I've been reading Mark of the Fool and one thing I've been thinking about is MC's power and failure's overdetermination. Alex has the Mark of the Fool, a divine inscription that prevents him from being useful in combat but allows him to learn new things quickly. The way he learns is that whenever he has to do something, the mark flashes him memories of him doing that thing before, focusing on the times or portions of it that he did it correctly. So, if he wants to learn how to run properly, whenever he's running the mark will flash him memories of him running in the past, but only his correct moves, and in no time he will learn proper running posture. Same for writing, learning new languages, detecting people lying, etc. It's a very overpowered ability.

What got me thinking was this passage, in book 3, page 403:

"All the predictions in the world don't matter in battle," Tyris continued. "In the end, the only thing that matters for victory is victory itself. And the only ones people want to hear from afterward are the victors."

"I dunno about that, sometimes we learn more from failure than from victory."

Alex would say this because every time he fails at something, he gains more data the mark can use to make him fail less in the future, that's how he learns so fast. But he doesn't learn from failure necessarily, he learns from just doing and giving the mark anything to draw from. If he tries to learn an ability, say, lockpicking, but he has never done anything near lockpicking at all in his past, the mark doesn't activate, it has no memories to draw from.

I have said in the past that one can learn way more from success than from failure, and that, in fact, focusing on learning from failure instead of success is a losing strategy. From Son of a Serpent:

There's this trend among indiedevs who are more business savvy of looking at games that failed on Steam to try and gain knowledge from them. All such endeavors are pointless and, as Eric would say, are examples of people hexing themselves with negative thoughts.

Failure is overdetermined. Most things fail for multiple reasons. So when a game doesn't do well, you can't really learn anything about it because you will think it failed for reason 1, but it actually failed for reasons 1 through 5. And then another game you look at might fail for reasons 2, 3, etc. You never actually know. If you focus on looking at successes instead your ratio of looking at what works, what is actual signal, versus what is noise, is much higher, and over the long run that will lead you to better decisions.

Not only that, it's easy to look at a game that failed and go "yea it failed because of the graphics" or "it failed because they didn't do enough marketing", which is just... This is a noobtrap. If you find yourself doing this in your head to any game understand that it's a thought pattern that you must stop, it won't lead anywhere good.

Let's say that you extract 5 reasons for why a game failed/succeeded after looking at it, and you do this for every game you look at. You don't know if those reasons are valid because you can never run this experiment again while changing only 1 variable, so in a sense you're always blind.

In this state of blindness, your hit rate when picking those reasons from successful games will be higher, because due to failure's overdetermination those failed game reasons will more often than not be wrong or incomplete, whereas the reasons from successful games will more often be right, as success often only takes a few right things to get going despite the bad ones. (See Oskar's tweet above about how you only need one or two good things in a game and the rest needs to only be good enough)

Let's use Vampire Survivors as an example. If Vampire Survivors failed you could extract many reasons out of it on why it failed, looks like trash, repetitive gameplay, etc, etc. But it succeeded, and once something succeeds it's a signal from God/the universe that despite all the obvious bad things about it, there are some specific things about it that just work. And in this case those few specific things were copied by lots of people to varying levels of success, but it's to some extent clear what they are.

Now multiply this effect by lots of games over a long period of observation and you'll have a much more useful bundle of facts on what you should do that come from successful games, compared to if you only looked at unsuccessful games, where you'll be full of facts like "your game shouldn't look like trash"... well, obviously, but that's not very helpful.

If you're in a low/high risk environment and your goal is surviving, a map of which things to avoid is useful because it will keep you safe. But if you're in a high risk environment (gamedev) and your goal is thriving and flourishing, you don't want to only avoid damage and survive, you want to go hardcore and take risks because the upside is infinite, and in that case a map of avoidance - things you shouldn't do, taken from unsuccessful games - is less useful than a map of action - things you should do, taken from successful games.

I still believe every word I said here is true. In fact, after SNKRX succeeded, I thought the main reason was due to the Auto Chess formula, but it was only after Vampire Survivors released a while later that it became clear it was due to the auto-shooting instead. What this means is that even in success, it's hard for a project's own creator to understand why it succeeded! Imagine trying to get real insights from someone else's failures, it makes no sense at all.

However... Alex's mark is kind of like an LLM, or maybe like the Whispering Earring. Hm, kind of a mix of both in some sense. I do wonder if the author will go more in the earring direction as the story progresses, probably not, but it would be an interesting darker turn. Well, even if he doesn't, said darker turn is in a way already realized in Don't Make Me Think so it's fine.

In any case, one of the things LLMs are really good at is finding good solutions when you give them a problem with lots of constraints and they have a lot of data to go from. Like, they're extremely good at that specifically. Most people probably don't have such problems to solve, but if you ever do, it becomes kind of undeniable to grant that these tools are actually super intelligent. I think the fact that most people don't have such problems is primarily why so many of them seem to think AI is a bubble. Well, it may be, but it's like the realest bubble there ever was. But yea, this property of LLMs is kind of exactly how the mark works, and what does Alex conclude after living with the mark for a while? That you often (actually he said sometimes) learn more from failure than from victory.

While the success argument above is still true, LLMs existing makes the learning from failure path more viable. But it's important to remember, just like with Alex, it's not learning from failure itself, it's learning from doing! The more data and constraints the robot has, the better it will do. So actually what you want is not to fail, but to just do various things to give it more examples and more constraints, exactly how Alex does it. Once you have everything, it will find a fortuitous path, as that's what it was built to do.

Reviewing this argument with Claude, he had two additional interesting insights, or perhaps reframings of the same insight that seemed good to me:

The original "learn from success, not failure" advice is a heuristic compensating for a specific human limitation (we are bad pattern-extractors over noisy corpora) by selecting a corpus with higher signal density. Once you have a better extractor — Mark, LLM, or a sophisticated domain-specific pattern matcher built from years of experience — the success/failure axis becomes secondary to the doing/not-doing axis.

It also implicitly proposes a third position that's distinct from both "learn from success" and "learn from failure": learn from variety. Bias toward expanding the corpus, not toward replicating winners or avoiding losers. That's actually a different optimization target and I think it's the one your argument is reaching for.

"Learn from variety" is aligned with what I mentioned earlier regarding references being more important than taste. Because you want to give the LLM examples and constraints, it's best to give it as many references as possible. So "doing" can also actually be "knowing when/how to reference what has been done." While the mark only draws from Alex's memories, the LLM can draw from humanity's memories, guided by you, the human.