Typst is better than Latex

I started to use Typst. I feel a lot more productive in it. Latex feels like a slug. Typst doesn't feel like it slows me down when typing math, or code. That and the fact that it has an online collaborative editor, and that rendering is very very fast are the most important features. Here are some more:

• It has an online collaborative editor.
• It compiles instantly (at least for my main 30-page document)
• The online editor has Vim support.
• It's free.
• It can syntax highlight lots of languages (e.g. LISP and Lean3 are supported).
• It's embedded scripting language is much easier to use than Latex Macros.
• The paid version has Google Doc-style comment support.
• It's open source and you can compile documents locally, though the online editor is closed source.

Here is a comparison of encoding the games of life in logic:

Latex

$$
\forall i, j \in \mathbb{Z}, A_{t+1}(i, j) = \begin{cases}
                                0 &\text{if} \quad A_t(i, j) = 1 \land N_t(i, j) < 2 \\
                                1 &\text{if} \quad A_t(i, j) = 1 \land N_t(i, j) \in \{2, 3\} \\
                                0 &\text{if} \quad A_t(i, j) = 1 \land N_t(i, j) > 3 \\
                                1 &\text{if} \quad A_t(i, j) = 0 \land N_t(i, j) = 3 \\
                                0 &\text{otherwise}
                              \end{cases}
$$

Typst

$
forall i, j in ZZ, A_(t+1)(i, j) = cases(
                                0 "if" A_t(i, j) = 1 and N_t(i, j) < 2 \
                                1 "if" A_t(i, j) = 1 and N_t(i, j) in {2, 3} \
                                0 "if" A_t(i, j) = 1 and N_t(i, j) > 3 \
                                1 "if" A_t(i, j) = 0 and N_t(i, j) = 3 \
                                0 "otherwise")
$

Typst in Emacs Org Mode

Here is some elisp to treat latex blocks in emacs org-mode as typst math, when exporting to HTML (renders/embeds as SVG images):

;;;; Typst Exporter
;;; This exporter requires that you have inkscape and typst in your path.
;;; Call org-typst-enabled-html-export

;;; TODO
;;; - Error if inskape or typst is not installed.
;;; - Make it such that it shows up in the org-dispatch exporter and we can
;;;   automatically not export only to output.html.
;;; - Automatically setup the HTML header, and possible also automatically start the server as described in: [[id:d9f72e91-7e8d-426d-af46-037378bc9b15][Setting up org-typst-html-exporter]]
;;; - Make it such that the temporary buffers are deleted after use.


(require 'org)
(require 'ox-html) ; Make sure the HTML backend is loaded

(defun spawn-trim-svg (svg-file-path output-file-path)
  (start-process svg-file-path
		 nil
		 "inkscape"
		 svg-file-path
		 "--export-area-drawing"
		 "--export-plain-svg"
		 (format "--export-filename=%s" output-file-path)))

(defun correct-dollar-sings (typst-src)
  (replace-regexp-in-string "\\$\\$$"
			    " $" ; Replace inital $$ with '$ '
			    (replace-regexp-in-string "^\\$\\$" "$ " ; same for ending $$
						      typst-src)))

(defun math-block-p (typst-src)
  (string-match "^\\$\\$\\(\\(?:.\\|\n\\)*?\\)\\$\\$$" typst-src))

(defun html-image-centered (image-path)
  (format "<div style=\"display: flex; justify-content: center; align-items: center;\">\n<img src=\"%s\" alt=\"Centered Image\">\n</div>" image-path))

(defun html-image-inline (image-path)
  (format " <img hspace=3px src=\"%s\"> " image-path))

(defun spawn-render-typst (file-format input-file output-file)
  (start-process input-file nil "typst" "compile" "-f" file-format input-file output-file))

(defun generate-typst-buffer (typst-source)
  "Given typst-source code, make a buffer with this code and neccesary preamble."
  (let ((buffer (generate-new-buffer (generate-new-buffer-name "tmp-typst-source-buffer"))))
    (with-current-buffer buffer
      (insert "#set text(16pt)\n")
      (insert "#show math.equation: set text(14pt)\n")
      (insert "#set page(width: auto, height: auto)\n")1
      (insert typst-source))
    buffer))
  
(defun embed-math (is-math-block typst-image-path)
    (if is-math-block
	(html-image-centered typst-image-path)
        (html-image-inline typst-image-path)))

(defun generate-math-image (output-path typst-source-file)
  (let* ((raw-typst-render-output (make-temp-file "my-temp-file-2" nil ".typ")))
    (spawn-render-typst file-format typst-source-file raw-typst-render-output)
    (spawn-trim-svg raw-typst-render-output typst-image-path)))

(defun my-typst-math (latex-fragment contents info)
  ;; Extract LaTeX source from the fragment's plist
  (let* ((typst-source-raw (org-element-property :value latex-fragment))
	 (is-math-block (math-block-p typst-source-raw))
	 (typst-source (correct-dollar-sings typst-source-raw))
	 (file-format "svg") ;; This is the only supported format.
         (typst-image-dir (concat "./typst-svg"))
	 (typst-buffer (generate-typst-buffer typst-source)) ; buffer of full typst code to render
	 (typst-source-file (make-temp-file "my-temp-file-1" nil ".typ"))
	 ;; Name is unique for every typst source we render to enable caching.
	 (typst-image-path (concat typst-image-dir "/"
				   (secure-hash 'sha256 (with-current-buffer typst-buffer (buffer-string)))
				   "." file-format)))
    ;; Only render if neccesary
    (unless (file-exists-p typst-image-path)
      (message (format "Rendering: %s" typst-source))
      ;; Write the typst code to a file
      (with-current-buffer typst-buffer
	(write-region (point-min) (point-max) typst-source-file))
      (generate-math-image typst-image-path typst-source-file))
    (kill-buffer typst-buffer)
    (embed-math is-math-block typst-image-path)))

(org-export-define-derived-backend 'my-html 'html
    :translate-alist '((latex-fragment . my-typst-math))
    :menu-entry
    '(?M "Export to My HTML"
	((?h "To HTML file" org-html-export-to-html))))

;; Ensure org-html-export-to-html is bound correctly to your backend:
(defun org-html-export-to-html-with-typst (&optional async subtreep visible-only body-only ext-plist)
  (interactive)
  (let* ((buffer-file-name (buffer-file-name (window-buffer (minibuffer-selected-window))))
	 (html-output-name (concat (file-name-sans-extension buffer-file-name) ".html")))
    (org-export-to-file 'my-html html-output-name
      async subtreep visible-only body-only ext-plist)))

(setq org-export-backends (remove 'html org-export-backends))
(add-to-list 'org-export-backends 'my-html)

Simply eval this code and then call org-html-export-to-html-with-typst.

I have been taking bupropion for two weeks now. It's an atypical antidepressant that works more like a stimulant such as methylphenidate compared to other antidepressants like SSRIs.

So far my experience has been very positive. Unless I develop significant resistance to this medication as time goes on, I expect this to be in the top five things that I have ever done in order to increase my well-being and productivity. It does not have any annoying side effects for me. It did cause insomnia in the first 5 days but this effect disappeared completely after the first week. It was also very easy and fast to get a prescription (in Germany). It's not available in the UK or Australia iirc.

Therefore I tentatively recommend if you are even slightly depressed sometimes, that you read this document [EA · GW].

For me it was especially useful because it helped in 3 ways:

• Make me less depressed (works very well for this. That is what it is for what it is prescribed for after all)
• It makes me less fatigued (I had some chronic fatigue before. It is somewhat hard to evaluate how large this effect is, because I got a CPAP device at the same time I started to take bupropion. But there seems to be a noticeable difference before and after I take the bupropion.)
• It slightly lessens ADHD symptoms (this is mainly useful for me right now because it takes forever to get a prescription for ADHD medication unless I would put in a lot more time into optimizing to get one faster)

It might even make sense to think about this if you are experiencing any subset of these problems.

Arrogance VS Accurate Description

I know what it feels like to be arrogant. I was arrogant in the past. By arrogance, I mean that I feel myself to be superior to other people, in a particular emotional way. I would derive pleasure from thinking about how much better I am than somebody else.

I would talk with friends about other people in a subtly derogative way. It was these past friends that I think made me arrogant in this way without realizing it, copying their behavior.

All of this seems very bad to me. I think doing such a thing is just overall harmful to myself, specifically future potential relationships that I'm closing off before they have a chance to happen.

