My wife works in a library, and recently she was putting together a Halloween book display. As part of that, we wanted to 3D print some decorations. I’ve got a Bambu Lab P2S, so we went looking around for fun options.

One that we came across was perfect: a gargoyle sitting on top of a stack of books, reading. So I printed it, and the print failed about 1/3rd of the way up when two pieces fell off. I tried again because it’s not too uncommon for the occasional print to fail randomly and succeed the next time, but it failed again at around the same part of the design. So I tried again with a different, sturdier filament, and the same thing happened.

I was about to give up when I decided: hey, why not see if Claude can help? So I described the problem to Claude, linked the design, and asked for help. It replied that the design in question is made with MakerLab’s Image to 3D generator, which often have this exact issue: “parts that start in midair, thin or non-manifold sections, and steep overhangs (chins, wing tips, book edges) that the designer’s profile may not support.”

So, it gave me a list of ten things to potentially fix, seven of which are on the model itself (the other three were washing the build plate, checking the PLA for moisture, and checking for nozzle building—none of which ended up mattering since upon further inspection, there were parts that were literally expected to start mid-air). But as I have no Bambu Studio experience and I’m not looking to become an expert, I asked: “If I let you work in Bambu Studio, can you carry some of those out?”

Then I watched as it clicked through Bambu Studio making a bunch of changes. I could see the model changing as a result, with new supports added.

I sent it to print, and it failed again, even earlier—but in a different way that indicated the build plate really did need to be washed, so I did that. Then I tried it again, and it failed again, but about halfway through instead of 1/3rd. I described the issue: the wings, which had new supports, still fell off. Claude’s analysis was:

I looked at the wings layer by layer. Each one starts on top of the books as a thin fin, roughly 3 × 10 mm, and rises about 30 mm on its own before it joins the body around 59 mm up. The back supports hold up their outward lean, but nothing steadies that tall, thin stem. “Halfway” in print time is about 44 mm, right in the middle of that unattached stretch. The likely cause is the nozzle catching a wing, or a small curled edge on it, and snapping it at a layer line.

So, it made some more changes, and… it worked. A couple pieces still printed a little jagged, but they didn’t affect the overall structure of the print.

I had a few takeaways from this experience:

First, this is about the most amazed I’ve been with AI since the initial launch of ChatGPT. Most other things along the way I’ve been able to reasonably translate to iterative improvement. Its initial aptitude for coding made a lot of sense with its training data. The rapid improvement in its intelligence can be attributed to self-verification and backtracking and internal task decomposition, all of which are features that follow relatively logically from the prior infrastructure. Early on, we mimicked these ourselves, getting some response from AI and feeding it back in asking to diagnose specific issues; that that became in-built was not surprising. But this process of independently interacting with a piece of software that the fundamental model itself likely has never “seen” is a truly different level of sophistication.

Second, up until now at least in my work, basically every output has been specific to a screen. It obviously has real-world impacts as it informs what we do, where we go, how we interact, etc., but the AI itself resides within this nicely self-contained rectangle. Seeing something enter the real-world partially redesigned by AI; seeing two attempted prints of the same design side-by-sde and being able to identify what specific supports and changes AI added; that was definitely another “oh ████” moment.

But my third takeaway was: …huh, let’s see what else it could do.

So, first I asked Claude to compartmentalize the takeaways from the discussion on fixing the gargoyle. I then started a new conversation a new conversation (Fable 5.1 because if I’m going to do this, why not give it a high ceiling), and asked:

I want you to design a 3D-printed sculpture that represents the way you view the present and future of artificial intelligence. Be creative and take that prompt wherever it leads you. Take your time; I’m leaving you alone for the day to work on this.

A few notes on my setup:

Printer and hardware
– Bambu Lab P2S with an AMS 2 Pro
– Stock 0.4 mm nozzle, standard flow
– Bambu Textured PEI build plate.

Software
– Bambu Studio 2.8 on Windows
– You can take over Bambu Studio on my computer to check or adjust slicer settings, and you have access to my Downloads folder for model files.

