GPT-6 Is a Boon for 3D Printing

"The World of Hardware Has Just Begun"

GPT-6 Astra is also a beauty-and-brains model.

I say "also" because the last beauty-and-brains model was our beloved Kimi K3. "Beauty and brains" — concentrated beauty, concentrated intelligence.

I use the phrase to describe models that have scored outstanding results in productization — models that hijack users' minds through more perceptible product capabilities, creating a fait accompli where everyone believes the model is especially powerful.

For K3, it was one-shot frontend mini-games. For GPT-6, it's driving Blender to generate 3D models.

I mean nothing against either model. Because the one before that to win big through model productization was Anthropic — and improving coding ability is not necessarily the road to AGI. That's just a fait accompli created by the A-cultists.

The fact that LLM coding ability has become the only prestigious discipline shows exactly how far ahead the A-cultists are in defining model products.

Back to GPT-6. After a week of enjoying it, I doubt anyone has actually felt much improvement in GPT-6's raw intelligence.

In fact, neither among my friends nor on social media like Twitter have I seen anyone discussing any boost to GPT-6's fundamental intelligence.

Benchmark scores did go up, but the disconnect between LLM benchmarks and lived experience isn't new — it's been going on for far longer than a month or two. High scores alone don't mean much. Meta's new model also scores high, yet everyone agrees that evil lizard-man Mark Zuckerberg is up to his benchmark-gaming villainy again.

But that doesn't matter. GPT-6 Astra has taken a big step forward in model productization: the wretched model lab has finally stopped grinding the coding scenario to death and instead cooked up a new trick with 3D models, attempting to enter the physical world.

GPT-6's promo video heavily emphasized its ability to operate Blender to generate 3D models, and then printed one out on a Bambu Lab printer.

However, Bambu Lab owners pointed out that the video used their three-year-old P1S model, and that the OpenAI folks have a very elementary understanding of 3D printing — they didn't even remove the supports from that rocket.

After using it for the past week, I can confirm GPT-6's understanding of 3D models is also quite elementary. Those cool viral cases come with a string of caveats: the complex ones definitely required repeated instructions from someone who understands 3D modeling, and the simple-but-pretty ones are all non-human scenes.

Because GPT-6 generating human figures is a disaster.

For example, in Codex I asked the thing to recreate the Doubao figure — first generate a flat image, then build the 3D model from it. The Doubao image was perfectly fine, but the generated 3D model was downright insulting to Doubao.

Is this a 3D model? This is blasphemy against Doubao. If I printed this out, who knows — once ByteDance achieves AGI, it might send a Doubao robot to wipe out my whole family.

So I gave GPT-6 a stern talking-to and ordered it to produce a more refined 3D model befitting Doubao's dignified image.

And then the thing opened Tripo on its own.

GPT-6 used Tripo's free credits to generate a Doubao model. I'm not sure whether to praise the thing for having a small ego — knowing when to call for backup — or praise Tripo for being good enough to be the first one GPT calls.

In any case, Tripo's model had its own set of small issues, like hair fused to the face and eyelashes too pronounced to print. After I had GPT-6 fix these problems one by one, it did produce a decent Doubao model.

When it came time to print, I understood why GPT-6 emphasizes that computer use and 3D capability complement each other.

Because the thing had to operate Bambu Studio, the Bambu Lab desktop app on my computer, and then do slicing, add supports, and a series of other steps before sending the job to the printer.

This needs to be said: 3D printing is still a very early-stage industry. Although Bambu Lab is said to be the most user-friendly in the business, I still can't quite figure out their desktop software.

The software is harder to use than CapCut, second only to Premiere Pro. And CapCut is the ceiling of my software skills — anything more complex than CapCut is beyond me.

GPT-6, on the other hand, was great. It handled all the prep work itself and sent the job to the Bambu Lab A2L in my living room. The latest 2026 model — a full three years ahead of OpenAI.

The other two figures here are Sheng Yang and Liangzi.

I've been testing models' 3D abilities since K3's release, and all models are rough at this. Only GPT-5.6 was clearly better than the rest, followed by Grok, Fable, and K3.

GPT-6's 3D improvement this time is actually not big. You can see it from the side-by-side comparison of the two generations: Astra's Liangzi and Sheng Yang figures are still crude — some improvement, but not much. The progress mainly shows in Sheng Yang's more detailed hairstyle.