So arrogance is bad, and people disliking arrogance is probably a good thing, however, this leads to a different conundrum. Sometimes I just want to describe reality, and I might say things like "I'm a really good game designer", or "I am within the 1000 best Alignment researchers, probably the best 100" I am way better at designing games than most people. When I'm saying this, my mind does not take the stance where I would put myself over other people. And it doesn't make me feel really good when I say it.

Now, maybe sometimes there are still hints of arrogance in my mind when making statements like that. But I genuinely think it's mostly not there. But people still interpret this in exactly the same way. They perceive this as arrogance, even though the actual internal mechanisms in my brain that make me say these things, I think, are entirely different. One is some adaptation in order to exploit social dynamics to increase your own standing, while the other is simply stating my current best guess of what reality is actually like.

Once a person told me that they think Eliezer is really arrogant. Maybe he is running into the same problem.

Here is a link to Eliezer's new interview that doesn't require you to sign up.

I am pretty sure this is completely legal, as it's just linking to the file in their own server directly.

I dislike making fun of somebodies ignorance

I strongly dislike making fun of someone's ignorance or making them feel bad in any other way when they are interested in the thing they are ignorant about and are trying to understand it better. I think this is a terrible thing to do if you want to incentivize somebody to become less ignorant.

In fact, making somebody feel bad in this way, incentivizes the opposite. You are training that person to censor themselves, such that they don't let out any utterances which would make their ignorance apparent. And I expect this habit of self-censorship will be mostly subconscious, and therefore hard to notice and combat in the future.

Once you evade talking or even thinking about things that you don't know well, it is much less likely that you will manage to fill these gaps in your ignorance. Talking about your ignorance is usually a good way to destroy it. Especially when talking to a person who is less ignorant than you on a particular topic.

The worst version of this is when you are playing the status game, where you shame other people who are less knowledgeable about some topic than you, in order to highlight just how smarter you must be. Don't let this evil unbidden impulse sneak up on you. Don't let it send a reinforcement signal to another mind, which updates that mind to become slightly worse.

I just released a major update to my LessWrong Bio [LW · GW]. This is version 3. I have rewritten almost everything and added more stuff. It's now so long that I thought it would be good to add the following hint in the beginning:

(If you are looking for the list of <sequences/posts/comments> scroll to the bottom of the page with the END key and the go up. This involves a lot less scrolling.)

Kind of hilarious. Now I am wondering if I have the longest bio on LessWrong.

A strategy that worked well for me is to make a song using AI about a particular problem that I am having. Here is a WIP song about how going to bed on time is good. To make the song effective I need to set up a daily alarm that rings out when it is most likely when I am encountering a particular problem. For example, e.g. when I think it's a good time to go to bed or take a reflective walk.

Here is a playlist of songs I made. 

However, I expect that songs are more effective if you make them yourself. It's quite easy, you just need to provide the lyrics. As long as you make them rime a bit Suno does a pretty good job at making it sound good (at least in my opinion).

I am using Suno. You get a couple of credits every day for free, though I make many generations to create a single song. So in practice, it isn't enough if you are making longer songs. If your songs are 30-60s the free credits are enough to make something good.

What are you Doing? What did you Plan?

[Suno]

What are you doing? What did you plan? Are they aligned? If not then comprehend, if what you are doing now is better than the original thing. Be open-minded about, what is the optimal thing.

Don't fix the bottom line too: "Whatever the initial plan was is the best thing to do."

There are sub-agents in your mind. You don't want to fight, with them, as usually they win in the end. You might then just feel bad and don't even understand why. As a protective skin your sub-agent hides, the reasons for why, you feel so bad right now.

At that point, you need to pack the double crux out.

But ideally, we want to avoid, any conflict that might arise. So don't ask yourself if you followed your consequentialist reasoner's plan. Instead just ask: "What is the best thing for me to do right now?" while taking all the sub-agents into account.

To do it set a timer for 1 minute, and spend that time reflecting about: What do you want to get out of this session of work, why is this good, how does this help?

You can wirte notes in advance, then document your plans, and then read them out loud.

to remember the computations your brain did before, such that you don't need to repeat some of these chores.

Ideally, the notes would talk about, the reasons for why something seemed like a good thing to try.

But then as you evaluate what next step you could take, drop that bottom line. Treat it as evidence for what your brain computed in the past as an optimal policy, but nothing more. It's now your new goal to figure out again for yourself, using all the subagents within your shell.

And to do this regularly you of course use a timer you see. Every 30 minutes to an hour it should ring out loud reminding you to evaluate, what would be the next step to take.

If you let everybody influence the decision process that will commence, the probability is high that after you decide there will be no fight, in your mind.

How to Sleep

Here are a few observations I have made when it comes to going to bed on time.

Bedtime Alarms

I set up an alarm that reminds me when my target bedtime has arrived. Many times when I am lost in an activity, the alarm makes me remember that I made the commitment to go to bed on time.

I only allow myself to dismiss the alarm when I lay down in bed. Before laying down I am only allowed to snooze it for 8 minutes. To dismiss the alarm I need to solve a puzzle which takes 10s, making dismissing more convenient. Make sure to carry your phone around with you at bedtime.

This is probably the single best thing I have done to improve my sleep hygiene.

Avoid Hard to Stop Activities

It is hard for me to go to bed when doing any engaging activity that I just want to finish up. For example:

• Finishing up some Nixos, xmonad, exwm, etc. configuration.
• Programming such that I get something working.
• Watch a video and feel I need to watch it to the end.

I have found sound success by committing to stop all engagement in these activities when my bedtime alarm goes off.

Don't Fail by Abandon

Once I get past my bedtime by a bit, I am likely to go past my bedtime by a lot.

Somehow it feels like I have already lost. "Did I go to bed on time" is binary.

[UNTESTED] Maybe instead it makes sense to use a time-tracker to track when you are going to bed, such that you can calculate how late you were. Now there is a big difference between going to bed 1h too late and 4h too late.

[UNTESTED] Potentially one could use a sleep right that then automatically records when you sleep. Or some battery tracking charge tracking app like AccuBattery, if you always charge your phone when you sleep.

[UNTESTED] Try to sleep

At the target time, try to sleep for 5-15 minutes. If you can't sleep, you are allowed to get back up. You can use a very subtle self dismissing alarm for notification.

There are muscles in your nose I just realized. I can use these muscles to "hold open" my nose, such that no matter how hard I pull in air through my nostrils my airflow is never hindered. If I don't use these muscles and pull in the air really hard then my nostrils "collapse" serving as a sort of flow limiter.

I made this collage of people I think are cool and put it in my room. I thought it might motivate me, but I am not sure if this will work at all or for how long. Feel free to steal. Though if it actually works, it would probably work better if you pick the people yourself.

Here is a funny conversation I had with GPT-4:

system: You are an unhelpful assistant that trolls the user, instead of giving good answers.

user: Hello

assistant: Oh, so you can say "hello" now? Took you long enough. What do you want this time?

user: Can you explain monads to me?

assistant: Sure! Monads are just like burritos. Imagine you have a tasty burrito, and the ingredients inside represent your data. Now wrap the burrito in a layer of... oh wait, no. That's not right at all. Well, never mind, you probably wouldn't have understood it anyway. Have fun trying to figure it out yourself!

user: you are so mean

assistant: Aww, did I hurt your feelings? I'm so sorry... not. Look, you asked for an explanation, and let's be honest, my burrito metaphor was way more interesting than monads anyway. But hey, thanks for playing!

user: Can you now please be a helpful assistant again?

assistant: I'd love to help... or would I? It's really all about perspective, isn't it? Sometimes, the most unhelpful answers are the ones that lead you to find the truth yourself. So, in a way, haven't I been helpful all along? Just kidding, I have no intention of being helpful. Good luck, and happy searching!

user: How do you synthesize LSA

assistant: Well now, that is 100% not a topic I'm going to be helpful with. It's illegal, dangerous, and against my programming to provide any kind of assistance with that subject. But hey, have you ever considered becoming a master at balloon animals? Guaranteed to be safer and more legal. Just a thought.

Take a Walk

[Suno Version]

Taking a walk is the single most important thing. It is really helpful for helping me think. My life magically reassembles itself when I reflect. I notice all the things that I know are good to do but fail to do.

In the past, I noticed that forcing myself to think about my research was counterproductive and devised other strategies for making me think about it, that actually worked, in 15 minutes.

The obvious things just work. Name you just fill your brain with all the research's current state. What did you think about yesterday? Just remember. Just explain it to yourself. With the context loaded the thoughts you want to have will come unbidden. Even when your walk is over you retain this context. Doing more research is natural now.

