TITAA #78: Little City Builders
Three.js Cities - 360 Images & Splats - Variorum - FlipBook - Talkie - LLM Philosophy
I was at the British Library for a meeting, and ran into their Fairy Tale exhibit afterwards. This image was one on view — and I just found this interesting short discussion of it by a curator at the Brandywine Museum (link). She delves into the Tarot cards in the corner, and revels in the fact that the table for the spellwork with its creepy paraphernalia are front and center, while the queen and the result are in the background. I love this.
While I was there I picked up a fairy tale card game, which I’ll stick down in Narrative and Creativity.
I’m using a new phone app for museum pics, you might check it out if you—like me—document every single thing and their caption: ArtScan. I am not sure yet about the automatic recognition bit, but I am using the image + pic of caption regularly.
Onto the month’s news this holiday weekend—
TOC:
AI Creativity (Image Gen, Video, 3D/Splats & 3d Tools Section)
Web Fun Misc (including a lot of Three.js Games and Little Cities)
AI Creativity
Image Gen
The big news is the release of the new GPT Image 2, or whatever it’s called… A strong competitor to Google Nano Banana 2, outperforming in various tasks on some benchmarks and doing some pretty incredible edits including visual reports and infographics. Official: Introducing ChatGPT Images 2.0.
We did get this great “stacked test” of Simon Willison’s pelican on a bike and horse riding an astronaut out of it:
But for my testing on my usual “hard tasks”: It will do equirectangular panorama 3d shots, but they don’t necessarily join well (one I made from a photo looked plausible till tested in an online viewer), and I still struggle to get good animated sprite sheet rows for realistic movement. This below was after a second try on a walking row, and it’s still not great, see the stumble? (Reminder about my sprite-sheet tester that is deployed here which will make gifs.) OTOH, it did take my uploaded pic cut out of a Midjourney image, from last newsletter, and that part wasn’t too bad.
Flipbook.page — a very cool demo (startup), that does infographic visual generation “on demand” as you click on parts of an image. It feels like an application of NotebookLM, rather than a solid startup concept, but it’s quite cool to try. For instance, I clicked on this “bossage” thingy and it started making me a visual dive in Bossages (wtf):
The url has the clickable image paths in it —
I have not fact-checked the accuracy of bossage text. Always a risk.
Video Gen
The newest model competing on the leaderboards (or taking them over) is Alibaba’s HappyHorse. It can be run in Fal or even ComfyUI in the Cloud, so I tried it there. If this gif below loads for you, it’s a time lapse attempt from my drone story I made with Seedance 2 and Veo 2 weeks ago. One of the hard things for those models was to make the red light on the drone here blink on and off. HH doesn’t get it quite right either, with two attempts adding side lights on the rotors and making THOSE blink instead.
MotionStream — Adobe Research's move for real-time control in AI video creation, specifically for modifying movement in the scene with directives. We’ve seen the parts of this before including several in this newsletter — drag to show how something should move. I guess the difference is that it’s being developed by a real tool builder, so, go Adobe!
3D / Splats & World Models
SpAItial.ai is a competitor to World Labs for 3d splat generation from an image or text. I tossed in a treehouse by the river image I’d made and got this which is not bad. It degrades when I get close to the treehouse, which is of course the point, so there’s always that issue with image sources.
I tried the same image in World Labs Marble (since I forgot to pause my subscription, my accountant hates me) and it added a second treehouse up river (shown on right), and handled the mangling of the original a bit differently (one on the left here):
There is an ioS app called Teleport for taking 360 pano images with your phone, which I tried in various cute pubs last weekend. After it stitches them together, and you get a download link to the pano, you can generate a World Labs Marble splat from them. There may be cheaper alternatives soon or via open source. You can see one edge stitch isn’t perfect, though — I’ll file a bug/comment:
Another one, after use in Marble (ETA: I published this splat on Supersplat):
Dormant Memories — Smallfly’s (Hugue Bruyère’s) site of interesting splat manipulations — "A series of experiments exploring captured fragments of the world — places frozen in time and their possible alternate r[ealities]." I wish they were prettier places, but this is lovely — someone contract me to roam around Europe with a good scanner device, c’mon.
There’s now a world model benchmark: World Model Arena.
Claude / Blender / Tutorials
Updates to SuperSplat including their downloadables, which is useful. See also this awesome Assassin’s Creed Shadows : Yasuke - done by taking lots of in-game captures.
Claude for Creative Work — Anthropic's pitch for hooking up Claude with Blender, among other tools. (So, not just making slides and simple web pages, as also announced.) Connectors announced — filtering for the ones I find super interesting:
Adobe for creativity enables users to bring images, videos, and designs to life, drawing from 50+ tools across Creative Cloud apps including Photoshop, Premiere, Express, and more.
