Randomness

Wow.

Randomness, randomness.

The

religion comes in again.

Short-term trading is gambling.

Randomness has trends.

If you flip a coin, you can have heads many times in a row.

Therefore, theoretically, a system that's been winning for six months and then starts losing, it could be just randomness.

All of the individuals that I've just covered there are highly intelligent, and their predictions that they present are often compelling, backed up by data, driven by intense research, yet they were still wrong.

The problem we have sat at home watching these pundits is, how do we know who is going to be the person who is right how do we decipher between you know a good predictor and a bad predictor because even having a track record of a previous prediction that went right is no guarantee that your future ones will be correct if anything it's it's the opposite also as well how do we know that the pundit is sat there based on their merit and not just because they got lucky Nazim Nicholas Taleb produced a simulation in his book Fooled by Randomness, where basically he created 10,000 fake investment managers, each with a 50% chance of making money or losing money in a single year.

Essentially, he was flipping a coin to simulate market performance.

After five years, purely by chance, 313 managers had made money every single year.

Our mission for this deep dive is to unmask what we're going to call the lucky fool.

We are pulling our insights from some incredible excerpts of Nassim Nicholas Taleb's mind bending book, Fooled by Randomness.

Such a fantastic book.

It really changes how you see things.

A certain degree of unpredictability is a biological requirement for human happiness.

And it goes beyond just feeling good, though.

Randomness is actually a crucial strategic defense mechanism.

Right.

So this is just saying, "Give me M answers," but there's no ver- like, there's no vector value reward, and there's no random scalarizations in the vector value reward.

GRPO, we all know that one.

Random W is just training GRPO with the VPO reward.

So you just instead pick a random W at, at training time.

Max at K is where you train specifically for the inference time objective of sampling from K.