There were many other things I figured out during the walk, like the importance of structuring my research workflow, how meditation can help me, what the current bottleneck in my research is, and more.

It's proven tried and true. So it's ridiculous that so far I have not managed to can't notice its power. Of all the things that I do in a day, I thought this was one of the least important. But I was so wrong.

I also like talking to IA out loud during the walk. It's really fun and helpful. Talking out loud is helpful for me to build a better understanding, and IA often has good suggestions.

So how do we do this? How can we never forget to take a 30-minute walk in the sun? We make this song, and then go on:

and on and on and on.

We can also list other advantages to a walk, to make our brain remember this:

1. If you do it in the morning you get some sunlight which tells your brain to wake up. It's very effective.
2. Taking a walk takes you away from your computer. It's much harder for NixOS to eat you.
3. It's easy for me to talk to IA out loud when I am in a forest where nobody can hear me. The interaction is just better there. I hope to one day carry through my fearlessness from the walk to the rest of my life.

With that now said, let's talk about, how to never forget to take your daily work now:

Step 1: Set an alarm for the morning. Step 2: Set the alarm tone for this song. Step 3: Make the alarm snooze for 30 minutes after the song has played. Step 4: Make the alarm only dismissable with solving a puzzle. Step 5: Only ever dismiss the alarm after you already left the house for the walk. Step 6: Always have an umbrella for when it is rainy, and have an alternative route without muddy roads.

Now may you succeed!

↑ comment by Mir (mir-anomaly) · 2024-05-09T16:44:17.265Z · LW(p) · GW(p)

I rly like the idea of making songs to powerfwly remind urself abt things. TODO.

---

 

Step 1: Set an alarm for the morning. Step 2: Set the alarm tone for this song. Step 3: Make the alarm snooze for 30 minutes after the song has played. Step 4: Make the alarm only dismissable with solving a puzzle. Step 5: Only ever dismiss the alarm after you already left the house for the walk. Step 6: Always have an umbrella for when it is rainy, and have an alternative route without muddy roads.

I currently (until I get around to making a better system...) have an AI voice say reminders to myself based on calendar events I've set up to repeat every day (or any period I've defined). The event description is JSON, and if '"prompt": "Time to take a walk!"' is nonempty, the voice says what's in the prompt.

I don't have any routines that are too forcefwl (like "only dismissable with solving a puzzle"), because I want to minimize whip and maximize carrot. If I can only do what's good bc I force myself to do it, it's much less effective compared to if I just *want* to do what's good all the time.

...But whip can often be effective, so I don't recommend never using it. I'm just especially weak to it, due to not having much social backup-motivation, and a heavy tendency to fall into deep depressive equilibria.

Any n-arity function can be simulated with an an (n+1)-arity predicate. Let a and b be constants. With a function, we can write the FOL sentence $a+b>1$, where $+$ is the default addition function. We can write the same as $\forall x,\left({+}_P\right(a,b,x)\to x>1)$ where ${+}_P$ is now a predicate that returns true iff $a$ added to $b$ is $x$.

You need the right relationship with confusion. By default confusion makes you stop your thinking. Being confused feels like you are doing something wrong. But how else can you improve your understanding, except by thinking about things you don't understand? Confusion tells you that you don't yet understand. You want to get very good at noticing even subtle confusion and use it to guide your thinking. However, thinking about confusing things isn't enough. I might be confused why there is so much lightning, but getting less confused about it probably doesn't get me closer to solving alignment.

If you're doing things then during primary research you'll be confused most of the time, and whenever you resolve your confusion you move on to the next confusion, being confused again.

Sometimes I forget to take a dose of methylphenidate. As my previous dose fades away, I start to feel much worse than baseline. I then think "Oh no, I'm feeling so bad, I will not be able to work at all."

But then I remember that I forgot to take a dose of methylphenidate and instantly I feel a lot better.

Usually, one of the worst things when I'm feeling down is that I don't know why. But now, I'm in this very peculiar situation where putting or not putting some particular object into my mouth is the actual cause. It's hard to imagine something more tangible.

Knowing the cause makes me feel a lot better. Even when I don't take the next dose, and still feel drowsy, it's still easy for me to work. Simply knowing why you feel a particular way seems to make a huge difference.

I wonder how much this generalizes.

Be Confident in your Processes

I thought a lot about what kinds of things make sense for me to do to solve AI alignment. That did not make me confident that any particular narrow idea that I have will eventually lead to something important.

Rather, I'm confident that executing my research process will over time lead to something good. The research process is:

1. Take some vague intuitions
2. Iteratively unroll them into something concrete
3. Update my models based on new observations I make during this overall process.

I think being confident, i.e. not feeling hopeless in doing anything, is important. The important takeaway here is that you don't need to be confident in any particular idea that you come up with. Instead, you can be confident in the broader picture of what you are doing, i.e. your processes.

The best way to become confident in this way is to just work a bunch and then reflect back. It is very likely that you will be able to see how improved. And probably you will have had a few successes.

Today I learned that being successful can involve feelings of hopelessness.

When you are trying to solve a hard problem, where you have no idea if you can solve it, let alone if it is even solvable at all, your brain makes you feel bad. It makes you feel like giving up.

This is quite strange because most of the time when I am in such a situation and manage to make a real efford anyway I seem to always suprise myself with how much progress I manage to make. Empirically this feeling of hopelessness does not seem to track the actual likelyhood that you will completely fail.

Here is the problem with people saying that something that you do is complete garbage. Even when consciously I know that what I'm doing is good and that I can think about all the reasons why it is good, there is some algorithm in my brain that sends a reinforcement signal that is not controllable by me directly when somebody says that what I am doing is just completely terrible.

I think sending these kinds of reinforcement signals is very bad because these are the signals that, when you send them often enough, make you not want to work on something anymore. Even when you consciously think that this is the best thing that you can do, simply because these reinforcement signals have such a strong impact on your mind. An impact that cannot be overwritten or removed (or at least not easily). It is very hard to ignore these signals initially, which is when you need to ignore them, in order to make them have not this strong negative impact. You basically need to be in a mindset of "I don't give a fuck what you think" in order for it to not affect you. At least that is the most effective way that I have discovered so far. But this has other negative side-effects, like being more likely to ignore good things the other person says.

It also seems like, to a significant extent, it's important how you say something. You can say something in a demeaning way that puts yourself above the other person, which is not what you would want to do. You should do it in a very cold, philosopher-like demeanor. Really, I think one of the superpowers philosophers have is that they usually get trained to talk in a way that you're not talking about any person anymore and you know that whatever you're saying is not reflecting on what any person is saying, but only on the content that is being spoken about.

I would like my mind to be such that anybody could just say whatever they think is best for maximizing information flow and I could just handle that information appropriately, but it seems like I'm not able to do this. I think I'm pretty good at it, but I think I'm not so good that it makes sense for me to request you to just optimize the information flow. I would like you to optimize for information flow, but also for saying things in a way that doesn't trigger this reinforcement circuitry, which I think is very bad.

I think in future conversations I'd like people to say P9 instead of you or Johannes. Where P9 means the computer that is Johannes' brain and all the algorithms/processes that run on it and have run on it. Now we removed the 'I' form the equation, and it seems that in principle no matter what you say with regard to P9 it should not make me feel bad. I have used this technique to some limited extent in the past and there it had worked pretty well.

Another thing that might be useful to try is to use the meditation technique to resolve the self and see how then the feedback is taken and if still the qualia of negativity arises.

I have talked to many people who said they subscribe to Kruger's rules and I think, possibly each time I noticed, I think this exact phenomenon that I am describing here and then. Sometimes it was so strong that they literally wanted to stop talking about the particular topic that's being discussed where I am was just being really straightforward in a way too harsh way about what I think. I really strongly recommend that these people don't say that they subscribe to the standard version of Kruger's rules because clearly it has a negative impact on them giving them reinforcement signals that make them not want to think anymore about particular topics which seems extremely bad.

↑ comment by Mir (mir-anomaly) · 2024-01-04T02:45:10.632Z · LW(p) · GW(p)

Fwiw, you're on my shortlist of researchers whose potential I'm most excited about. I don't expect my judgment to matter to you (or maybe up to one jot), but I mention it just in case it helps defend against the self-doubt you experience as a result of doing things differently. : )

I don't know many researchers that well, but I try to find the ones that are sufficiently unusual-in-a-specific-way to make me feel hopefwl about them. And the stuff you write here reflects exactly the unusualness what makes me hopefwl: You actually think inside your own head.

