Then I was doing poorly in grandmaster, until I realize you can ace grandmaster by just picking the longest explanation every time.

Literal reparations:

https://en.wikipedia.org/wiki/War_reparations

> War reparations are compensation payments made after a war by one side to the other. They are intended to cover damage or injury inflicted during a war.

Another war lost by the USA, despite its infinite military supremacy. Weird.

One observation that comes to mind is that a lot of people in the world realize that the interests of the US and of Israel are not aligned, but American policy makers keep pretending they're one and the same. This was completely self-inflicted and it's not the first time. (The Iraq war cost what, trillions?) It's quite incredible.

When the volume of content goes up, the depth goes down.

What is happening is that we're consuming more and engaging with everything more superficially. I think that's a concerning trend.

I don't dispute that in aggregate the effect was positive. But I spend more time thinking about things which impact me directly, and I assure you that in my personal life it used to be a problem, and fixing it was an improvement.

https://en.wikipedia.org/wiki/Stanislav_Petrov#Aftermath

> Petrov underwent intense questioning by his superiors about his judgment. Initially, he was praised for his decision.[2] Colonel-general Yuri Votintsev, the then-commander of the Soviet Air Defense's Missile Defense Units, who was the first to hear Petrov's report of the incident (and the first to reveal it to the public in the 1990s), states that Petrov's "correct actions" were "duly noted".[2] Petrov himself states he was initially praised by Votintsev and promised a reward,[2][22] but recalls that he was also reprimanded for improper filing of paperwork because he had not described the incident in the war diary.[22][23]

> Petrov has said that he was neither rewarded nor punished for his actions.[24] According to Petrov, he received no reward because the incident and other bugs found in the missile detection system embarrassed his superiors and the scientists who were responsible for it, so that if he had been officially rewarded, they would have had to be punished.[2][24][22][23] He was reassigned to a less sensitive post,[23] took early retirement (although he emphasized that he was not "forced out" of the army),[22] and suffered a nervous breakdown.[23]

On Centaur (hybrid LLM + classic HPO) the LLM is only called to give its opinion a fraction r=0.3 of the time (the remaining is plain HPO). But that means that:

A) the compute used by Centaur is not directly comparable to the compute of the other methods. Centaur had the advantage the r was itself hyperparam-optimized with a cost that is not budgeted on the main graph. Centaur cheated by getting free compute under the table.

B) it's not even clear that the advantage of choosing r=0.3 is real and not noise. If you look at Figure 11, it's not clear that the stuff in between 0.1 and 0.5 isn't noise. It could well be noise. And if you believe the variation is noise and fit a line or a parabola to smooth out the noise, you'd conclude that the optimal is don't use an LLM, so it's not clear that the LLM contribution is even positive.

C) another reason why the LLM contribution doesn't look positive: again on Figure 11, how do you explain that r=0.8 is horrible? If the LLM is principled in some way, if it can reason through "I see such and such therefore I try such and such" then asking it more would mean that it can experiment more and exclude bad regions faster. And if there's no input for it to give, it could just accept "I'll use the optimizer's suggestion this time" over and over. Hybrid should always be strictly better than just classic, but in reality this is more false the larger the r.

Overall, I don't think the conclusion follows from the paper. However, as humans the idea that "reasoning + classic HPO should be classic HPO" is very appealing. I also like the idea of exposing the opimizer internals to the LLM.

This isn't the case in most areas. For example in Law, where everythign is text, you can RL so that an LLM produces an argument which a human would believe to be more reasonable, but you can't get a really fast loop of: make an argument, test it in front of a judge, refine the argument until you win the case.