The appeal for me is simple: I'm not a professional writer. I have a lot of ideas and concepts that I want to explore. But writing is so toilsome and time-consuming that I basically never do it (I mean fiction; i still write my own posts.)

So all these ideas would go unexpressed if I didn't go Centaur. I do it for myself: I'm usually content to keep the story to myself after I write it. But if I ever published, I would credit Sonnet with 50% of the work (although I know it's extremely unpopular to claim having used AI for creative writing.)

Thank you for sharing. I've enjoyed this story. 

I've been writing fiction collaborately with Sonnet since 3.6 came out. I found that 4.5 is better at writing. But the biggest jump I've seen is in literary analysis and criticism. When I gave 4.5 my story collection, it was able to identify the most promising one and it also dismissed a few that it said were failures (I mostly agree). I think I could get some interesting results by putting the model through ideate-generate-criticize loops. 

I want to point out that getting the model to one-shot a story without even a plot or a concept is the hardest possible task. Much like in other fields (like programming) you can see steady linear progress by working collaboratively with the model and supervising it. Whereas if you just ask it to one-shot a story (or an app, or a legal document) you will see little progress for a long time, until all the right pieces come together. 

Therefore, it's all the more impressive that you were impressed a this one-shot story (although it's informed by the works in the context window).

Here's a thought experiment.

In version A, I have a button that non invasively scans my brain and creates 10 perfect copies of my brain state in a computer. I press the button. For an instant, 11 identical mind states exist in the universe. Then each mind starts diverging along different causal chains.

Intuitively, I expect the following:

• I won't experience anything unusual after pressing the button (eg, I won't wake up in a computer). I will still feel that I am in my physical body, in the room with the button
• each of the mind copies will feel that they are the 'one true version of vlad' and won't experience the other minds 'from the inside'. presumably, they will be surprised to be in a computer and not in the room?
• if I shut down the computer and kill the 10 minds, I won't experience anything unusual

In this case, I identify myself with the embodied mind.

In version B, the setup is identical except the scan is destructive. The second I press it, my physical body is destroyed.

Now, what happens to me? There's no specific reason for me to end up in one of the minds and not the others. But I cannot go to all 10 minds at the same time — I am a single mind with its own casual chain, not a collection of minds.

For instance, imagine each of the 10 minds is caused to feel a different sensation at the same time. There's nobody to feel all 10 sensations at the same time because the minds are causally isolated. Yet I cannot say that I am feeling a particular sensation and not the others.

So in version B, I still identify myself with the embodied mind, which is destroyed — hence oblivion. Conversely, what happens to the 10 minds if I delete them from the computer? Oblivion.

(This is just my attempt to map my naive intuitions. I have a sense some version of no-self could be the solution, but I'm not there yet. I also feel that naive intuitions fail for Everett branches which is another reason to be suspicious.)

Nice! Last weekend I expanded https://www.gptrim.com/ to allow the user to a) see savings in both characters and tokens; b) determine their own combination of word processing functions. Then I saw, like you said, that to save tokens you only want to remove stopwords. I will next add the option to remove punctuation. I also want to give users two general recipes: optimize for saving tokens vs. optimize for saving characters. Always happy to take more ideas.

I will probably write again on this, on my new personal Substack or other websites, reporting what I've learned. Would you like me to cite you and link to your profile? My DMs are open! 

P.S: Due to my speedrunning the coding, the website now has an issue where it sometimes adds spaces to the text. I am aware of this and will fix it latest next weekend. The Python package that runs all this is accepting PRs: https://github.com/vlad-ds/gptrim. 

I see your point. I think the existing tokenizer is designed to keep all parts of text, while the idea here is to sacrifice some information in favor of compression. But writing this, I also realized that this approach is more effective at saving characters than tokens.

This is what I was hoping for when I wrote this post. Thank you for your insight. 

New position: sometimes when using ChatGPT, you only care about the number of characters, because of the character limit in the chat message. In that case, you want to get rid of spaces. But if you want to save on tokens, you probably should keep spaces. I think the solution is: a) allow the user to choose the mix of transformations for their use case; b) show them how much they are saving in characters and tokens so they can optimize for their use case. 

Free, universal financial tracker.

I wrote an article on this subject (i.e. why do we play zero-sum games while praising positive-sum games?)

https://native-wonder.blogspot.com/2020/12/things-people-want.html

Thank you, this is very useful. Lately I've been interested in programs that are fully online and could be completed in a year. Would you have any recommendations for that?

Wonderful, thank you!