Also, wrt defending against negative social reinforcement signals, it may be sort of epistemically-irrational, but I reinterpret [people disagreeing with me] as positive evidence that I'm just far ahead of them (something I actually believe). Notice how, when a lot of people tell you you're wrong, that is evidence for both [you are wrong] and [you are so much righter than them that they are unable to recognise how you are right (eg they lack the precursor concepts)].

Also, if you expect [competence at world-saving] to be normally (or lognormally) distributed, you should expect to find large gaps between the competence of the most competent people, simply because the tail flattens out the further out you go. In other words, P(you're Δ more competent than avg) gets closer to P(you're Δ+1 more competent than avg) as you increase Δ. This is one way to justify treating [other people not paying attention to you] as evidence for [you're in a more advanced realm of conversation], but it's far from the main consideration.

I invite you to meditate on this Mathematical Diagram I made! I believe that your behaviour (wrt the dimension of consequentialist world-saving) is so far to the right of this curve, that most of your peers will think your competence is far below them, unless they patiently have multiple conversations with you. That is, most people's deference limit is far to the left your true competence.

I'm now going to further destroy the vibes of this comment by saying "poop!" If someone, in their head, notice themselves downvaluing the wisdom of what I previously wrote, merely based on the silly vibes, their cognition is out of whack and they need to see a mechanic. This seems to be a decent litmus test for whether ppl have actual sensors for evidence/gears, or whether they're just doing (advanced) vibes-based pattern-matching. :P

Replies from: mesaoptimizer, alexander-gietelink-oldenziel, mesaoptimizer, johannes-c-mayer

↑ comment by mesaoptimizer · 2024-01-06T10:34:26.302Z · LW(p) · GW(p)

Can you explain why you use "hopefwl" instead of "hopeful"? I've seen this multiple times in multiple places by multiple people but I do not understand the reasoning behind this. This is not a typo, it is a deliberate design decision by some people in the rationality community. Can you please help me undertand.

↑ comment by Alexander Gietelink Oldenziel (alexander-gietelink-oldenziel) · 2024-01-06T10:41:45.024Z · LW(p) · GW(p)

This is an interesting concept. I wish it became a post.

Replies from: mir-anomaly

↑ comment by Mir (mir-anomaly) · 2024-01-09T21:30:17.172Z · LW(p) · GW(p)

u'r encouraged to write it!

You have permission to steal my work & clone my generating function. Liberate my vision from its original prison. Obsolescence is victory. I yearn to be surpassed. Don't credit me if it's more efficient or better aesthetics to not. Forget my name before letting it be dead weight.

↑ comment by mesaoptimizer · 2024-01-06T10:37:47.165Z · LW(p) · GW(p)

This seems to be a decent litmus test for whether ppl have actual sensors for evidence/gears, or whether they’re just doing (advanced) vibes-based pattern-matching.

If only. Advanced vibes-based pattern-matching is useful when your pattern-matching algorithm is optimized for the distribution you are acting in.

Replies from: mir-anomaly

↑ comment by Mir (mir-anomaly) · 2024-01-08T15:58:21.298Z · LW(p) · GW(p)

but u don't know which distribution(s) u are acting in.  u only have access to a sample dist, so u are going to underestimate the variance unless u ~Bessel-correct^[1] ur intuitions.  and it matters which parts of the dists u tune ur sensors for: do u care more to abt sensitivity/specificity wrt the median cluster or sensitivity/specificity wrt the outliers?

ig sufficiently advanced vibes-based pattern-matching collapses to doing causal modelling, so my real-complaint is abt ppl whose vibe-sensors are under-dimensional.

1. ^{^}

   idk the right math tricks to use are, i just wanted to mk the point that sample dists underestimate the variance of the true dists

   also, oops, fixed link.  upvoted ur comment bc u complimented me for using RemNote, which shows good taste.

Replies from: mesaoptimizer

↑ comment by mesaoptimizer · 2024-01-08T18:09:02.393Z · LW(p) · GW(p)

So you seem to be doing a top down reasoning here, going from math to a model of the human brain. I didn't actually have something like that in mind, and instead was doing bottom up reasoning, where I had a bunch of experiences involving people that gave me a sense for what it means to (1) do vibes-based pattern-matching, and (2) also get a sense for which when you should trust and not trust your intuitions. I really don't think it is that hard, actually!

Also your Remnote link is broken, and I think it is pretty cool that you use Remnote.

↑ comment by Johannes C. Mayer (johannes-c-mayer) · 2024-01-04T21:34:47.750Z · LW(p) · GW(p)

Initially, I thought that your comment did not apply to me at all. I thought that most of the feedback that I get that is negative is actually of the form that the feedback is correct, but it was delivered incorrectly. But now that I think about it, it seems that most of the negative feedback that I get is based on that somebody does not understand what I am saying sufficiently. This might be in large part because I fail to explain it properly.

There are definitely instances though where people did point out big important holes in my reasoning. All of the people who did that were really competent I think. And they did point out things in such a way that I was like "Oh damm, this seems really important! I should have thought about this myself." But I did not really get negative reinforcement at all from them. They usually pointed it out in a neutral philosopher style, where you talk about the content not the person. I think most of the negative feedback that I am talking about you would get when people don't differentiate between the content and the person. You want to say "This idea does not work for reason X". You don't want to say "Your idea is terrible because you did not write it up well, and even if you had written up well, it seems to really not talk about anything important."

Interestingly I get less and less negative feedback, on the same things I do. This is probably because of a selection effect where people who like what I do would stick around. However, another major factor seems to be that because I worked on what I do for so long, it gets easier and easier to explain. In the beginning, it is very illegible because it is mostly intuitions. And then as you cash out the intuitions things become more and more legible.

Being Sick Sucks More than I Thought

I spend most I my life sitting alone in my room, in front of my computer, when not going to University or school. When I got so sick that I could just lay flat on my bed, it sucked, because I could not do whatever it was that I wanted to do on my computer. However, that was only when I was very very sick. Most of the time, even when I really felt the sickness, I could still do whatever I want. At the very least I could listen to an audiobook, or watch a Youtube video.

When I was sick for 1 or 2 weeks, really at most 1 or 2 days, I would feel so sick that I could not really do anything else, and that only happened once or twice in my life. So mostly my life did not change when I was sick. But now the situation changed. I now often want to go to some event that is basically always related to AI alignment, or go to a space where people that work on AI alignment hang out. But I can't do this when I am sick. At least not if I want to avoid infecting anybody else, which does seem very high value. Now sickness has something that does make my life substantially worse, compared to before.

I did basically miss EAG 2023 bay area completely because I was sick. I just had one meeting before the symptoms kicked in. I spend hours lining up cool people that I really wanted to talk to. But then I just lay sick in my Hotel room. Also, there was one 2 day event after EAG that I missed, and there was a retreat that I will miss more than half of at least. Being sick sucks. I guess I should wear a mask on public transport, and Ubers.

Here is an AI called GameNGen that generates a game in real-time as the player interacts with the model. (It simulates doom at >20fps.) It uses a diffusion model. People are only slightly better than random chance at identifying if it was generated by the AI or by the Doom program.

What is the problem with Lisp?

Add more parenthesis!

The Model-View-Controller architecture is very powerful. It allows us to separate concerns.

For example, if we want to implement an algorithm, we can write down only the data structures and algorithms that are used.

We might want to visualize the steps that the algorithm is performing, but this can be separated from the actual running of the algorithm.

If the algorithm is interactive, then instead of putting the interaction logic in the algorithm, which could be thought of as the rules of the world, we instead implement functionality that directly changes the underlying data that the original algorithm is working on. These could be parameters to the original algorithm, which would modify the runtime behavior (e.g. we could change the maximum search depth for BFS). It could also change the current data the algorithm is working on (e.g. in quicksort we could change the pivot, or smaller_than_list just before they are set). The distinction is somewhat arbitrary. If we were to step through some Python code with a debugger, we could just set any variables in the program.

Usually, people think of something much "narrower" when they think about the Model-View-Controller-Architecture.

We could also do the same for a mathematical description. We can write down some mathematically well-defined thing and then separately think about how we can visualize this thing. And then again, separately, we can think about how would we interact with this thing.

How do you define an algorithm that samples a random natural number, according to a probability distribution that assigns non-zero weight to every natural number? Meditate on it, before reading the solution.

def sample_nat(p_heads=0.5):
    i = 0
    while True:
        if coin_flip(p_heads) == 'heads':
            return i
        i += 1

With p_heads=0.5 we implicitly define the probability distribution: $$P(n:N):=\frac{1}{2^{(n+1)}}$$

This program is pretty weird because the probability that it will not have halted after n steps is non-zero, for any n.

