Yuval Avidani
Author
A 3D printer is a machine that builds a physical object from scratch, one thin layer at a time, based on a digital file. That's already cool enough - but the part that actually changed this year isn't the printing itself, it's where the file comes from. Until recently, to print an original object you had to know how to model it in 3D software like Blender - hours of work and an artist's skill. Today, Meshy generates a 3D model from a single sentence: the geometry in about 20 to 30 seconds, with a full texture in under a minute.
Turns out this isn't just an experiment anymore: according to Meshy, as of GDC in March 2026, over 100 million models have already been created by more than 10 million users. Let's break down everything from zero: how a 3D printer actually works, how a file becomes plastic, and how you get from a sentence in Hebrew to an object you can hold in your hand. We'll walk through the whole thing with one small example: a statuette of a hoopoe, the symbol of our newsroom, that we want to print.
The big idea: printing is building, not carving
Let's start with the principle, because it's simple and elegant. There are two ways to make an object. The old way is called subtractive manufacturing: you take a block of material - wood, metal, stone - and remove everything unnecessary, just like a sculptor carving marble until the shape remains. The way a 3D printer works is completely opposite: additive manufacturing - instead of removing material, you add it layer by layer, from bottom to top, until the object is complete.
That's exactly why the field's official name is additive manufacturing. Its huge advantage: you can build shapes that can't be carved - internal cavities, lattices, complex geometry - because we're building them from the inside out. Think of it like building a LEGO tower brick by brick, except the bricks are as thin as a sheet of paper, and each layer bonds to the one beneath it until a whole object accumulates.
Three families of printers: filament, resin, and powder
Now, how do you actually stick layer to layer in practice. There are three main technologies, and the difference between them is simple: what the raw material is made of, and how you harden it.
The first, FDM (Fused Deposition Modeling), is the most common and cheapest, and probably what you have at home or at school. It works like a robotic hot glue gun: a plastic filament goes into a heated head, melts, and comes out of a thin nozzle that moves through space and draws each layer in molten plastic that hardens immediately. The filament is usually a plastic called PLA - cheap, easy to work with, and perfect for learning.
The second, SLA (Stereolithography), goes for precision. Instead of filament, there's a vat of a light-sensitive liquid called resin, and a precise laser hardens it point by point. Wherever the light hits, the liquid turns solid. Think of it like drawing with a beam of light on jelly, and wherever you touched hardens. The result is much smoother and more precise than FDM, which is why it's the choice for figurines, jewelry, and dental models.
The third, SLS (Selective Laser Sintering), is the powder method. You spread a thin layer of nylon powder, and a laser melts (sinters) the grains exactly in the right spots so they fuse together. And here's the clever part: the powder that didn't melt stays around and supports the object during the build, so you can print complex shapes without support structures - and that's exactly why SLS is popular in industry.
From file to object: how the digital model becomes plastic
So we have a printer. But it doesn't understand "hoopoe" - it only understands movement commands. Between the idea and the object there's a short, fixed chain, and it's worth knowing, because that's exactly where AI is about to step in.
Everything starts with a 3D model: a digital representation of the shape, built from a network of small triangles called a mesh. The more triangles there are, the smoother the surface. This mesh is saved in a file format, and the most common one for printing is called STL - a file that describes the outer shell of the object as a collection of triangles, with no color and no material, just shape.
Now the critical step: slicing. You take software called a slicer - for example Cura, PrusaSlicer, or Bambu Studio - and it cuts the model into hundreds of thin horizontal layers, and translates each layer into a language the printer understands: G-code, a list of commands that tells the nozzle exactly where to move, how fast, and how much material to release.
The slicer is also where you set the parameters that determine whether the print succeeds. Layer height (the thickness of each slice - lower gives a smoother object, but slower). Infill (how much the object is filled inside; usually you don't fill it completely but rather 10 to 20 percent of an internal lattice, which saves material and time without giving up strength). And supports (temporary scaffolding that holds up parts sticking out into the air until they harden, and which you remove at the end).
The real obstacle: it's not the printing - it's creating the model
Now to the point that, to me, is the heart of the story. For years, the barrier to entering the world of printing wasn't the printer - a good FDM printer today costs less than a mid-range phone. The barrier was the first step: where do you even get the 3D model from.
