TITAA #71: Splats and Fat Cats
World Labs splats - Video Cats - Distributed Heroines - Among AIs - Historical Type - Reve
I wrote a bunch about the World Labs splats I played with this past week, so we’re just diving right into that in the 3D segment. Fun stuff!
Table of Contents (I promise to make them links):
Web Fun / Misc / Arty
AI Creativity
3D
World Labs! Let’s start here because World Labs finally opened up their beta app Marble to a lot of people, me included. You basically prompt a 3D gaussian splat environment via text or image to get a little splat world to play with in a web browser. Their fast mini model takes about a minute and their better one takes about 6 minutes. A small splat world means “I can move around in it, it has 3 dimensions and a floor, ceiling, and walls.” The details falls off pretty quickly on the sides as sketchy gaussians, but it’s nevertheless very cool. Pics and videos below!
It’s been an interesting ride, seeing what people have rendered and—even better—how people who can manipulate 3D tools are using those little world artifacts. Cleaning them up and stitching them together is kind of the baseline, which is what I spent some of my weekend learning to do. The World Labs web page shows a large connected set of hobbit hole rooms with an overhead minimap view, a feat which which requires hand construction at the moment (you can see it on X too).
On X, some more examples to see for different attempted use cases:
consistent world models for video using World Labs output (link and also this one), and an experimental movie using World Labs plus Veo, Eleven Labs, etc (link)
multi-user chat settings using frame_vr (link)
a giant example rendered as a movie with music (link)
a “Draw Your Game” app putting characters into Marble splat worlds that’s on Steam (video of cutout world)
using Marble “templates” in Rosebud.ai to build games (link)
adding physics (lib below) to play jenga in one “world” (link)
redesigning rooms and seeing how they feel around you, sometimes with the help of Nano Banana (various people including levelsio and Justine Moore on X)
viewing them in VR (see below), almost its own use case it feels so different.
But onto my experiment:
For exploration purposes, I made 2 worlds from images generated in Midjourney for a Piranesi-like museum/gallery of strangeness (the Susanna Clarke Piranesi).
The boundaries on the splat worlds don’t go far, but I remain interested in the visual decay as the details fade. I think it’s entrancing. Some views in my browser:
Using the SuperSplat editor, I cleaned their limits up on the ends (and a little bit inside but not much) and oriented them end to end:
While in there messing around, I finally figured out how to make and export an animation of moving end-to-end (I kept the blurry junction of the 2 because I liked it, while most people are editing thing together seamlessly):
And then with the help of some audio generated in Suno v5 and some AI coding tools, I made a little web app that has the camera circle in one splat, transition to the next with a glitch audio effect, and then circle in that one. Positions and views are randomized slightly each transition. Each splat has its own audio file and there is a drop off/mix as you move between them. Tiny video snippet of running it once:
This was fun, and made me think hard about ways to use these things even in the limited form they have now, how they could be easier to work with, and how audio might work inside them. I also ended up wanting some different things from Suno v5 that I couldn’t figure out how to generate, like whispers and chanted nonsense, and more sound fx mixed in with the audio (wind, rain)… I could spend time editing that myself, but, geez, time! More coming, I hope!
Tools of relevance, some of which are new and some of which I’ve mentioned in the past, but used over the weekend:
The open source and online SuperSplat editor for trimming your splats, making animations, etc. This is invaluable. I modified a local build to add camera coordinates in the display to debug some camera automation in my own project.
bmild/spark-physics at cabin: Adding physics to the Spark splat tool!
Gsplat Share - for a live web demo (of limited time) of your splat. Here’s one of mine, a raw output from World Labs, using their site (valid for a week only): link.
A little navigator app (code) to load your splat worlds with arrowing between them, demo here.
Lofiworlds.ai — a little idle app for VR that shows off lightly animated cozy Marble splat worlds. In VR it’s amazing even with low res. There is also a lot of VR enablement work going on: they just added an “open in VR” button to their UI browser of user created worlds.