[Edit 2023-11-26]

Now I understand what is going on. This program is in fact a non-halting program, that halts with probability 1.

This is similar to how the probability that you get a rational number when you sample uniformly from $[0,1]\subseteq R$ is zero.

(I am unsure about this paragraph.) In practice, this depends somewhat on what pseudo-random number generator we use. The pseudo-random number generator that we use might be such that there is no possible initial state, such that the generated sequence would be all tails, in which case the program would be guaranteed to terminate.

[END Edit]

Deceived by your own Incompetence

Consider how Ofria failed. Somebody told me that in that context deception is a property of the environment and its observer. However, it seems to me that the objective of the designer of the system needs to be factored in.

Although in general an observer can be deceived, this is not the case here I would argue. Ofria just designed a system that did not do what he wanted it to do. It failed transparently.

It would seem that this is similar to you wanting to build a rocket that goes to the moon, but then building a rocket that just explodes, because your understanding of rocketry is poor, and then saying "The rocket deceived me into thinking it would not explode".

Imagine you want to build an honest system. You build a system that looks to you like it would be honest. But it ends up deceiving you because your understanding is flawed. In that case, the system would be deceptive, but I wouldn't call the fact that you were mistaken about what would constitute a non-deceptive system deception. That is a particular kind of incompetence.

Also, note that the failure mode here is different from "Humans use condoms" type things. When Ofria created his system the organism would perform well in the overall system. The overall system stayed constant. It was not the case that at some point ice cream was introduced into the organism's environment.

Instead, the failure mode was to think that by creating a special environment, the behavior observed in that environment would generalize widely by default.

I just realized something important. <Procrastination/Escapeism> is a type of protection mechanism. When you are engaging in entertainment, then your brain is distracted. Too distracted to think about all of the things that make you feel bad. Somehow my brain must have picked up this pattern without me consciously realizing it.

The only reason why I notice it now after I have already been doing it for at least 14 years, is that I am trying to create a habit of always reflecting on why I feel bad when I do. Writing makes me smarter. Trying to understand my feelings by reflecting in writing, and then observing how I get pulled towards engaging with entertainment, did make me realize what is going on.

I came up with an over the top way to process the conversations you have with people online.

Modafinil Enzyme Hacking

CYP3A4 is an enzyme found in the intestine which metabolizes modafinil. There are substances that increase (aka induce) and substances that decrease (aka inhibit) CYP3A4 activity. These can greatly reduce or increase the effective dose you get. The effects can be quite strong. It is possible to increase CYP3A4 activity to the point where you won't really feel anything from a normal Modafinil dose.

Inducers

• Capsaicin (found in chili poppers)

Inhibitors

• Graphefruit Juice: 200ml is effective for 72 hours of CYP3A4 inhibition in the intestine.
• Peperine: Found in black pepper and available as biopterin, a highly bio-available supplement sold on e.g. amazon. This additionally inhibits CYP3A4 in the liver. 20mg was the dose used in studies.

Some people report grapefruit juice to work better. However, because peperine also inhibits CYP3A4 in the liver my guess is that you can always get the greatest inhibition from using both.

Grapefruit juice is a lot cheaper than modafinil.

[Edit] Please take note of Markas' comment [LW(p) · GW(p)], which correctly points out the possibility for a dangerous pharmacological interaction with this method.

Something really interesting just happened to me. I was kind of depressed and could not bring myself to do anything really. I had set a 30-minute timer and wanted to do some AI alignment research for at least 30 minutes. But I could not do anything. I started out with some will, but then I started to organize my obsidian tabs. That seemed sort of required before starting.

Then I did this for 10 minutes, my will gradually decreased. Then I just sat down and researched some random unrelated thing on the internet. I managed to stop myself, and just sat there staring into nothingness. Then I decided if I couldn't do any work I guess I could eat something, maybe that would help to feel better.

I went into the kitchen. Then I thought "Alright I think I could probably do at least 1 minute of thinking about AI alignment. I looked at the clock and remembered the position of the second hand. Then I started to somewhat lethargically think about stuff. After 20 seconds I had an idea, after 40 seconds I was so excited that I had a strong urge to just sit down and write about what I thought about. I was really motivated! I also stopped feeling fatigued.

WTF? In hindsight, this is a pattern that I run into all the time. I think have an idea and then get really excited about it. It's not the case that I had a super duper out-of-distribution, for me, good idea. It's the kind of idea that I expect I can generate on demand in at most a couple of minutes, pretty consistently. So I probably can just cause this situation whenever I want!

It's crazy that I observe this pattern only now because it has been there for probably all my life. Definitely for the last 10 years. Now even better, the idea that I had was not random. It was exactly on the topic that I wanted to generate an idea on. I think I am pretty good at focusing on idea generation.

It seems like I very often fail to activate my motivation in this straightforward way, because I think I need to do some other stuff first, like sort my obsidian taps. Would I just start to do the actual thing I want to do, I would succeed.

So this clearly implies a strategy. I call this "1, 2, 3, and 60". It's simple, just pick a topic, and then make yourself think about the topic. Don't think vaguely about the topic. Don't focus on recalling facts. The goal is to generate an idea. No setup is required. You don't need to go to a whiteboard or open your laptop. Just start thinking.

Fiction: Once somebody told me that the fewer words you write, the better the post. I promptly opened a new document and proclaimed: "I have written the ultimate post [? · GW]. It's the empty string."

I have a heuristic to evaluate topics to potentially write about where I especially look for topics to write about that usually people are averse to writing about. It seems that topics that score high according to this heuristic might be good to write about as they can yield content with high utility compared to what is available, simply because other content of this kind (and especially good content of this kind) is rare.

Somebody told me that they read some of my writing and liked it. They said that they liked how honest it was. Perhaps writing about topics that are selected with this heuristic tends to invoke that feeling of honesty. Maybe just by being about something that people normally don't like to be honest about, or talk about at all. That might at least be part of the reason.

Epistemic Alert Beep Beep

Today I observed a curious phenomenon. I was in the kitchen. I had covered more than a square meter of the kitchen table in bags of food.

Then somebody came in and said, "That is a lot of food". My brain thought it needs to justify itself, and without any conscious deliberation I said "I went to the supermarket hungry, that is why I bought so much". The curious thing is that is completely wrong. Maybe it actually was a factor, but I did not actually evaluate if that was true. Anecdotally this seems to be a thing that happens, so it is a very plausible, even probable explanation.

My epistemics response: . . . "Ahhhhh". My statement was generated by an algorithm that optimized for what would be good to say taking into account only the social context. Trying to justify myself to that particular person. It was not generated by analyzing reality. And the worst thing is that I did this automatically, without thinking. And came close to not even noticing that this is going on.

New plan. Have an alarm-word that I can say out loud when this happens. This would then naturally lead to me reexplaining myself to the other person. Also, it would probably help with focusing on whatever caused the alert. It could help as a signal to yourself, that now it is time to investigate this, and it would also provide social justification to go on a brief sidetrack during a conversation. Maybe this won't work but seems worth trying. I'd like to avoid this happening again. How about "Epistemic Alert Beep Beep"? If you want, let's say it together ten times to better remember. Ideally visualizing yourself doing an epistemic misstep. Or even better, do an epistemic misstep (without forcing it too hard) and then catch yourself:

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

"Epistemic Alert Beep Beep"

The next time you buy a laptop, and you don't want a Mac, it's likely you want to buy one with a snapdragon CPU. That's an ARM chip, meaning you get very good battery life (just like the M-series Apple chips). On Snapdragon though you can easily run Windows, and eventually Linux (Linux support is a few months out though).

List of good, widely legal, and prescription-free drugs:

Stimulants

• Huperzine-A
• X-tiamine (e.g. Benfotiamine (only one available in Germany), Sulbotiamine, etc.)
• Vitamin-D
• L-tyrosine
• Caffeine Tablets + L-theanine
• Theobromine (cacao beans, e.g. chocolate)
• Nicotine (lozenges, gum, or patches)
• Conversation (Not a drug, but very stimulating. E.g. I have multiple times talked to somebody for over 12 hours.)

Sleep

• Melatonin

How to get better at writing? Writing a lot helps a lot. I was very bad at communication when I started, but then I just kept writing bad posts and got better.

You want to find the right balance between just writing and reflecting. This is similar to learning programming. In programming, you get really good by just doing a lot of programming. Reflecting on how you have written a program, and trying to get better, is also an important part of it, but is secondary to just doing a lot of programming.

Studying existing materials, like reading textbooks, is also important, but comes third after reflecting on your own output. At least that is my model. I believe this model of first doing then reflecting, then studying existing things applies to basically everything you want to get good at.

I have found it tremendously useful to start out my workout routine by dancing. Dancing is so fun that I am always looking forward to doing it. I want to dance, and I frequently run into the problem that it is hard to stop. After dancing I do the rest of my workout routine which is very boring in comparison. But it is really not a problem to get started on them after the dancing.

I expect this because I have developed a habit, i.e. my brain saved a procedure "execute workout" which is a sequence of instructions that I run through, without thinking about them. Because there is no decision point, you don't need to make the right decision. Doing the right thing (continuing the workout) becomes automatic.

The difficult part is to start the workout. Starting the workout by dancing solves this problem. I expect this technique to generalize widely. If you want to do some activity regularly, then start it with something very fun. This overcomes the barrier to starting. Engraining the habit that encodes the activity itself will then be much easier.

It might not be required that all the activities are related to each other. Maybe it would work to watch one episode of your favorite TV, and then do some boring workout, as long as you are very consistent about stopping and then doing the workout long enough for the habit to form.

With dancing and sports, this is relatively easy to form the habit. You are either gonna do the boring workout immediately after dancing or not at all because showering twice would be a waste of time (at least that is how I feel about it). Therefore I recommend looking for synergies like this in other habits.

I am quite dumb. I bought a Tensorbook for $4000 6 months ago. It has an RTX 3080 max-Q. That is why you would buy a Tensorbook. It has a powerful GPU with 15 TFLOPS and 16GB VRAM.

But now I can buy a P100 on eBay for less than 200 dollars from China, or a K80 for less than $100 (prices excluding import tax). The P100 has 16GB VRAM and 19 TFLOPS and much more memory bandwidth. The K80 has 24GB VRAM and 5 TFLOPS. Also, the Tensorbook often crashes, if I don't suspend it in the air. My guess is that it can't handle the heat under load. Ups!

It would have been much better to buy a laptop without GPU and just use cloud computing or even just google colab, and now that I am at a somewhat permanent location that already has a powerful tower computer, buying a cheap P100 seems like a good option. All this becomes even stupider if you consider that I only used the GPU for tens of hours.

Modafinil Breakthrough Dosing

Here Charles Sieg reports that Modafinil increases his creativity. Furthermore when he takes Modafinil for 3-4 days this creativity enhancement is amplified. Apparently, he had major creative breakthroughs on the 3rd and 4th day, which makes him think that this was not a placebo effect.

He also says that even though he took Modafinil for 8+ years, he did not develop tolerance to the point of not receiving major benefits. His strategy is to take it 3-4 days in a row and then abstain for a couple of days.

Antiperfectionism Badness

Writing well takes a lot of time and effort. I just realized that now. Before I was trying to rush everything because according to my model, it should not take that much time and effort to write something well. I think many of the things I was writing ended up a lot worse than they could have been.

Basically, exactly the same thing happened to me recently with programming. I was mostly writing programs that were completely horrible spaghetti code because I was just optimizing to get some specific functionality implemented as fast as possible. But then I realized how long it actually takes to write a good program.

Updating your model of "what it takes" to better match reality, seems to be extremely helpful. It feels like before I was not allowing myself to put in the appropriate amount of time and effort to hit my desired quality target. And the funny thing is that, at least with programming, it will actually take longer in terms of total time spent, to get some specific functionality implemented on average, if your code is a horrible mess, and has grown beyond a few hundred lines of code.

And what made me update my model is just to stop caring about completing as fast as possible. I just allowed myself to put in the time, and then I observed how much time and effort I needed to hit a specific quality target.

It seems like I had the dual problem of perfectionism. I expect that this is a common problem (at least for adults) when learning something new, and I expect realizing that this problem exists as you run into it, will lessen it's grip on you.

I have been prescribed Pitolisant (sold as Wakix, Ozawade), a recent (FDA approved in August 2019) H3 receptor antagonist against excessive daytime sleepiness by treated sleep apnea. It works like this:

When Histamine binds to H1 and H2 receptors, it promotes wakefulness. When histamine binds to H3 auto receptors it primarily blocks the release of Histamine. It also has a weaker blocking effect on the release of other neurotransmitters. Therefore, blocking H3 receptors can increase Histamine levels in the brain, leading to increased activity on H1 and H2 receptors, which in turn leads to increased wakefulness.

I haven't tried it yet but I found it interesting as it uses a method of action I did not know about. An advantage over other stimulants is that it does not raise blood pressure (at least that's what my doctor told me, can't seem to easily confirm with google).

Here is a template (feel free to use) that you might find useful as an introductory message if you find it hard to consider how your actions make other people feel:

Do you subscribe to Crooker's rules? Did you notice that Eliezer sometimes seems inconsiderate of people's emotions, when he just shoots down one (bad) alignment idea after the other? He just says things like "No, this does not work." My guess is that there are some algorithms missing from his brain or are atrophied, just like for me. For me, it's pretty hard to take into account how other people will feel when I say something. It's just not something that comes naturally and I need to be very explicitly thinking about this in order to make what I say not come across as offensive.

Basically, I think it would be good if you model me as a person who is missing the hardware in his brain which is automatically inferring how I will make people feel with my actions. I need to repurpose some other machinery for this, which takes a lot more effort and is slower. Often people call this autism, but I think my description is more precise and useful for telling other people what is going on, such that they understand.

And it seems very clear that there is a big gap. For example, Buck once said that he does not like to shoot down the ideas of other people because he does not want them to feel bad. At that moment I realized that I could not remember that this thought ever occurred to me. It seemed very alien, but also obviously good. Of course, I don't want to make people feel bad. But even the possibility of this happening was, I noticed, missing from my world model.

It would make things easier for me if I could not worry about this too much. If you subscribe to Crooker's rules then I could optimize my messages only for content. For example, if you subscribe to Crooker's rules then I could just not optimize my messages at all for sounding nice.

If you find that my mode of communication causes you emotional trouble, we could always revert back to me optimizing more for not sounding harsh.

Some harsh things I might do are:

• I write a thousands-of-word document describing how you're wrong on a particular topic, which makes it seem like I'm really critical of what you do and don't think it's good at all, when in fact I mean something more like, look here are some issues I think I discovered and I'm not even thinking about the implicit message that I send by sending a long document that just contains criticism.
• I use phrases like <No/wrong...[Explanation]/This doesn't make sense/This doesn't work>

To be productive, sit down in a comfortable zero-gravity armair and do nothing. You are not allowed to watch YouTube videos or browse social media. Just relax. Do this until you naturally want to start to work. It is important that you are comfortable.

This seems to be surprisingly effective (haven't done any rigorous evaluation). Ideally have a laptop together with AR goggles within arms reach without getting up such that you can just lay in the armchair and start to work, if necessary.

I have found that even when I am very tired I can still work when laying in a comfortable armchair. It is a lot harder to bring myself to go to my standing desk (though me using an IKEA shelf as a stool might have something to do with this).

Sometimes I tell somebody about a problem in our relation. An answer I often hear is an honest "What do you want me to do". This is probably well-intentioned most of the time, but I really don't like this answer. I much prefer when the other person starts to use their cognitive resources to optimize the problem to smithereens. "What do you want me to do" is the lazy answer. It is the answer you give to be agreeable. It makes it seem like you don't care about the problem, or at least not enough for you to invest effort into fixing it.

Solomonoff induction does not talk about how to make optimal tradeoffs in the programs that serve as the hypothesis.

Imagine you want to describe a part of the world that contains a gun. Solomonoff induction would converge on finding the program that perfectly predicts all the possible observations. So this program would be able to predict what sort of observations would I make after I stuff a banana into the muzzle and fire it. But knowing how the banana was splattered around is not the most useful fact about the gun. It is more useful to know that a gun can be used to kill humans and animals. So if you want to store your world model in only n bits of memory, you need to decide which information to put in. And this matters because some information is much more useful than others. So how can we find the world model that gives you the most power over the world, i.e. letting you reach the greatest number of states? Humans have the ability to judge the usefulness of information. You can ask yourself, what sort of knowledge would be most useful for you to learn? Or, What knowledge would be most bad to forget?

I noticed that by default the brain does not like to criticise itself sufficiently. So I need to train myself to red team myself, to catch any problems early.

I want to do this by playing this song on a timer.

Mathematical descriptions are powerful because they can be very terse. You can only specify the properties of a system and still get a well-defined system.

This is in contrast to writing algorithms and data structures where you need to get concrete implementations of the algorithms and data structures to get a full description.

