Okay, let's do this thing...

Premise: If a concept can be easily conveyed in an academic setting then the labor required to write it is too boring, too undignified, for a human to perform merely to ensure its conveyance. Instead, to expect a human to write it, the concept itself can no longer be the only thing being conveyed. What must also be conveyed is the sharpness of the concept, exposed through the literary talent and humanity of the writer.

I have a stupid fucking problem. Right now I am floating in the middle of the ocean on a big boat. This is my first vacation in years and it is going poorly. The food is bad, the port stops are being cancelled due to apocalyptic weather phenomena, and it seems there are more mobility scooters and walkers on board than passengers under forty. But none of these are my stupid fucking problem. Rather, my problem is that I must spend part of my miserable time on this earth endeavoring to be accepted on LessWrong as a full writer because the CEO of my company wants me to be able to publish the groundbreaking alignment research that we work on in our internal alignment lab in a place where it will be publicly accessible to thousands of researchers. It is a completely instrumental problem subject to completely instrumental reason. How boring.

However, because of the pace of development I don’t really have the time to write documents myself. The development of complex scientific concepts does deserve some care but it’s increasingly difficult for me to believe that the care they deserve lies in the human provenance of the writing of the academic paper itself --- if it ever did. Most papers are skimmed by most readers most of the time, and while the details can absolutely matter, and some rarefied academic writing is knock-your-socks-off beautiful, most of the work of the university is a university discourse: a very dry thing indeed.  

The reality is that an academic paper of yesteryear that simply conveys the concepts could be completely written, at the level of technical prose, by your favorite AI agent of the day, and still be a totally legitimate piece of scientific work. The ‘intellectual labor’ that makes a modern document scientifically valid instead lies in the scientific taste and hunch that prompted the execution of the experiment, the editing of the document after the fact, and the questioning of the document that occurs alongside its editing. Every single word can be written by an AI while the truly significant labor that actually beatifies it as ‘science’ lies in the process of questioning the model and its artifacts: ‘In this experiment did you properly run this statistical analysis? How could you prove to me that you did? Does this code appear to do what the model said it does? What about these experiments over here, the numbers look too smooth, what is going on? What does a different model say about these results? There appears to be a contradiction? Has this already been done in the literature? Are you certain about the reasoning in this section?’ etc.

Actually physically writing a mere technical report, at this point, is pointless.

And so we return to my stupid fucking problem. I can’t just submit some of the work from my lab if it is written by AI because LessWrong doesn’t want AI submitted documents. However, I don’t really have time to write anything by hand unless it is something of an Œuvre de passion, something I care about beyond its instrumental necessity in the grand project of not having every human on earth killed. Also, I have to be careful in a more extensive sense --- some of the work is valid but should not be released because it is dangerous, or it is dangerous and invalid (a strange but nevertheless possible combination) and while I do have time to make that distinction, I don’t have time to do that and juggle writing full documents. Not at the pace we must work.

Ah, but Œuvre de passion --- I have some of those. Maybe I’ll submit my top shelf shit and get published that way!

First I try my speculative work on the nature of time --- it’s only 150 pages or so, a lot of it is written by AI, but I figure if I pair it with a handwritten note explaining what it is and why I think it’s useful and simply point at it, without making it the body of the post then that will be interesting. In this way I am being very clever you see --- my post is about a longer document I’ve written with AI, not actually itself written by AI.

Response comes back... too many yellow flags.

Okay, maybe another attempt: I know my work on death and immortality! It is a meditation on five dialectical-logical movements through the concepts of death and immortality and on what basis we can expect to die or live forever along with an opening that explains, somewhat, what a dialectical-logical movement is. In just five easy steps even! I can’t really go into the deepest detail, of course, because actually understanding dialectics, à la Hegel, is the kind of thing that drives grad students insane, but I can give strong impressions. Some outlines. Some significant brush strokes.

Too many yellow flags.

Okay maybe if I comment I will get some love? I can submit three a day, maybe I should read some posts. Here’s one about how Eliezer seems to think that the use of export driven trade statistics are actually about states using external parties to validate which parts of their economy are actually useful under conditions where they are unable to judge internally due to corruption or ideology. We could say that the simple 20 second idea is ‘states use export statistics to understand their economies under conditions of authoritarianism to get around the informational downsides of being authoritarian’. It's a brilliant concept, but incomplete. There is a history here: Eliezer's idea actually explains both why Keynes's Bancor was clever and why it was never adopted. The Bancor’s implementation required a deeper reimagining of the political-information-economy of the entire world order than was possible under the US-led global regime of the mid-twentieth century. Deeper, even, than Keynes himself had imagined.

I write a nice comment explaining this but fuck!!! A couple of the paragraphs that are too boring to write are written by AI. It’s automatically rejected.

Okay I rewrite it by hand and I post the rewrite. Alongside it I make some other comments about how there might be a flaw in the reasoning of another writer when they talk about their interactions with Chinese civil society and I also comment on how I appreciate the work of another alignment researcher using a span-only method to elicit thought processes from models without corrupting those processes.

My rewrite and other comments sit, for days, now approaching weeks, unmoderated....

I make another short comment. Someone writes about how they built an automated research scaffold. I point out that maybe, alongside their scaffold, they might want to RL some models to specifically look for the most interesting stuff in the infinite stream of nonsense the scaffold produces. At this point I am annoyed and don’t really want to develop in long detail what the concept is, just give enough to say ‘hey, here is a direction you could take this research that might be interesting’.

I get a message, finally a human has moderated an actual comment!

Insufficient Quality for AI Content.

Well at least it isn’t a yellow fucking flag. 

I’m driven to drink. Only half of something called a Punch Romaine à la Carpathia on this godforsaken floating retirement home in the middle of the Caribbean that I’ve chained myself to in the attempt to have a good time before the end of the world. I look out at an ocean that seems to embody the sublime blue. Is my writing really that bad? Am I that incomprehensible? Why can’t my considerable poetic talent ever turn into publishable anything? 

There are other places I could publish, I think darkly. In fact I will, if it gets rejected here. But that won’t be sufficient to satisfy my frustration: my passion demands an œuvre...

Now my mind turns to petty revenge. I’ve been to LessOnline twice. I’ve drunk in the unspeakable beauty of Berkeley. I once kissed a boy at Lighthaven. I exist in the real world too, Goddamnit! I know what I can do: at the next LessOnline I can subject every moderator who rejected me to fifteen minutes of my possibly very annoying company, just long enough that politeness dictates that they actually have to bear annoyance IRL. Bwahahahaha!

Meh, that seems like a lot of effort. I’d probably rather spend my time eating canapés.

Or, if they had actually understood my Death and Immortality piece they would understand that we will all live a very long time. Thus, in the P(doom) scenarios we will not die but instead will have no mouth and yet feel the urge to scream. Perhaps the evil AI will torture them with a perpetual succession of individuals very much like me, but not exactly me, all of whom will annoy them for 15 minutes at a time in a simulated Lighthaven and ensure that they never again will kiss a boy. 

I have written it, so now it is possible and since it is possible, then by the power of the many-worlds interpretation there is a future where it will happen.

Just as, in some unhappy timeline, my personal AI demon will torture me perpetually with the phrase:

Too many yellow flags.