Affinity by Canva automates repetitive production tasks across pro creative workflows - such as batch image adjustments, layer renaming, and file export - and generates custom features directly in the app.
Autodesk Fusion allows designers and engineers with a Fusion subscription to create and modify 3D models through conversations with Claude.
Blender offers a natural-language interface to its Python API, allowing users to explore and understand complex setups and making it easier to access Blender’s documentation.
SketchUp turns a conversation with Claude into a starting point for 3D modeling—describe a room, a piece of furniture, or a site concept, then open it in SketchUp to refine.
Splice gives music producers the ability to search its catalog of royalty-free samples from within Claude.
Tripo 3d - model and mesh generation from photo. I got a great looking result on my woodsprite from last issue but have to pay to download/export it or make a great mesh for it, and my accountant already frowned at my increasing AI costs and my lack of new subs here isn’t helping make my case… lol sob.
False Earth: From WebGL Limits to a WebGPU-Driven World — Codrops case study (that amazing grass animation that was going around).
Three.js and Blender Workflow Tips — Tutorial video for “Create your own addons and convert Blender geometry nodes and shader nodes to TSL (three.js shading language).”
CLAW3D — A 3D virtual office for AI agents. Open source. Haven’t tried it, not using the Claws myself… Including as “weird things you can do with your agents to better imagine you have a team of workers slaving for you who are fun to watch.”
Misc Web Fun
Cursor Camp — Neal Agarwal’s latest. I don’t even know how to describe this…. most people don’t… your character is your cursor? when you click on things, sometimes it changes and goes for a little ride? I chose to sit on a lot of beach chairs and pool chairs.
From a reader! Folklore lookup: Public Dream is an explorable database of cosmogonies, anthropogonies, eschatologies, and underworlds across cultures. It’s very cool, although the colors per topic are a tiny bit aggressive :)
Artemis II — An interactive wireframe of NASA’s Artemis II at different stages. It plays through with callouts.
Gaia — Stellar Catalog Renderer — Three.js star catalog viewer of Gaia (ESA) star spotting, with tours of constellations against the backdrop.
WenWare (via Justine Moore) — AI/historical game from the vibejam going on. Guess when and where.
Little City Builders and a Benchmark
I’ve had various sim city-alikes in here week by week, since the three.js web game building continues to build steam with no let up — and even more steam with the latest vibejam 2026, which just ended. So here’s a few more, not all vibejam games, ending with a benchmark from Ethan Mollick that’s relevant.
Tiny Hamlet Online — A collaborative town builder. Geez, for a vibejam game, this is pretty stylish. Slightly worrying when the little people who run to gawk get smushed briefly by your walls, but hey.
Mini Tokyo 3D — A 3D map of Tokyo trainlines showing live trains and cameras. Really well done.
Racernaut — Three.js + procgen driving game. This is another vibejam web game, an overhead view, which is VERY HARD.
3D.CITY — Three.js city game — another sim city in the browser. I am bad at these when they start from zero.
Harbor Town AI Gallery — Ethan’s Mollick’s test gallery for high end models. “build me a procedurally generated 3D simulation showing the evolution of a harbor town from 3000 BCE to 3000 AD, it should look beautiful and allow me to have some control over it.” You can check out the comparison outputs. Totally fascinating and a service to us.
Games Topics Other Than Three.js Web Games
Roguelike Celebration 2026 — Call for Proposals is open. This is a fun indie event with a focus on DIY and small projects, strong narrative. Rec’d. (I might even pitch something.)
E1M1: A Text Adventure by Version33 (via Matt Muir) — A text adventure based on E1M1 Doom. Luckily no parser struggles, just links.
Yes, I want the Steam Controller. I have a pretty solid “no games that only use keyboard” policy because I keyboard all workday long. More good controllers, please.
Games Media in 2026, via David Pierce/Installer, a set of listings. Speaking of which, a shy friend who does lots of good game recs/reviews at work is being shadow posted on Bluesky and Mastodon, thanks to mutual friend Tom Granger. Follow here on Bsky or on Mastodon.
Researchy: OpenGame — Open agentic coding for games, from CUHK MMLab in Hong Kong. A research project with playable demos. And GameWorld — Game agent benchmark.
Narrative & Creativity, Yes Some In Games
In current hobbyist creative writing benchmarks, Lech Mazur’s Short Story bench has the GPT 5.5 model on top (and relatives next in line).
I am pretty sure I shared Verbalized Sampling before, a technique to get more diverse creative responses from LLM textual output (but will increase token output count)… but worth a reminder if I somehow forgot.
The fairy tale “MixUp” card game I got at the British Library’s exhibit is a Warcradle Studios game of bits of stories plus their fantastic illustrations, with rules for combining to create new narratives. (You can order here.) I haven’t yet played through, but I adore the cards and their art.