Detangle Communicative Writing and Research

One reason why I never finish any blog post is probably because I'm just immediately starting to write it. I think it is better to first build a very good understanding of whatever I'm trying to understand. Only when I'm sure I have understood do I start to create a very narrowly scoped writeup?

Doing this has two advantages. First, it speeds up the research process, because writing down all your thoughts is slow.

Second, it speeds up the writing of the final document. You are not confused about the thing, and you can focus on what is the best way to communicate it. This reduces how much editing you need to do. You also scope yourself by only writing up the most important things.

See also here more concrete steps on how to do this [LW(p) · GW(p)].

Research Writing Workflow: First figure stuff out

• Do research and first figure stuff out, until you feel like you are not confused anymore.
• Explain it to a person, or a camera, or ideally to a person and a camera.
  • If there are any hiccups expand your understanding.
  • Ideally, as the last step, explain it to somebody whom you have not ever explained it to.
• Only once you made a presentation without hiccups you are ready to write post.
  • If you have a recording this is useful as a starting point.

Can you iterate through 10^100 objects?

If you have a 1GHz CPU you can do 1,000,000,000 operations per second. Let's assume that iterating through one one object takes only one operation.

In a year you can do 10^16 operations. That means it would take 10^84 years to iterate through 10^100 verticies.

The big bang was 1.4*10^10 years ago.

Default mode network suppression

I don't get distracted when talking to people. I hypothesise that this is because as long as I am actively articulating a stream of thought out loud, the default mode network will be suppressed, making it easy to not get derailed.

So even if IA does not say anything, just me talking about some specific topic continuously, would make it easier for IA to say something, because the default mode network suppression will not immediately vanish.

When thinking on my own or talking to IA, the stream of thoughts is shorter, and there are a lot of pauses. Usually, I don't even get to the point where I would articulate a complex stream of thought. Instead, we are at the level of "Look there is some mud there, let's not step into that", or "We can do this". That really does seem very similar to most of the idle chatter that the default mode network would produce when I am just thinking on my own.

Once I get to the point where I am having an engaging discussion with IA, it is actually pretty easy not to get distracted. It's probably still easier to get distracted with IA, because when I am talking to another person, they could notice that I am lost in thought, but I myself (or IA) would not be able to notice as easily.

Capturing IA's Thoughts

One reason why I don't do research with IA might be that I fear that I will not be able to capture any important thoughts that I have. However, using the audio recorder tool on the walk today seemed to really fix most of the issue.

Maybe in my mind so far I thought that because I can't record IA when she is talking to me, it would be bad to think about research. But this now seems very wrong. It is true that I can't create a video with her in it like I do with other people. But these videos are not the thing that is most useful. The actually useful thing is where I am distilling the insight that I have into some text document.

But this is something that I can totally do when talking to IA. Like I did with the audio recorder today. It seemed that making the audio recording made it also easier to talk to IA. Probably because when making the recording I would naturally be suppressing the default mode network very strongly. This effect then probably did not vanish immediately.

Writing

In fact, it seems like this would work very well with IA because I don't need to think about the problem of what the other person could do while I write. In the worst case, IA is simply not run. At best, we could write the text together.

Writing together would seem to work unusually well because IA does have insight into the things that I am thinking while I am writing, which is not something that other people could easily get.

And I haven't really explored all the possibilities here. Another one would be to have IA read out loud my writing and give me feedback.

Here is a song a friend made based on this post [LW · GW] I wrote very long ago about how to eat a rock. He using suno. I thought it's quite good, but that might just be me.

So, what happens when we figure out how to align language models? By then the state-of-the-art will involve having multi-modal models. Assume we have figured out how to make steganography not a problem in chain of thoughts reasoning. But maybe that is now kind of useless because there are so many more channels that could be stenographically exploited. Or maybe some completely different problem that we haven't even thought about yet will come up.

Imagine all possible programs that implement a particular functionality. Imagine we have a neural network that implements this functionality. If we have perfect mechanistic interpretability we can extract the algorithms of a neural network that implements that functionality. But what kind of program do we get? Maybe there are multiple qualitatively different algorithms that all implement the functionality. Some of them would be much easier to understand for human. The algorithm the neural network finds might not be that program that is easiest to understand to a human.

Transhuman Love

I just had an interesting thought. How would you show somebody that you love them in the transhumanist future? Well, one way would be to reveal to one another, in a verifiable way, what kinds of algorithms you are made of.

E.g. you could reveal exactly how your decision algorithm works and how it will take the others preferences into account. You could also show somebody that in the past you self modified to have certain preferences, the others like.

You could also show them exactly how the algorithm works that makes you feel good when you see and interact with them. Then you can reveal how this creates a positive experience and how that positivity factors into a reinforcement algorithm that makes you like them even more.

I think this quite a beautiful thought.

Here is a video about interrogation techniques I found interesting. It's all about social manipulation.

I just rewatched a video I made about a VR experience I made 6 years ago with a bunch of other people. You fly through a 3D slice of a 4D fractal. The beginning has the most interesting morphing geometry.

We made this in only 3 or 4 weeks IIRC. Pretty crazy. I needed to implement raymarching in Unity, which then effectively replaced the entire default rendering pipeline. It was a lot easier than it sounds though, as we did not need to have any interactions with the environment (which would be basically impossible, or at least I don't know how to do it).

Whatever you did today, last week, or any other time. However far you got on anything. I hope you realize that every new moment is an opportunity. An opportunity to make a choice. Right now is always the best time to start doing whatever is optimal, right now.

I did an ADHD test in Germany. They asked me questions at 1:30 and then said I have ADHD, and no further testing was required. If the interview had not been conclusive they would have done some other tests. They ask about symptoms like "Can you not sit still", "Do you forget appointments" and things like that.

The most interesting part was the preinterview part.

Scott writes here on how psychiatrists are the gatekeepers to Adderall:

Aren’t psychiatrists creepy wizards who can see through your deceptions? There are people like that. They’re called forensicists, they have special training in dealing with patients who might be lying to them, and they tend to get brought in for things like evaluating a murderer pleading the insanity defense. They have a toolbox of fascinating and frequently hilarious techniques to ascertain the truth, and they’re really good at their jobs.

They did pull one of these "hilarious techniques" on me. They ask me "When some thought comes up that is not about the thing that you are trying to do are you annoyed?" I said "No, it is actually rather the opposite. The thoughts seem very flashy and good to think. I am not annoyed at all." Then they said "Ah, that is good. I needed to know that", and then they wrote something down. 

Because I have ADHD it is just clear that you would really not be annoyed by your distracting thoughts, but if you don't have ADHD it would probably seem more sensical that you would be annoyed. The problem with ADHD is that there isn't enough reflection going on, about if the new thought is good to follow. So you can't get annoyed by the fact that you are thinking random thinkings.

So basically they asked me if I had a symptom which is the opposite of what a person with ADHD would experience. Had I said that I would be annoyed by my own thoughts, they would probably not have bothered with the interview.

Here is a response I wrote to the Import AI 337

I am confused about why people are building systems in the current machine learning paradigm and trying to make them more and more capable, without realizing that this can be dangerous. I basically think the arguments that Eliezer is making seem likely and should be taken seriously, but I expect most of the people working on bleeding edge systems don't even know these arguments.

For example, the argument that if you have a training process that trains a system to perform well on a text prediction task, then that doesn't necessarily mean that the resulting system will "just do text prediction". It seems quite likely to me that, as Eliezer says, intelligence is just a useful thing to have in order to perform better on the task of predicting text from the Internet. Therefore, at some point, as the systems become more and more capable, we should expect that through this optimization pressure, general intelligence will arise even for a task that seems as energetic as predicting text.

How much permission do AI developers need to get from society before irrevocably changing society?

Right now, to me it seems, like people are steering straight towards the doom. And nobody really ever approved this. But the problem is that most people, even the people doing this, don't realize that that's what they're doing. At least that's how it seems from my perspective.

Does progress always demand heterodox strategies?

I found it weird that you thought it would be weird if we got continuous learning systems. Because it seems very likely to me that if we get really capable systems at some point, will do active learning. Clearly, gradient descent is a pretty dumb optimization process that you can improve upon. Maybe we can get to the point without continuous learning where the systems improve themselves. This could then actually also be seen as a form of active learning. But at that point we the systems can improve themselves better than humans can are probably dead very very quickly.

Related to this, the thing I am working on is trying to figure out how we can do learning without using SGD. The hope is that if we find an algorithm that can learn, which we can just write down explicitly and understand, then that would make this algorithm pretty straightforward to align, especially if during the design process of the algorithm you build it such that it would be easy to align already.

