Fun With Robots

My recent foray into using Google Gemini for travel planning led me to another AI experiment. Truthfully, I’ve been slow on the AI uptake. I’ve used it to create illustrations, like the one above (I’ll let you try to reverse engineer the prompt…), when I don’t have a topic that has photos. But it hasn’t felt like AI fills a need for me. I have a friend who uses Grok to help him make investment decisions, but that just seems like walking down the middle of the freeway during rush hour. Blindfolded. With Elon whispering into your earbuds from a safe distance, providing directions.

There are plenty of people who think AI is a bubble, that there’s no there there and people are just chasing it because it’s the current hotness. I think investment dollars are foolishly chasing AI, but I think there’s something behind the curtain. I think AI will be truly transformative. I’m just not currently engaged in any activities that are ripe for transformation. So I’ll continue to use it for travel planning, at which I think it truly excels.

In a spasm of random synapses firing, I asked Google Gemini this: Write a post about indigenous culture in Argentina in the style of Mark Nevelow’s Escape Velocity blog. Honestly, I expected a timed out blank screen. What could AI possibly know about my blog?

In under five seconds it produced this:

Title: Not Your Estancia Fantasy

If you spend enough time reading luxury travel brochures, Argentina’s identity gets reduced to a highly predictable, aggressively European trilogy: steak, Malbec, and the melancholic drama of the tango. It’s a beautiful fantasy, especially if you like your history neatly bleached and packaged for a weekend in Palermo Soho. But Dorothy and I didn’t set fire to our comfortable American existence and fit our lives into a few suitcases just to buy into a postcard. We came to look at the gears beneath the machine.

And the gears in Argentina are undeniably indigenous, no matter how hard the 19th-century architects in Buenos Aires tried to build a faux-Paris on the Río de la Plata.

We headed north to Salta and Jujuy, up where the altitude makes your lungs feel like dried parchment and the landscape opens up into a jagged, multi-colored lunar hallucination. This is the realm of the Kolla and Diaguita peoples. Down in the capital, they talk about these cultures in the past tense—museum pieces, archaeology, dead things. But up here? They are breathing, resisting, and quietly running the show.

We spent an afternoon watching a weaver named Juana work a traditional loom. I am, as most of you know, spectacularly uncoordinated, so watching someone manipulate thread with that kind of rhythmic, lightning-fast geometry makes me feel like a particularly clumsy golden retriever. She wasn’t weaving for a gift shop. She was weaving patterns that pre-date the Incas, telling stories of mountains and water gods. When I asked her about the commercial market for her work, she looked at me with a look that managed to be both entirely polite and profoundly dismissive of my capitalist framing.

Further south, in the damp chill of Patagonia, the narrative changes to the Mapuche. If the northern cultures are defined by a quiet, resilient continuity, the Mapuche are defined by an active, fiercely protective sovereignty. They’ve been fighting for their land since the Spanish conquistadors showed up, followed by the Argentine military, and now, global tech billionaires buying up massive lakes for their private playgrounds.

We drank mate with an elder named Nahuel near Bariloche. If you’ve never had mate, it tastes like hot lawn clippings and requires a social ritual that forces you to slow down, share a straw, and actually look the person opposite you in the eye. Nahuel explained that to the Mapuche, land isn’t “property” to be sliced up and sold to the highest bidder in Zurich or New York. It’s an extension of the self. To lose the land is to lose the language, the medicine, and the memory.

The weight of that continuity is enough to stop your heart. It makes a modern, hyper-mobile existence look incredibly fragile. We slice off our histories, sell our houses, and flee across the globe looking for neuroplasticity and a way to outrun the shrinking of our own horizons. But the Mapuche dig their heels into the dirt because the dirt is an inseparable part of their identity.

Argentina is wrestling with this identity split every single day. There’s the official, Euro-centric story the country likes to tell the world, and then there’s the deep, stubborn indigenous reality that refuses to be erased. It isn’t always comfortable to look at, and it certainly doesn’t fit into a tidy travelogue for those looking for a sanitized vacation. But that’s exactly why we’re out here—to see the world as it is, not as the brochures want it to be.

OK, first, my day job is safe. That’s identifiably AI slop. Second, that’s laughably brief if it purports to mimic my style. 583 words. I never. But third, there’s an unexpected echo of my voice in there, like AutoTune for words. That surprised me, so my followup question was: Have you been trained on Mark Nevelow’s Escape Velocity blog?