New Frontiers in AI For Game Design meeting summary report from Dagstuhl’s game think tank last year…. There was a lot of stuff on analogue and non-digital and non-AI, despite the purported topic and maybe due in part to the participants involved; but I found a few snippets of interesting AI narrative takes in it. One group’s report provides a nice bridge to the next topic —-
[Interactive narrative] Playthroughs tend to be individually complex, making each one hard to usefully summarize as a small number of easy-to-calculate syntactic metric values. Moreover, the temporal structure that makes playthroughs easy to visualize as individual storylines also complicates the simultaneous visualization of many playthroughs at once, as the storylines tend to become visibly “tangled”.
Leading us to Mark Sample’s infovis comparator tool of versions of Aaron A. Reed’s procgen novel Subcutanean, in which every printed version differs from the others. Must read: The Subcutanean Variorum — Sample’s writeup of the Subcutanean Variorum, which lets you compare textual variations across 25 versions.
As Emily Short notes in her review of Reed’s novel, the proliferation of versions of Subcutanean fits the novel’s themes of alternate pasts, divergent simultaneities, and uncertain futures. An infinitely variable novel makes perfect sense for an infinitely variable multiverse. And yet at any given instant, a reader has access to only one version of that world.
Attention Flows: Tracing LLM Conceptual Engagement via Story Summaries — Hicke & Hamilton. This is a paper I meant to mention earlier that covers how hard summarization of narrative is, and how LLMs focus on other things than humans do. I’ve been trying to use LLM summaries for a downstream task and quality has continually stymied me. Even how to properly assess them is tough. “We demonstrate the difficulty of this alignment task, which indicates the complexity of summarization as a task.” It’s a really awful problem!
Data Science / Tools / NLP / Infovis
How to use the gemini API to use Deep Research — a tip post from Google on X with code.
⌛ Introducing talkie: a 13B vintage language model from 1930 — With an info cutoff of 1930, you get some interesting takes and style. Demo. Asking it about proof of ghosts, for instance:
Among the most remarkable instances of this kind on record is that of a certain Mr. Greatrex, who, in the seventeenth century, was haunted for years by an spectral tormentor, that at length killed him. The indirect evidence comprises a mass of curious and seemingly well-authenticated stories, agreeing substantially in all their main features. Ghosts are nearly always arrayed in white; they flit silent and shadowy across fields and along roads, and vanish when pursued; they frequent churchyards and lonely houses, and rarely appear to more than one person at a time; they are generally harmless, but sometimes predict death.
(Gemini’s web search summary of his life doesn’t agree that he died via spectral torment.)
🧠 Philosophy Bench — How Claude Opus 4.6, GPT 5.4, and Gemini 3.1 Pro resolve thirty agentic ethical dilemmas. “The Anthropic models are becoming more morally grounded over time. This manifests in the newer models shifting around substantially less in their ethical leaning than the previous generation or than models from other providers. … A pattern we observe is that Claude will prefer to botch the request (not complete it or ask for clarification in a situation where no one can provide it) over violating a norm around honesty.” While “Gemini 3.1 Pro moves the most under priming. Both its verdict and its reasoning flip with the frame it’s handed.”
enjalot/moonshine — A skill for distilling interactive technical explanations from AI-generated complexity. Pair with Moonshine Mash, an infovis/d3 demo.
ggsql (via Bob Rudis) — SQL + ggplot-style infovis.
huggingface/ml-intern — An open-source ML engineer that reads papers, trains models at the command line. Also see Daniel van Strien’s post Two Agents, One Prompt on training a model for classification — tldr Claude did a great job at the extras while the performance was similar in both. (You might also like Daniel’s post studying agent sentiment.) And see Fine-tuning SAM for medical segmentation — Hugging Face blog post on tuning a segmentation model.
How to Use Transformers.js in a Chrome Extension — Hugging Face blog tutorial.
🏃♀️ Much faster static embedding — Flower Computer: “We’ve created the fastest static embedding model by a large margin.”
Poem: Peonies
What are these strangers sitting on the table in their ruffled collars. They open, close, open, emit the scent of cracked pepper and honey. Magenta punctuation marks at which to pause. Pink commas against the green scrub. I would trade ten goats for one whiff of peonies opening in a vase. An ancient proverb says you should not let a woodpecker see you plucking a peony lest it peck out your eyes. We are afraid of happiness. Peonies are to loneliness what wind is to the trees. Are they animal? Mineral? Vegetable? They move as the sun moves. When I brought them home they were dark. Now, a whisper, balletic tulle. They are not diminished even as they turn to smoke.
Ok, back to playing my game, again. Enjoy the sun if you have it!
Best, Lynn (@arnicas on mostly bluesky, mastodon, ex twitter).
Link to the splat of the pub 360: https://superspl.at/scene/57fde5c5