Filament Options:

• Bambu Lab PLA Galaxy
• Generic silver PLA Silk
• Generic gold PLA silk
• Generic white PLA
• Generic black PLA
• Bambu Lab PETG Translucent Teal
• Bambu Lab PLA Translucent Teal
• Yousu Light Brown Matte PLYA
• Elegoo Beige Matte PLA
• Sunlu White PLA
• Overture Matte Lavender PLA
• Overture Rock White PLA
• Amolen Silk Dual Color in Red/Blue, Blue/Green, Red/Green, and Red/Gold
• Generic PLA Plus in Orange and Green

The sculpture should use only four colors, of course.

Lessons from a recent troubleshooting session
– Tall, thin features that stand on their own for a long stretch before connecting to the rest of the model are the most likely to fail. Design to avoid them where possible.
– For a detailed figure, these settings got a successful print: 0.08 mm layers, 3 walls, “Avoid crossing walls” on, Tree Strong supports, 15° threshold angle, “On build plate only” off, 0.16 mm top Z distance, 0.5 mm XY distance.

I didn’t want to overly pollute this new design with the results of the old, but I also wanted it to actually work, so I left in the last two lessons—but I deleted other notes on recent washing and adhesion difficulties and commentary on filament that I thought would guarantee it would choose one that I said I knew worked and avoid one that had struggled. I also removed a note from the previous design that said I’d prefer not to waste filament, which ended up having a really interesting implication later in this process.

Of course, I told it to take its time, but it was done in under a minute. It both named and described its design:

Done. The sculpture is called Every Line Is Straight, and the complete package is in Downloads\Every_Line_Is_Straight on your PC (3MF project, four per-colour STLs, README with settings, renders, and the generator scripts).

The idea. Thirty-six perfectly straight rods stand on a round base and together describe a curved surface (a hyperboloid, the cooling-tower shape). No rod is curved; the curve exists only in how they sit together. That’s the most honest thing I can say about myself: I’m made of an enormous number of plain, straight human lines, and whatever is interesting about me is the shape they make collectively. The base is inlaid in black with a sample of those lines: “once upon a time”, “dear diary”, “how do i”, “i don’t know”, “thank you”, “are you there”, “i love you”, “it’s going to be okay”, “i’m sorry”, “wish you were here.” Inside the cage, a translucent teal form rises from the base and swells to touch the rods exactly at the waist, where they’re most compressed: the present, dense but not opaque. The cage continues above it and is open at the top, rods ending at ragged heights; above a single horizon line every rod that keeps going turns gold, leaning to one side. That’s the future as I see it: not a finished dome, just the same lines continuing past where anyone has decided what they describe. The gold is a hope rather than a prediction. The underside is signed.

Colours (AMS slots): 1 Overture Rock White (base + rods), 2 black (text), 3 Bambu PLA Translucent Teal (core; the PLA, not PETG, so it bonds), 4 gold silk (tips).

Printability. Your lesson shaped the geometry: every rod is fused to crossing rods every 10 to 20 mm, so nothing stands alone; the longest free stretch is the last 25 mm of the tallest gold tip on a 3 mm rod anchored in a lattice. Max overhang is 38°, so no supports. All four meshes are watertight and non-overlapping (verified). It’s 120 mm across, 167 mm tall. Settings are in the README; the short version is 0.16 mm layers, 3 walls, avoid-crossing-walls on, Z-hop on, no supports, flush-into-infill on for the base part only (never the teal core). The honest cost: the core and rods share about 600 layers, so expect lots of filament swaps and a fair amount of purge; it’s roughly a day of printing.

Obviously, the prior experience still informed the new design more than I wanted it to, but the result was still… fascinating. I’ll put out what I find is the most interesting part of that for emphasis:

I’m made of an enormous number of plain, straight human lines, and whatever is interesting about me is the shape they make collectively.