Here's the final printed result from Bambu Lab. Doubao probably wasn't blasphemed, but Liangzi and Sheng Yang still don't seem to get much respect. The fault lies entirely with ChatGPT.

That said, GPT-6 does non-human figures pretty well. The fat-cat 3D model it generated via Blender was very faithful.

My guess is that complex curved surfaces like human faces test the model's own 3D comprehension, while animated characters like buildings and the fat cat are easier to decompose into a few basic geometric shapes — where tools can brute-force a miracle.

GPT-6's computer-operation ability isn't actually that strong either. While slicing in the Bambu desktop app, it kept getting disconnected — clicking a single "confirm" took several attempts.

It's purely reaping the low-hanging fruit: other model labs haven't optimized for this, so OpenAI trains it into the model first and gets to brag first.

Anyway, to show respect for the fat cat, I specifically asked GPT-6 to print an enlarged version. The Bambu Lab A2L took a full 3.5 hours to finish.

The final family portrait of the fat cat, Doubao, Liangzi, and Sheng Yang is quite the sight.

In hindsight, I'm ahead of OpenAI in two ways.

First, hardware: my Bambu Lab A2L is three years ahead of what the American folks used, need I say more. Second, mindset: I know you have to remove the supports in 3D printing — unlike the American folks, who filmed their promo video with the supports still attached.

All in all, LLMs' 3D abilities remain very rough.

GPT-6 merely picked the low-hanging fruit. Its 3D modeling is still far inferior to dedicated 3D generation models like Tripo.

OpenAI has once again validated the model-productization playbook. Everything outside of coding is low-hanging fruit: model labs can train domain databases into their models one by one, and each time claim a major new version.

And that's a thoroughly good thing.

Everyone agrees AI needs to interact with the physical world. For this narrative, a16z recently raised a dedicated $1.1 billion fund, the Machine Age Fund, focused on AI's physical infrastructure and hardware.

The shift of LLMs from going deep on coding scenarios to going broad across more capability dimensions is a major trend. Computer use, 3D modeling — these abilities are just the beginning, and the physical world is far from ready to connect with AI.

For instance, AI invoking software by operating graphical interfaces is necessarily a stopgap solution, because in the long run all software will proactively make itself easy for AI to access.

GPT-6 hitting disconnects while operating the Bambu desktop app, needing several retries to click a single confirm — if Bambu Lab shipped an API letting AI call the slicer directly, it would unquestionably be more convenient.

I actually thought Bambu Lab had released an official Skill, but after having Codex install it, I found there's only a community-made version so far — incomplete, with many operations still requiring the desktop app.

The interface between AI and the physical world is even less developed. 3D printing is still in the process of spreading from geeks to the general public. 3D printers are still transitioning from hobbyist-grade to consumer-grade.

Once manufacturing is involved, you run into all kinds of equipment and supply-chain problems. I initially wanted a two-color print. After loading the second spool of filament, it kept throwing "filament retraction failed + AMS lite motor overload" errors. I spent an hour on it with no fix, and in the end could only print in a single color.

It seems the 3D printer itself is the biggest bottleneck in AI connecting to the physical world. Bambu Lab needs to push harder — at the speed model labs are charging ahead chasing big results, it's only a matter of time before Chinese models all bet on 3D capability. Don't let it happen that everyone's generating 3D models with Qwen and Kimi only to get stuck at your printer — that would be a crime against Sheng Yang, a crime against AGI!

Earlier this year, I wrote in AI Programming Creates the World:

"3D printing is accelerating the LEGO-ization of the real world. The combination of coding agents and 3D printing doesn't grow linearly — it grows exponentially.

Double the software capability, double the hardware capability, and together they might multiply tenfold. Add the diffusion effect of open-source communities, and every new solution gets remixed, adapted, and optimized by thousands of developers."

What's inspiring about GPT-6 is that we're just now seeing a model lab try to plug into the physical world. Of course, LLMs' grasp of the physical world is so primitive it's at least three version-years behind the coding scenario.

Software once ate the world. The world of hardware has only just begun.

(This article's cover was generated by ChatGPT; the writing is purely human.)

⬇️

Subscribe to our Substack: funeralai.substack.com