I’ve posted this before but I want to make it more clear that I want feedback.
I’ve written an essay on the effects of interactive computation as an improvement for Solomonoff-like induction. (It was written in two all-nighters for an English class, so it probably still needs proofreading. It isn’t well-sourced, either.)
I want to build a better formalization of naturalized induction than Solomonoff’s, one designed to be usable by space-, time-, and rate-limited agents, and interactive computation was a necessary first step. AIXI is by no means an ideal inductive agent.
If you want people to read what you write, learn to write in a readable way.
Looked at your write-up again… Still no summary of what it is about. Something along the lines of (total BS follows, sorry, I have no clue what you are writing about, since the essay is unreadable as is): “This essay outlines the issues with AIXI built on Solomonoff induction and suggests a number of improvements, such as extending algorithmic calculus with interactive calculus. This extension removes hidden infinities inherent in the existing AIXI models and allows .”
I’m in the process of writing summaries. I replied as soon as I read your response.
If you want people to read what you write, learn to write in a readable way.
You are pretty much the first person to give me feedback on this. I do not have an accurate representation as to how opaque this is at all.
In algorithmic representations:
Separate hypotheses are inseparable.
Hypotheses are required to be a complete world model (account for every part of the input).
Conflicting hypotheses are not able to be held simultaneously. This stems mainly from there being no requirement for running in a finite amount of space and time.
There are other issues with Solomonoff induction in its current form, such as an inability to tolerate error, an inability to separate input in the first place, and an inability to exchange error for simplicity, among others. Some of these are addressable with this particular extension of SI; some are addressable with other extensions.
There is a similar intuition about nondeterministic hypotheses and a requirement that only part of the hypothesis must match the output, as nondeterministic Turing machines can be simulated by deterministic Turing machines via the simulation of every possible execution flow, but that strikes me as somewhat dodgy.
Feedback (entirely on the writing): The first goal when editing this should be to eliminate words from sentences. Use short and familiar words whenever possible. Change around a paragraph’s structure to get it shorter. Since this is for English class, cut out every bit of jargon you can. If there’s a length requirement, you can always fill it with story.
The best lesson of my dreadful college writing class was that nonfiction can have a story too—and the primary way you engage with a nontechnical audience is with this story. Solomonoff induction practically gets a character arc—the hope for a universal solution, the impotence at having to check every possible hypothesis, then being built back up by hard work and ancient wisdom to operate in the real world.
When you shift gears, e.g. to talk about science, you can make it easier on the reader by cutting technical explanations for historical or personal anecdotes. This only works once or twice per essay, though.
You can make your paragraphs more exciting. Rather than starting with “An issue similar in cause to separability is the idea of the frontier,” and then have the reader go in with the mindset that they have to hear about a definition (English professors hate reading about definitions), try to give the reader a very concise big-picture view of the idea and immediately move on to the exciting applications, which is where they’ll learn the concept.
I’ve posted this before but I want to make it more clear that I want feedback.
I want to build a better formalization of naturalized induction than Solomonoff’s, one designed to be usable by space-, time-, and rate-limited agents, and interactive computation was a necessary first step. AIXI is by no means an ideal inductive agent.
Had a look at your link, but couldn’t make sense of it. Consider writing a proper summary upfront.
This seems an ambitious task. Can you start with something simpler?
Sorry, my writing can get kind of dense.
It doesn’t quite strike me as ambitious; I see a lot of room for improvement. As for starting with something simpler, that’s what this essay was.
If you want people to read what you write, learn to write in a readable way.
Looked at your write-up again… Still no summary of what it is about. Something along the lines of (total BS follows, sorry, I have no clue what you are writing about, since the essay is unreadable as is): “This essay outlines the issues with AIXI built on Solomonoff induction and suggests a number of improvements, such as extending algorithmic calculus with interactive calculus. This extension removes hidden infinities inherent in the existing AIXI models and allows .”
I’m in the process of writing summaries. I replied as soon as I read your response.
You are pretty much the first person to give me feedback on this. I do not have an accurate representation as to how opaque this is at all.
Separate hypotheses are inseparable.
Hypotheses are required to be a complete world model (account for every part of the input).
Conflicting hypotheses are not able to be held simultaneously. This stems mainly from there being no requirement for running in a finite amount of space and time.
How’s that? Every few lines, I give a summary of each subsection. I even double-spaced it, in case that was bothering you.
Your essay was interesting. What did you think of a similar post I recently wrote?
Feedback (entirely on the writing): The first goal when editing this should be to eliminate words from sentences. Use short and familiar words whenever possible. Change around a paragraph’s structure to get it shorter. Since this is for English class, cut out every bit of jargon you can. If there’s a length requirement, you can always fill it with story.
The best lesson of my dreadful college writing class was that nonfiction can have a story too—and the primary way you engage with a nontechnical audience is with this story. Solomonoff induction practically gets a character arc—the hope for a universal solution, the impotence at having to check every possible hypothesis, then being built back up by hard work and ancient wisdom to operate in the real world.
When you shift gears, e.g. to talk about science, you can make it easier on the reader by cutting technical explanations for historical or personal anecdotes. This only works once or twice per essay, though.
You can make your paragraphs more exciting. Rather than starting with “An issue similar in cause to separability is the idea of the frontier,” and then have the reader go in with the mindset that they have to hear about a definition (English professors hate reading about definitions), try to give the reader a very concise big-picture view of the idea and immediately move on to the exciting applications, which is where they’ll learn the concept.
Thanks for the in-depth critique! I haven’t read your post yet, but it piqued my interest.
Also, moving on to the “exciting applications” isn’t very effective when there aren’t any. :I
Bah humbug.