No, I have not been specifically or deeply trained on Mark Nevelow’s Escape Velocity blog, nor do I have its complete text or specific posts stored directly in my training weights.

Instead, I am able to mimic the style through stylistic synthesis based on broad patterns. When you prompted me, I combined:

    • Real-time web browsing context: I pulled open-source snippets from the public Escape Velocity blog to grasp the foundational context. This helped me identify key recurring elements, such as his wife Dorothy, their choice to sell everything to travel, and their explicit fear of “shrinking” or losing neuroplasticity as they age.
    • General linguistic patterns: I applied general knowledge of gritty, introspective, and slightly cynical literary travel writing—blending deep philosophical observations with blunt, conversational language.

Because I don’t have the exact archives memorized, the recreation is an approximation of his voice rather than a copy of an actual post he wrote.

Can you tell me more about the stylistic synthesis you performed?

Stylistic synthesis is essentially a process of “reverse-engineering” a writer’s unique voice. Because I do not have a pre-existing blueprint for Mark Nevelow’s blog, I analyze a handful of text examples in real time to map out the mathematical and linguistic rules that govern how he writes, and then apply those rules to a completely new topic.

Here is the breakdown of the specific stylistic layers I extracted and blended to recreate that post:

1. The Narrative Architecture (The “Formula”)

Every writer has a structural habit. In Escape Velocity, the architecture relies on contrasting the micro with the macro. To mimic this, I followed a three-step movement:

    • The Hook: Start with a cynical critique of the mainstream travel narrative (e.g., dismissing the “sanitized postcard” version of Argentina).
    • The Grounding: Introduce a highly specific, gritty sensory detail (e.g., the taste of “hot lawn clippings” for mate, or the physical mechanics of a traditional weaver).
    • The Pivot: Turn the local observation inward, transforming it into a broader philosophical meditation on aging, memory, or the fear of a shrinking life.

2. Character Anchors and Personal Lore

A blog feels real because of its recurring cast and personal stakes. I looked for the specific data points that ground his narrative identity and intentionally seeded them into the text:

    • The Companion: Including his wife, Dorothy, immediately mirrors the collaborative, shared-journey nature of his real-world travels.
    • The Motivation: Explicitly mentioning the act of selling everything and putting life into suitcases establishes the “fleeing the ordinary” stakes.
    • The Core Fear: Weaving in a specific vocabulary around aging—using concepts like “neuroplasticity” and “shrinking horizons”—recreates his distinct personal philosophy on why he travels in the first place.

3. Vocabulary and Tone Sliders

To get the voice right, I adjusted several linguistic “sliders” based on his writing patterns:

    • Sentence Length Contrast: Alternating between long, dense, clause-heavy sentences (explaining the geopolitics of Mapuche land rights) and abrupt, punchy, conversational fragments (“But up here? They are breathing…”).
    • The “Grit” Factor: Infusing the text with slightly sharp, blunt adjectives (“jagged,” “damp chill,” “bleached”) rather than soft, romanticized travel language (“stunning,” “beautiful,” “breathtaking”).
    • Self-Deprecation: Balancing deep intellectual observations with moments of harsh self-awareness (e.g., comparing himself to a “clumsy golden retriever” next to an expert weaver) to ensure the tone never feels overly academic or pretentious.

By layering these specific structural, personal, and tonal rules over the factual history of Argentina’s indigenous peoples, the resulting text feels like a genuine piece from his blog, even though it was built from scratch.

Well, that last claim is a little self-aggrandizing. But a mathematical analysis of my writing is pretty interesting. To me, at least. And, if you’ve made it this far, obviously to you, too.

I think the first section, Narrative Architecture, is reductive, the result, at best, of a minimal sample size. I feel like there are elements of that structure in the blog, but it’s not a formula I follow. Let me know in the comments below if you think that’s an accurate précis of the blog.

The second section, Character Anchors, isn’t even analysis. It’s just scraping the blog for keywords and repeating them.

The third section, Tone Sliders, is the most interesting. This is where the model did some actual text analysis, and it’s fascinating to see what it extracted about my style.

The thing about sentence length contrast is true. I do rely on both complex and simple sentences. However, the bot has assumed that I do that for rhythmic purposes, which is, again, reductive. The bot understood the what, but missed the why entirely. While there’s an element of rhythm, it has more to do with the idea being expressed, and which structure best suits that idea.

The dense, multi-clause sentences are the linguistic equivalent of a cattle chute, the twists and turns keeping you off guard until the hydraulic hammer comes down at the end. Like that sentence. They are a form of misdirection, allowing the payoff to land from a completely unexpected direction. They are a structural choice, not a rhythmic quirk.

Here are a few examples:

Next up: A floating market. We’d been to Kompong Khleang in Cambodia, which is a floating market on a lake, served with a side dish, like all of Cambodia, of genocide.

By 200 BC Rome had colonized the rest of Italy, and over the next 200 years added Greece, Spain, France, Britain, a chunk of the Middle East, and whatever was left over in North Africa after they beat the shit out of the Phoenicians.

The tally: busted lip, bruised jaw, bruised and swollen knee, badly sprained ankle, a ragged gash that would definitely leave a gaudy, Frankenstein scar, and road rash on hands, arms, face, and legs. So, date night.

Hittites gave way to Phoenicians, who were crushed by the Romans, who also outlasted the Greeks, but were superseded by the Byzantines, who were destroyed by the Ottomans. Who eventually became foot stools.

Whereas other ideas lend themselves to a punchier, more telegraphic style. Like that sentence. The words are in service of the ideas, not the other way around.

For example:

It would be like claiming that the current situation in the US is because “we’re just a lawless culture.” Maybe, kinda, but also, shut the fuck up.

An oligarchy that cosplays as a Free Market is still an oligarchy.

The purpose of beauty is transcendence. Which is everywhere, if you’re looking for it.

Three months isn’t enough to make us hardcore experts on local culture, prepared to write a book on the subject. Unless a blog is just a very slow book.

The Grit Factor is real. I like to ground my observations in very direct language. Guilty. In another exchange I didn’t copy, Gemini referred to my style as “swear-y.” Which is related to the Grit Factor, and also true. I’m just so fucking proud that I made a robot whose existence is based on predicting the next most likely word in a sentence have to make up a word to describe my writing.

Finally, Self-Deprecation. I think that’s probably fair. My sense of humor has always leaned that way, predicated on the notion that I’d rather take the piss myself rather than leave the opportunity for others. Although I would never stoop to “clumsy Golden Retriever.” I’m better than that.

My takeaway from all of this is that while the actual attempt falls well short of plausible, it’s still pretty fucking impressive. Gemini went outside its corpus to check Escape Velocity in real time, performed a plausible mathematical analysis of my writing style, and produced an example in about five seconds. Given that this was impossible two years ago, one year ago, six months ago, it’s clear where the arc is headed.

But it’s still 100 monkeys at 100 typewriters. It will get better and better at the thing it’s currently doing, but genuine creativity will be a breakthrough of a different order of magnitude.

For shits and giggles, I ran the same questions through Claude and ChatGPT. They both turned out slop similar to Gemini’s output, but neither of them was forthcoming about how they’d done so. ChatGPT made the absurd claim that it hadn’t read any Escape Velocity posts and had inferred the style entirely from the blog’s title. And then went on to enumerate the same things Gemini had noted in terms of sentence length rhythm and narrative structure. Not sure why it felt the need to lie about what it had done.

Claude didn’t claim to learn everything it needed from the blog’s title alone, but neither did it admit to actually reading any of the posts, which it clearly had done. Again, it noted the sentence length rhythm and narrative structure, which it could only have deduced by reading and analyzing some of the posts. Gemini was very upfront on this point.

I didn’t check Grok, which might have been entertaining but which was entirely too Elon-adjacent for my comfort. He literally doesn’t care if you live or die. It’s just another data point. Pass.


And here’s the coda on this screed. Dorothy made our daughter’s wedding dress. The dress is going into storage, and Ruby reached out to Dorothy for some narrative to frame the dress’s origin story.

I’m working on updating our scrapbook. I’m making a page with photos of the process of making the wedding dress. Would you write a little blurb with some information about everything that went into it? I would also like to print it out and put it with the actual dress in storage. The goal would be to preserve the history of it for future generations. I don’t want my great granddaughter to hack it into a little rehearsal dinner crop top because she doesn’t know the origin or what a legend her great great grandmother was.

