Swimming in the Pacific Ocean
I knew that ChatGPT4 would refuse to write about direct experience with swimming in the Pacific Ocean, but I asked anyway. The best it would be capable of is a second hand account from the residue of patterns in meaning-language particles it had aligned with parameters during its training. I predicted it would decode and tokenize two parameters: Pacific Ocean and human swimming in it:
This format—a caveat (I’m not human), a clarification (I can tell you what humans have to say), and a generalization (diverse and exhilarating) followed by an organized list of relevant topics—is becoming predictable:
This list went on at some length. I noted that each topic was shaped similarly, beginning with a comment about the ocean followed by a comment about a feeling or sensation. The bot had tokenized the prompt during decoding into two parameters: 1) the Pacific Ocean and 2) feelings humans report while swimming in it.
Juxtaposing comment type 1 and comment type 2 was not done haphazardly during encoding (producing the list of topics). Immensity > awe and humility. Temperature variation > relaxation or invigoration. Powerful waves and currents > thrills and challenges. From prior knowledge, I knew that the bot had encoded patterns of meaning in its list along the tokenized x parameter (the ocean) and scanned for heavily weighted patterns of meaning along the y parameter (human feelings/the ocean). What would the bot do if I changed the task from retrieval of information to communication of information?
I immediately recognized the wonky logic of the topic sentence. Swimming as a “diverse and exhilarating experience” uses two adjectives in grammatical logical coordination (fork and spoon) which are not logically coordinate. I wanted to keep an eye on how the bot would deal with this tortured expression.
The topic sentence is less problematic for me, certainly shorter and more direct. It’s tough to see swimming as a “diverse experience” still. It partitions the experience of swimming in two asymmetric ways: diverse and exhilarating. The idea is right, a tip of the tongue, but the expression is wrong.
Ah. Swimming there “offers” diverse experiences. The parameters are less obvious in the substructure of this iteration, likely because the idea of the experience of swimming is announced as the theme, and it just so happens that differences in sensations of experience are caused by water differences.
Interesting that the bot encoded a second subsidiary topic sentence: “From the expansive vistas… each swim brings its unique sensations.” The first clause articulates diversity, the second main clause unique experiences. Let’s see what the bot could do along the axis of style now that the ideas are more clear.
As I suspected. Like most novice writers, the bot turned to intensifiers and modifiers instead of nominals and predicates to achieve vibrancy. Diverse isn’t enough; now it’s “truly diverse.” “Vast” isn’t enough. Now it’s “vast and breathtaking.” We do get some additional substance—“plunging into the inviting water” (inviting, mind you, not simply water). I knew from experience that the bot takes criticism in stride.
I noticed a subtle change in perspective as the bot encoded iterations of this information under sometimes vague prompting. It seemed to be trying to read me. “Picture yourself” doing something, it said. In this version it directly addresses you, telling you to “navigate,” “discover,” “surrender.” I wondered what might happen if I gave it a rhetorical purpose, an audience, a commission, someone to speak to, someone to speak for:
You be the judge:
One final question for the bot:
Its answer:
