Jonathan Herbert (@jmbherbert) header
Jonathan Herbert (@jmbherbert) profile picture

Geospatial PM & AI slop cannon. Quant dabbler. 3D printer.

Campbell, CA
Joined December 2010
Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)Photo by Jonathan Herbert (@jmbherbert)
Idea -> Model -> AR Experience with Gemini Demo Link: jmbherbert.github.io/schloss… Built from 1 prompt w/ zero manual 3D modeling. How it's made: (1/2)
1
45
(2/2) Carbon first used nano banana to imagine a model, then wrote procedural Blender Python (bpy / bmesh) while a 6-camera visual critic scored every view, hill climbing over 3 rounds and outputted a ~10MB AR .usdz, with <model-viewer> index.html. Served from Github pages.
41
Last week's Los Gatos model used Google imagery - so I wanted to see how Gemini 4 Carbon would do with consumer-grade imagery. I had a set of ~380 images taken in 2020 from a GoPro Max suction-cupped to my car roof. Here's what Gemini was able to do with it (1/5)
5
3
91
9,612
(4/5) For eval, the swarm built a WebGL split-wipe, multi-station 360° viewer against Google 3D Tiles, which let me inspect how well the geometry matched reality (pretty well, surprisingly - but led to some extra prompts to fix up road connectivity + building orientation)
2
1
5
1,158
(5/5) Overall, this showed me that combining aerial + vector + street-level imagery (as in my Los Gatos example) is a clear winner - particularly w.r.t. embedded semantics. That said, Carbon's ability to recreate a 3D scene from a car-roof mounted GoPro is still pretty solid!
1
13
922
Follow-up to my Los Gatos Blender build - tried turning the static 3D block into a ~20s live-action street shot using Gemini + Gemini Omni. Top: animated 3D scene in Blender. Bottom: grounded Omni video output. 🧵 (1/5)
4
2
15
935
Then Omni Flash translated that 3D pass into golden-hour video while keeping the actual buildings & signage intact. Also used a Gemini visual critic loop to catch subtle issues across cuts (e.g. crowd outfits changing, 3D-looking tree bark, flat road shading etc). (4/5)
1
66
Essentially - using an agent-built & controlled 3D scene as spatial scaffolding for Omni feels quite compelling: 3D handles exact camera physics, layout & timing, while generative video handles real light, foliage & people. Not perfect, but pretty good.
1
39
Gemini 4 Carbon is out today - spent a bit of time testing its long-horizon agentic coding on a personal experiment: rebuilding ~220m of downtown Los Gatos in Blender (~50 storefronts + 3D interiors), w/ ~zero manual modeling. Quick look under the hood 🧵 (1/5)
23
18
8
277
39,978
Most compelling part in practice: automated visual review loops. In an early pass, review agents inspected renders across the block and caught that ~45 shops were accidentally sharing the same 2 interior textures - then rebuilt unique 3D rooms & props per shop. (4/5)
2
24
2,256
Essentially - the big shift with Gemini 4 Carbon for me is moving from one-shot code generation towards workflows that can plan, inspect their own visual output, and fix subtle bugs over hours. More on today's release: blog.google/innovation-and-a…
1
24
2,176
I turned my handwriting into a functional font in an afternoon. No manual tracing. No expensive software. Just one sheet of paper, OpenCV, and Gemini 3.8. Here's what it looks like typed in real time:
1
1
279
The dead giveaway of a fake handwriting font is double letters. If both 'o's in 'look' or 'coffee' are identical, it looks robotic. Because I wrote each letter 3x, we built an OpenType rule: any repeated letter swaps to an alternate version automatically. 🖊️
1
27
I didn't write this from scratch - I paired with Gemini 3.8 in an agent harness. From the printable PDF and OpenCV vision to TrueType winding and compiling the TTF, Gemini handled all of it in minutes. 🚀
1
63