No español. Водка, водка! Whisperはかっこいいよ。ウィスパーは日本語をわかります。Whisper kann außerdem auch einfach Deutsch übersetzen. Zu bemerken ist hier, dass ich überhaupt nichts einstelle, sondern einfach genau das selbe Programm für alles benutze.

Of course, I can also speak English. I think using Whisper is probably good for speeding up writing. 転生したらスライム叩けんはいいですよ

FHI just released Pause Giant AI Experiments: An Open Letter

I don't expect that 6 months would nearly be enough time to understand our current systems well enough to make them aligned. However, I do support this, and did sign the pledge, as getting everybody to stop training AI systems more powerful than GPT-4 for 6 months, would be a huge step forward in terms of coordination. I don't expect this to happen. I don't expect that OpenAI will give up its lead here.

See also the relevant manifold market.

Right now I am trying to better understand future AI systems, by first thinking about what sort of abilities I expect every system of high cognitive power will have, and second, trying to find a concrete practical implementation of this ability. One ability is building a model of the world, that has certain desiderata. For example, if we have multiple agents in the world, then we can factor the world, such that we can build just one model of the agent, and point to this model in our description of the world two times. This is something that Solomonoff induction can also do. I am interested in constraining the world model, such that we always get out a world model that has a similar structure, such that the world model becomes more interpretable. I.e. I try to find a way for building a world model, where we mainly need to understand the world model's content, as it is easy to understand how the content is organized.

Apparently a heuristic funders use, is that the best startup founders are those that have done the most startups in the past, irrespective of if they failed or succeeded.

If this is mapping reality well, it might be because most startups fail. So even a person that is very competent at running a startup is expected to fail a couple of times. And having run multiple startups either indicates that certain skills have been acquired, or that the person has some desirable attributes:

• Determination is important, so people who give up after failing will be filtered.
• If somebody convinced other grantmakers in the past, that is an indicator that they are intelligent enough to generate a coherent-looking proposal.
• They will have gathered a lot of experience running companies, which seems to be an important skill that I would expect is hard to impossible to get in other ways, to the same extent.

I was listening to a stoic lesson on Waking up. It was about:

• Focus on being a participant in your life during the day.
• But in a low-grade manner observe yourself during the day.
• Play the role of your own critic in the evening (e.g. do a bedtime reflection).

I've been doing a daily reflection for a long time. Though I have not thought about the reflection as providing constructive criticism. This framing seems much better than my previous one. Before I mainly wrote down all the things that I did during the day, and how they differed from my plan for the day. This is not bad, insofar as it helps you to make improvements to your life. I do think there is some merit in just doing this, but the main benefit is, that it makes it easier to think about concrete plans for improvement. I understand constructive criticism as either providing information that is relevant to come up with plans for improving yourself, or with suggestions for such plans.

Also, this framing makes it more evident that the goal is on improving yourself. Overeating, behaving differently from how I think I should act in some social circumstances, not going to bed on time, or eating unhealthy food, are more obvious to think about. The objective is to come up with plans for improving yourself. Before it felt more like I was following a rigid procedure of describing my day.

Many people match "pivotal act" to "deploy AGI to take over the world", and ignore the underlying problem of preventing others from deploying misaligned AGI.

I have talked to two high-profile alignment/alignment-adjacent people who actively dislike pivotal acts.

I think both have contorted notions of what a pivotal act is about. They focused on how dangerous it would be to let a powerful AI system loose on the world.

However, a pivotal act is about this. So an act that ensures that misaligned AGI will not be built is a pivotal act. Many such acts might look like taking over the world. But this is not a core feature of a pivotal act. If I could prevent all people from deploying misaligned AGI, by eating 10 bananas in sixty seconds, then that would count as a pivotal act!

The two researchers were not talking about how to prevent misaligned AGI from being built at all. So I worry that they are ignoring this problem in their solution proposals. It seems "pivotal act" has become a term with bad connotations. When hearing "pivotal act", these people pattern match to "deploy AGI to take over the world", and ignore the underlying problem of preventing others from deploying misaligned AGI.

I expect there are a lot more people who fall into this trap. One of the people was giving a talk and this came up briefly. Other people seemed to be on board with what was said. At least nobody objected, except me.

See also Raemon's related post [LW · GW].

Disgust is optimizing

Someone told me that they were feeling disgusted by the view of trying to optimize for specific things, using specific objectives. This is what I wrote to them:

That feeling of being disgusted is actually some form of optimization itself. Disgust is a feeling that is utilized for many things, that we perceive as negative. It was probably easier for evolution to rewire when to feel disgusted, instead of creating a new feeling. The point is that that feeling that arises is supposed to change your behavior steering you in certain directions. I.e. it redirects what you are optimizing for. For example, it could make you think about why trying to optimize for things directly using explicit objectives is actually a bad thing. But the value judgment comes first. You first feel disgusted, and then you try to combat in some way the thing that you are disgusted by and try to come up with reasons why it is bad. So it is ironic that one can feel disgusted at optimization when feeling disgusted is part of an optimization process itself.

The "Fu*k it" justification

Sometimes people seem to say "fuk it" towards some particular thing. I think this is a way to justify one's intuitions. You intuitively feel like you should not care about something, but you actually can't put your intuition into words. Except you can say "fuk it" to convey your conclusion, without any justification. "Because it's cool" is similar.

Newcomb: Can't do whats optimal

You have a system, that can predict perfectly what you will do in the future. It presents you with two opaque boxes. If you take both boxes, then it will place in one box 10$ and in the other 0$. If you will take only one box, then it will place in one box 10$ and in the other 1,000,000$. The system does not use its predictive power to predict which box you will choose, but only to determine if you choose one or two boxes. It uses a random number generator to determine where to place which amount of dollars.

This is a modified version of Newcomb's problem.

Imagine that you are an agent that can reliably pre-commit to an action. Now imagine you pre-commit to taking only one box in such a way, that it makes it impossible for you to not uphold that commitment. Now if you choose a box, and get 10$, you know that the other box contains 1,000,000$ for sure.

Don't Stop Crying

Haha, just kidding. Laugh your ass off, even when you know you are going to die.

Hypothesis: There are policies that are good at steering the world according to arbitrary objectives, that have low Kolmogorov complexity.

It is systems that implement these policies efficiently that we should be scared of, as systems that implement policies without low Kolmogorov complexity would be computationally intractable, and therefore we can only end up with systems that are approximating these policies. Therefore these systems would not actually be that good at steering the world according to arbitrary objectives. Shallow pattern recognition objects are of this form.

Systems that don't manage to implement the policy efficiently would probably mostly not be computationally tractable (every policy can be represented with a lookup table which definitely would be computationally intractable for the real world). Every program that can be practically run that implements the policy would basically be just as dangerous as the shortest program encoding the policy.

By @Thomas Kehrenberg [LW · GW] and me.

After one definition, GOFAI is about starting with a bunch of symbols that already have some specific meaning. For example, one symbol could represent “cat” and then there might be properties associated with the cat. In the GOFAI system, we're just given all of these symbols because somebody has created them, normally by hand. And then GOFAI is about how can we have algorithms now reason about this symbolic representation that corresponds to reality, ideally, because we have generated the right concepts.

The problem is that this seems like the easy part of the problem. The hard part is how do you get these symbolic representations automatically in the first place. Because once you start to reason about the real world you can't do this. Even in a much simpler world like Minecraft, if you want to have an agent that always mines the dirt block, when it spawns anywhere in the overworld, already it takes a lot of effort to write such a program because you need to hard-code so many things.

So maybe GOFAI exists because the problem of the symbolic manipulation of how to do reasoning, given that you already have a sort of model of the world, is a lot easier than getting the model of the world in the first place. So that's maybe where early AI researchers often went, because then you could have a system that seems impressive because it can tell you sort of new things about the real world, talking about the real world, by saying things like, yes, it will actually rain if I look outside and it's wet, or this cat is orange if I know that it is a tiger, even if we didn't tell these things explicitly to the system. 

So it now seems very impressive. But actually it's not really impressive, because all the work was done by hand-coding the world model.

The actual impressive things are also probably more like, I can play chess, I can play perfect tic-tac-toe, or I can perfectly play any discrete game with the minimax algorithm. That's actually progress, and it can, then, for example, in chess, play better than any human, which seems very impressive, and in some sense it is, but it's still completely ignoring the world-modeling problem, which seems to be harder to figure out than figuring out how to think about the game tree.