke the Insurrection Act between October 15, 2026 and June 30, 2027?
- 40%
- Will at least one of OpenAI or Anthropic begin trading publicly on a U.S. stock exchange?
- 30%
- Will a federal law that explicitly preempts or bars state-level AI regulation be signed?
- 20%
