Techniques

How I am doing this is first I think a lot before writing anything down.

Examples of artefacts produced by this procedure:

• The JIT Brain Hypothesis
• Cut out the paper from the scratchpad and glue it in.
• I did put in a lot of thought per character.
• I write down thoughs only when they reach a certain level of clarity. That doesn't mean that what I write down is right. It's all about solving the problem.
• I say out loud what I want to write down next.
  • Often you notice holes in your understanding by trying to saying it out loud, that you wouldn't notice when only thinking in your head.
  • Often that makes me realize that it's not important. Therefore I can cut while generating.
  • Saying it out loud usually makes the text that ends up on paper better, because I change it while writing.
  • Sometimes you may say things out loud multiple times. First you figure something out at all, then say it again optimizing for compunicating it well. And then finally when writing it down you can still optimize. At each step
  • When saying something out loud
• The goal is to figure out a problem. This entire processes purpose it for helping you with this.
• Writing with non-erasible pen forces you to just keep on going.
• It get's you away from your computer

Try this: When you don't feel like planning (though really this is useful in general) ask your self what you actually want to do. That is actually think about it for 1-2 minutes (ideally set a timer for 1 minute minimum). The goal it to use your brain power to compute what you actually want. You will probably realize that, e.g. watching youtube, is not actually what you want when you think about it. That's only what you want if you follow your emotions blindly.

DON'T ask yourself "What is the best thing to do". You need to ask "What do I want". In my experience thinking about what I actually want is really easy. It's not imposing at all. But asking what is best does.

For this exercise I find it very helpful to write down a list of things that I migh want to do. This is not a todo list. I didn't fail if I don't complete these items. It's just a thinking aid, that helps me to consider many things, making it much more likely that I actually even consider the thing that actually want.

If you try it let me know if it worked.

Here is a much more detailed explanation from my notebook. Also see the example below.

John Wentworth said that if you can get to know what an expert keeps track of in their head, then that's a very big step towards becomming an expert yourself.

When learning from expert demonstration, you want to constantly have the expert describe what they are keeping track of in their head. Either tell them to do this, and querry them constantly.

E.g. when programming keep track of the runtimes of individual things such that you can notice when you call something linear in n, in a for loop n times. Then the runtimes becomes O(n²).

One general pattern seems to be, that if you care about the final result having certain properties, you want to pull in the evalution of the properties as much into the creation process as possible. That allows you to prune the creation process, minimizing wasted work.

Sometimes I am unsure what to say somebody. Maybe saying something makes them uncomfortable. Maybe it sends them down a bad path. E.g. maybe I am unsure about whether to tell somebody that they should work on alignment. Maybe they can't do it, and so it would be a bad experience for them, and they would waste their time. So should I convince them anyway of it if I think alignment is important.

Convincing is the wrong frame. Often what you actually want is for another person to simply understand reality. If reality was such that them working on alignment is good, then them understanding reality would likely make them want to work on alignment.

It's good default strategy to just take actions that make people understand reality. Then infofar as incentives are aligned (and they often are in that nobody wants a misaligned AGI), they will then do the action that you would like them to do.

(Of cause exfohazards need to be taken into account.)

Doing this is also a lot easier than coming up with some elaborate scheme of what you are going to say such that you manipulate the other person into doing exactly what you want them to do. You don't need to keep track of your elaborate schemes and lies. And when it comes to something like AI policy, such a strategy is very brittle. Unless you can "take over" the role of the policymaker and now control all important decisions that somebody will make, you need to actually give them a good model of reality that allows them to compute the right decisions themselfs.

Of cause sometimes it's good to lie, like when the Jihadist asks you if you are Muslem.

In my experience, how good/useful a research discussion is depends a lot on how well you can explain "the thing" (though explaining is easier in conversation).

Where it can be something you don't understand but try to understand through attempting to explain the thing you don't yet know. Produces multiple incoherent hard to understand explanations, before getting to something that makes sense, but a very good way to make progress in my experience.

I wrote many drafts, and few public posts. Public posts seemed to give better feedback for getting better at writing. Seeing people confused, and misunderstanding thing I thought where perfectly explained is the equivalent of building your perfect rocket design and have it on launch explode. That is, it's a reality check.

Are you too bad at programming to achive your objectives? Aim to become the best programmer ever! You may early terminate before being the best, but aiming to become the best is often the best way to make progress quickly, and identify what's actually important. If you don't aim to be the best you'll like generate inferior plans, that won't even make you good (or at minimum take much longer than neccesary to implement).

Inspired by How to do Great Work.

The best time to take a walk might be just after you have developed an idea, and the next step is to implement it one level deeper.

E.g. I did once design an advanced cannon for from the depths on paper. After doing that I first took a short walk before building the design in from the depths. This made a lot of relevant ideas bubble up in my mind.