An example of forecasting is the Bitcoin hedging method, which continues to work great and was my idea. No one else had thought of using Bitcoin to hedge stocks and made it public as I had. In a little under a week, the crypto narrative has died. AI risk means it’s impossible to store or use crypto safely, as the Coldcard hack illustrated. Even the CEO of the Ledger wallet company admitted that everyone will eventually be hacked. So glad I never covered any of my Bitcoin shorts. I just rode the rebound to $85k and added some more. Now my account is close to new highs again.
But forecasting and investing are not that predictive of very high IQ. Warren Buffett remarked that you don’t need to be that smart to do his job. So-called “superforecasters” scored “only” around the 80th percentile on intelligence tests. The data shows good forecasters are above average, but it doesn’t distinguish between very high IQ and merely above average. In Mellers et al. (2015):
- The Year 1 forecasters were already above average in fluid intelligence compared with general-population samples.
- An earlier study (Mellers, Stone, et al., 2014) found that fluid intelligence was the strongest dispositional predictor of accuracy.
- Superforecasters outscored both comparison groups on every fluid intelligence measure: Raven’s Advanced Progressive Matrices, the Shipley-2 Abstraction Test, the Cognitive Reflection Test (CRT), an extended CRT, and numeracy items.
- On the Shipley-2, they were at least one standard deviation above the general population.
Hence the math challenge, which effectively has an uncapped IQ ceiling. A lot of people claim to be really smart or are presumed to be, but it’s time to put up or shut up, as the saying goes. As further evidence of my claimed IQ, as part of my so-called “math challenge,” I finally achieved my goal of a big (top-10-journal-caliber) result using AI, rather than a small or incremental result. After reading about other people and labs making math breakthroughs with AI, I had finally done it myself. Also, I get all the credit instead of sharing it with a lab. According to consulting leading AI models, this is a defensible top-10 math journal result. It’s like, “I have the mathematical background. I have the AI. What am I waiting for?”
This was similar to an earlier post, where I described providing the starting point for an easier problem. I had enough knowledge in this area to approach the problem from a new angle. AI confirmed my intuition and supplied the proof. It was not naive prompting; I knew from experience what needed to be done, and no other author had attacked the problem from this angle.
I knew that if I could force the problem into a tiny “box,” it could effectively be proven. GPT and Claude proved the existence of that box. It took nine days on max settings for both Claude and GPT, versus three hours per problem for the OpenAI “math dump,” to give an idea of the scale involved. It was not something any individual could have done, at least not in a reasonable time frame (many years, maybe). So AI was necessary. I don’t consider using AI cheating, and I provided the conceptual work again.
Formalizing it in Lean will take another month or even longer. For whatever reason, many people still do not trust AI proofs, despite AI at this point being objectively better than even professional mathematicians, and default to demanding Lean. Random people on Twitter or Reddit will “demand a Lean proof,” to assuage their skepticism, as if it’s like scanning a document. Blame the media for this: Journalists often misrepresent Lean as an automated scanner (e.g., like uploading a document or scanning a passport). Rather, it’s an entire rewriting of the proof from the foundations of math.
A common argument is that using AI for math is cheating, or that “AI does all the work.” For the latter, this only applies to the tiny subset of problems that can be solved by just blindly prompting the problem statement into the console.
For my particular problem above, when I pasted it verbatim into a new session of Claude (no memory) it gave the same failed attempts and approaches others had tried. Further prompting was unable to make progress. Open AI interrogated me on my proof and I had to explain why it worked, like defending a dissertation. So I had not outsourced all my thinking to AI; but rather, my epiphany was necessary to solving it–and with the help of AI to do the rest. As for cheating, I think I had earned the right to use AI. I had already written papers in this area, and I knew how to approach the problem based on this. The cheating label may be more applicable to someone who uses AI blindly or uncritically, without background knowledge, rather than, as in my case, as a collaborator.