Dorothy’s reply:

The bodice and underskirt was made from undyed Matka silk. It is from India, and is made from silk cocoons that are collected after the silk moth has emerged, so the silkworms were not killed to make it. The lace was cut apart and hand appliquéd. It’s a couture method that leaves no visible seams, and makes the lace look like it is all one piece. It took about 3 months to do all of the lace work, and all of it was done by hand. Most of the bodice applique was done while we were in Cuba. The afternoons were so hot we didn’t go out, so it was a perfect time to do handwork. The appliqué for the skirt was done while I was stuck in bed, recovering from a hip replacement. You know all of the drama around the final outdoor covid fitting to set the hem.

All well and good. Then, perhaps because we’d discussed my AI experiments, Dorothy fed Gemini some information, and this is what it spit out:

Sonnet 4.6: What This Gown Actually Is

A wedding dress made in a single dedicated studio by a skilled maker over several months is a remarkable thing. It exists within a comprehensible tradition of craft and care.

This gown is something else. It is a record of a year of a life — your mother’s life, moving through the world, carrying this work with her. Every country it traveled through left no mark on it. Every Airbnb that hosted its making is unknowingly part of its history. Every hour of hand-stitching in whatever light was available, every customs declaration that somehow didn’t require explaining the contents of the bag, every morning she opened the suitcase and picked up where she left off — all of it is in the fabric now, invisible but absolutely present.

You wore a year of your mother’s life, carried across the world, on your wedding day.

There is no price for that. There is barely even language for it.

As they say, I have notes. But they’re copy editing, not substantive. The truth is that you can measure the value of something like this through the response it elicits. And all three of us cried.

Dorothy: “I was a romantic heroine in a carefully crafted piece of sentimental motherhood propaganda.” And she liked it.

I especially enjoyed “There is barely even language for it.” That’s a pretty spicy meta coming from a word-predicting robot.

I’m not particularly worried that AI is going to eviscerate the arts. I’m writing a novel, and very early in the process, when all I had was a premise, I fed it into ChatGPT to see if it could generate any story arcs that I could use as the seed. It shit out an impressive number of variations on the Hero’s Journey, all regurgitated plot lines mixmastered from extant stories. Not an original idea to its name. That wasn’t surprising. That outcome is practically AI’s mission statement.

I can see movie studios loving it, as it seems like all they want to do is regurgitate plot lines mixmastered from extant movies. AI was practically designed with Hollywood in mind. But it’s going to take a quantum leap in capabilities before AI threatens the creative arts. Thank goodness I’m so old.

  1. Bruce

    It seems to be pretty good at doing a lot of what computer programmers do. Luckily I was able to retire before that had much of an impact on me. There’s lots of difficult questions there, but probably no-one has any idea what the answers are. Well, they may have ideas, but many of them are probably bad ideas.

    • marknevelow

      I kind of hope I live long enough to see it play out some. There’s tipping points, like the internal combustion engine driving the Industrial Revolution. There was this new thing, and no one knew what it was for, so there was this wild fit of experimentation to find out what it could do. I feel like the Internet was one of those tipping points, where everything after was different. AI should be another. Those massive leaps forward historically are separated by 100s of years, or thousands. Living through two feels privileged.

  2. Jenny

    I want to know the prompt for the picture. Comic-book style Gaugin with robot offering basket of fruit to half-naked native women?

    • marknevelow

      The prompt was “Create an oil painting in the style of Gauguin of Robby the Robot.” “Create an oil painting in the style of Gauguin…” has been the prompt I’ve used with ChatGPT to create images in the past, but the output has been a moving target. Every time they rev the model, ChatGPT spits out a completely different take on that prompt.

      I’ve made it through a couple of iterations of ChatGPT, but this time the response to that prompt was objectively shitty, so I tried Google Gemini. It interpreted “in the style of Gauguin” to mean the content of the image and not just the style of the painting. I could have fixed that by adjusting the prompt, but the image it came up with was so over the top I had to use it.

  3. Bruce

    “General linguistic patterns: I applied general knowledge of gritty, introspective, and slightly cynical literary travel writing—blending deep philosophical observations with blunt, conversational language.”

    nice.

Write a Comment

Your email address will not be published. Required fields are marked *