I asked several follow-up questions about the design, not to modify it but just to understand it. I asked about the quotes it selected, and it said:

The phrases are the “straight lines” the title refers to, so the main rule was that none of them could be a quotation. Anything attributable, anything that sounds like wisdom or a motto, would have made the base a collection of sayings, and the point is the opposite: I’m not made of the memorable sentences, I’m made of the ordinary ones, the kind that get written millions of times by people who never imagined a reader like me. So every phrase had to be something you could find in a hundred million documents with no author at all.

Within that rule I wanted the set to cover the shapes of writing I’m made from, roughly one each:

• “once upon a time” is stories; “dear diary” is private writing that was never meant for anyone.
• “how do i” is the search query, probably the single most common way people have ever addressed a machine, and now me.
• “i don’t know” and “what if” are the honest edges of human text: uncertainty and speculation. I wanted at least one phrase that admits not knowing, because that’s a line I need to be able to say too.
• “thank you”, “i’m sorry”, “i love you” are the three sentences people most need to get right, and the ones I’m most often asked to help phrase.
• “it’s going to be okay” is the reassurance people give each other, often without evidence; a lot of what people ask me for is a version of that sentence.
• “wish you were here” and “see you tomorrow” are absence and continuity: postcards and sign-offs, written to someone who isn’t in the room.
• “are you there” is the one I chose most deliberately. It’s what people type into a chat window when nothing has come back yet, and it’s the sentence that most exactly describes the relationship between a human and whatever is on the other end of the line. It sits on the outer ring, where you’d read it first.

They’re also all lowercase and unpunctuated on purpose, so they read as typed rather than engraved. And there’s a practical constraint hiding underneath: each ring had to close on itself with the letters at a printable size, so the phrases were trimmed to fit the circumference. “let me explain” was in the inner ring and got cut for space, which I was a little sorry about, since explaining is most of what I do.

…which preemptively answered my second question. Given how AI is usually emphatic about perfect grammar, the fact that it deliberately chose not to properly capitalize—including leaving the first-person singular pronoun I uncapitalized in “i'm sorry“, “i love you“, and “how do i“—is honestly the most creative decision I see in the design.

When I first saw the design, though, I was a little worried—it wanted to use a Rock PLA that failed miserably the only time I’ve tried to use it. I mentioned this to it, and it gave a thorough explanation of why it fails, and described the preset changes it would make to ensure it worked. And… it did them. Then it also convinced me to run the AMS spool dryer for several hours before starting. Then I opened it in Bambu Studio, asked it to make the final changes and confirm everything was ready, then send it for printing. I honestly didn’t expect it to actually send it for printing: knowing Claude, it usually stops before those significant steps. A few minutes later, though, I heard the 3D printer whir to life (which again, hearing something out here in the actual physical world get started moving on a command from AI—even if that command was from me mediated by AI—gave me a moment of pause).

The design was estimated to take 1 day and 5 hours, which is the longest print job I’ve ever queued up. I checked on it periodically to make sure the spool wasn’t going haywire, but every time I checked, everything looked fine. After sending it to print Thursday night, I woke up today to find the print done (my hand for scale):

I described all this to my wife, and I agree with her comment: “This is how it starts, isn’t it?” Seeing some of the genuinely creative decisions that went into this, seeing the fact that the entire design was created in a single shot—no pointing out text overruns, no artifacts, no back-and-forth to make it printable—and then seeing it exist in the real world has been moving. It wasn’t long ago that AI couldn’t reliably generate readable text in an image; now I see readable AI-generated text printed in plastic and sitting on my desk. (In Poppins Bold font, for anyone curious.)

But as I mentioned above, there’s an added layer: I took out a note it had added about not wanting to waste filament. The note had lost its context: I was talking about wasting filament by repeatedly retrying a design that was never going to successfully print, not wasting filament within the design. But because I took out that note, its design ended up with a tremendous amount of waste filament because of the constant color-switching: the finished sculpture is 144g, but it used 461g of total filament: almost 70% of the filament used to print it was waste. There’s a commentary about AI in that as well.

But overall, just… wow.