Anyone who wanted an original object had to learn to model in software like Blender or Fusion, and that's a real skill that takes months to acquire: understanding meshes, topology, proportions, how a surface curves. The alternative was downloading a ready-made model from sites like Thingiverse, but then you get what someone else already made, not what you imagined. And that's the gap that closed the door on most of us: we knew how to print, we just didn't know how to create what we'd print.
The solution: AI that generates the model from text or an image
And here come two names worth knowing: Meshy and Tripo. Both do one thing, amazing in its simplicity: you give them a sentence or an image, and they return a full 3D model - a mesh of triangles with color and texture - in under a minute.
How does it work under the hood? Roughly, the same idea behind image generators, just in 3D. A model is trained on millions of description-and-shape pairs, and it learns the connection between words and geometry. Tripo even unveiled, in March 2026 at the GDC event, an original 3D diffusion model (H3.1) that generates the shape directly - not through intermediate images.
What can you feed them? Three types of input. Text (text-to-3D - "a hoopoe statuette with spread wings"). A single image (image-to-3D - you upload a photo, and get a volume from it). Or several images from different angles (multi-view - front, side, and back, so the model doesn't "invent" what it can't see). And the output exports exactly to the formats we already know: OBJ, GLB, FBX, and also STL for printing.
Let's put some numbers on the table, because that's where you feel how mature this really is:
From the AI's model to a real object: the catch nobody tells you about
But wait, before we run to the printer - there's an important catch, and this is exactly where most people fail on their first print. A model generated by AI looks perfect on screen, but it's not always "watertight" - meaning fully closed, without tiny holes in the mesh.
Let me explain the term, because it's critical. A watertight mesh (also called manifold) is a mesh where every edge belongs to exactly two faces - like a vessel you could fill with water without it leaking. A printer must know precisely what's inside the object and what's outside, so a single tiny hole in the mesh - the most common flaw in AI-generated models - can ruin the entire print.
The good news: these tools know how to fix this themselves. Meshy has a "printability check" that closes holes and returns a watertight mesh automatically. After that, two more things are still a must to verify: setting the actual size in centimeters (the model has no idea whether our hoopoe is finger-sized or hand-sized), and checking minimum wall thickness - about 1.2mm for FDM, about 0.5mm for SLA - otherwise the walls are too thin and will crack. Only then do you export the STL, slice it, and print.
And it doesn't end with the model: AI across the whole printing chain
And here's the surprising part: generating the model is just one place where AI enters printing - there are two more, no less important.
The first is called generative design, or in engineering terms, topology optimization. Instead of an engineer sketching a part and trying to shave weight off it by guesswork, we define the conditions for the AI - where the forces act, where the connection points are, how much weight is allowed - and it generates shapes a human wouldn't have thought of on their own: bone-like structures and lattices that are strong exactly where needed and hollow everywhere else. Strong and light at the same time, and exactly the kind of thing only a printer can build.
The second saves us expensive failures. An AI-powered camera watches the print in real time and catches faults before they ruin everything. The famous failure is called "spaghetti" - when the print detaches from the build plate, and the plastic keeps smearing into the air like a pile of noodles. A system called The Spaghetti Detective (today Obico) was a pioneer in this kind of detection, and modern printers like the Bambu Lab X1-Carbon already include built-in spaghetti detection and first-layer inspection, with reported accuracy above 90 percent. The printer simply stops on its own before 8 hours of material go up in smoke.
Bottom line
To me, what's happening here is similar to what happened with video and text: the creative part, which used to belong only to experts, suddenly opened up to everyone. The hardest thing about 3D printing was never the plastic - it was the gap between "I have an idea" and "I have a file." Meshy and Tripo close exactly that gap.
But let's be precise, because there's no magic here either. AI models are great for figurines, characters, accessories, and toys - but for precise engineering parts, ones that need millimeter-accurate dimensions or have to fit with other parts, you still need real modeling or CAD software. Besides that, there's almost always cleanup work afterward: closing holes, scaling, checking wall thickness. AI shortened the path from months to minutes, but it didn't eliminate the need to understand what you're doing.
So here's the question I'll leave you with: if today any sentence can turn into a physical object on your desk - what's the first thing we're going to print?