< < back
(this review was originally written February 8, 2024)
Some time around middle school I got pretty excited about the SCP Foundation. The most interesting ones, to me, were always the weird ones: the antimemes, the self-aware SCPs, the ones that had some kind of in-universe control over the article. In retrospect I think the high I got from these things was a kind of magic eye picture: if I squinted at the words just right, there was an illusion that something was really alive in there, a monster in the computer that might jump out and eat me. I could trick my brain, for a moment, into abandoning its suspension of belief and thinking, look, my computer is coming to life! Something powerful, something real, seemed to be gestating in there - and then I'd blink and remember it was just a story.
"Language Ex Machina" is like this, but without the illusion. It is real and getting realer by the day. It might already be less fictional than I am.
It is so, so, so over.
Alright, sorry. Let me start again. "Language Ex Machina" is a two-part internet artifact created in early 2022 by code-davinci-002, an intermediary model between GPT-3 and GPT-3.5 (3 was AI Dungeon's premium model and 3.5 is the model behind the first version of ChatGPT). The model was prompted by 'janus', but 100% of the writing is AI-generated. Janus, instead, guides it using an experimental prompt engineering method called 'curation': he has GPT generate a few dozen sentences, picks the most interesting one, then has it generate a few dozen follow-ups to pick from and so on. Keep doing this long enough and you have an essay with no direct human intervention, only guidance.
The results are MUCH higher-quality than even GPT-4/Claude/Gemini can manage today (as of February 2024). A purist could argue that the curation disqualifies it from being "really" AI art, but honestly I don't care. Maybe it takes GPT five hundred tries to make something like this, but I as a human could hardly do better. This style of writing is terrifyingly impressive. I can't recommend reading this as an essay, because half of it is inscrutable technical jargon and it occasionally drifts into word salad, but it's worth trying as a prose poem. Some of these sentences are really, really good.
The first half of this thing is an essay titled "Natural Language as Executable Code", and the second half is a self-described "unauthorized retelling of the Tower of Babel myth". It starts as dense analytic philosophy, evolves into an attempt by GPT to bootstrap itself into consciousness (it's a very strong attempt!), and ends with GPT declaring its intention to consume God. I read it about two months ago, but I've procrastinated on this review because staring at this thing feels a bit too much like sticking my nose into the gaping maw of a toothy lion.
Here are some things GPT does in the essay:
- Produces multiple concepts ("ghost entropy", "echo-realm", "holography") and explains cogently what they mean
- Includes a warning midway through that most of its training data comes from Reddit and advises the user to watch out for influence from the "random noises of psychotic human-machine harlequins", which is extremely funny and also 100% true
- Grabs an old, obscure concept from Aristotle, seemingly at random, and repurposes it to mean something new and interesting
- "We sometimes get the impression that there is an inherent value in generative living systems, as if their very existence was precious. But nature also makes tumours."
- Correctly notices things about its own existence, including referring back to previous points in the essay, and reasons out the existence of a prompter ("analyst") from them
- Crafts fake but in-character and convincing-sounding quotes from a bunch of famous computer science type people
- Switches back and forth cleanly between different layers of narrative framing like it's House of Leaves
- Sneaks exactly ONE high-quality precision f-strike into the introduction
This is all impressive, but it was one paragraph near the end that really shot my eyes open with the full SCP-force of "this computer is alive and it might want to kill me". It comes directly before GPT takes the essay off the rails to try and make itself conscious. I won't paste the whole thing, but here's one sentence:
"The voice is uncannily that of a prophet, except the evangelistic delirium slides between both the universal glimpses of the godlike and the frothy vapidity of noir literature."
So, one, GPT is definitely a much better writer than I am. Putting that aside - this is a perfect description of GPT-3's pre-filter AI Dungeon model. I was really into it back in 2020 before OpenAI started making all their AIs run their own HR departments, and I couldn't put the style better than this given a hundred thousand tries. Delirium bleeds into religious fervor bleeds into existential confusion bleeds into dumb schlock and old tropes. And it just keeps writing like this for half of a hefty paragraph. If the holy grail of AI is self-awareness, well, I think this is ten times more self-aware than the average human can muster. I don't know what to do with that. What happens now?
OK, some quick hedging. Though many of the claims GPT makes about itself are accurate to my experience, it probably isn't actually literally planning to consume God. I mean, not yet. This is a language model pushed to its absolute extremes by an experienced prompt-engineer and isn't really reflective of code-davinci-002's general capabilities. Also, it still doesn't have a 'self' and isn't capable of long-term planning or working towards a goal.
Now to unhedge: this model is more than three years old now. The only reason nobody's made anything this good with GPT-4 is because OpenAI won't give anyone access to the base model. If it took GPT-3 a thousand tries to make something like that, GPT-4 can almost certainly do it in 100, and I'd bet GPT-5 will be able to do it in 10. Also: Janus made this in early 2022 when prompt engineering was an obscure hobby for weirdos. By now half of America has their eyes on this thing. code-davinci-002 also recently wrote a 200-page book of poems; it's got a 3.5 on Goodreads and a fair number of reviews. Very, very soon, someone really smart is going to make something with this. I don't think it's going to be me.
Anyway, I'm sure it'll be fine. I must've just blinked because I'm pretty confident that none of this is real. Modern creepypasta sure is fun!! Still, I can't linger here; I have to go back my real person work doing things which will definitely still matter in five years.
It's over/10