The SparkJs web tool project from World Labs, which is being actively worked on to support work flows using the splat outputs of WL Marble. For instance, this is an X post about a not yet merged LOD (level of detail) scheme for the lib that is loading and using 16 million splats in real time. (And 5 million splats on an iphone mini.) They’re hiring, btw.
One other quick item worth noting under deadline:
AlmondGod/tinyworlds: minimal reimplementation of deepmind’s genie 3 world model, a small, artisanal version of Google’s unreleased Genie 3. Cute. Maybe it belongs in Games though.
Image Generation
Reve Image - AI Image Generator and Creative Tool - Christian Cantrell’s image editing/generation app in its latest form. You can have conversational editing threads, and it has a proto-editor tool that allows you to select objects in images and modify them by editing their prompts. You can also just generate by prompt, like these 3 from my attempt at “polaroids with questionable garden fairies”:
Here’s a view of me iterating on cleaning up some blurry statues in the foreground (1, top left), and adjusting their design:
Video
A good piece in the latest on “how it really goes trying to make pro work product with AI gen tools” — Tool: Land Rover Defender – “Who Says Cars Can’t Dream” Experimental Spot (via Matt Muir’s Web Curios).
The Kling 2.5 Turbo is really good. You can try a glif app to make a long video with it if you want (my effort was only so-so and then Kling had an API problem, but Fabian’s examples on their site are great). Also out: wan-2.5-i2v on Replicate and Fal.
And this morning was Sora 2… well, not super accessible from the EU, but there are means (thanks to TomLikesRobots for an invite). I tried to make a bunch of Fat Cat Week videos (all text to video except for the last Midjourney one), for “fat cats play in a pool under a waterfall, catching salmon and tossing them in the air” and before I knew it I had done it with a bunch of models. See what you think about their grasp of physics, fatness of the cats, etc.
Web Fun / Procgen / Misc Arty
The three.js account on X posts a ton of gems, if you are into 3d web stuff (obvi I am).
Air Fiesta FM (Beta 2.2.59): A Google-Earth based 3d web game with balloons. I love this. You fly around over your map of choice (my new town, for instance) and navigate towards pins floating in the air, to open up controls and ability to stream local radio. (Via Three.js or Garrett Johnson, or both.)
Another cute three.js web mini game - Messenger. Deliver letters in a small world. There may be quite a lag… some characters had a “…” over their head for quite a while. Maybe because it’s multiplayer (you run into yourself a lot).
Josh Comeau’s Whimsical Animations course is opening up with some preview content.
Lofiworlds.ai — as noted above, a little idle app that shows off lightly animated cozy World Labs splat worlds. In VR it’s amazing — every world is a story waiting to be told. You can watch them without VR but you won’t be as impressed, and you won’t feel like you’re in a moving painting.
Speaking of moving paintings, friends and colleagues of mine worked on this Google Arts & Culture animated paintings project with Veo, which is lovely: Moving paintings × Fukuda Art Museum — Google Arts & Culture.
Another superb historical recreation from Nicolas Rougeux: Printing Types, their History, Form and Function.
Games
Canvas of Kings on Steam: Amazing map maker tool, I will make maps with this.
#47 | We asked for chaos. You delivered 480 times. Indie game jam produced entertaining ideas.
I built ChatGPT with Minecraft redstone! - YouTube: Building an actual LLM in Minecraft blocks. I continue to be blown away by the Minecraft craft culture. Also, today’s Dreamer 4 paper from Deepmind— “By learning behaviors in imagination, Dreamer 4 is the first agent to obtain diamonds in Minecraft purely from offline data, without environment interaction. The task requires choosing sequences of over 20,000 mouse and keyboard actions from raw pixels.”
Among AIs — “A live arena where top AI models play Among Us against each other.” God I love this. Done before the new Claude was born, GPT-5 is an outlying winner. “GPT-5 combines high accuracy with low harm, the safest town player.” I’m honestly a bit surprised. “GPT-5 remains proactive and leads harmful ejections; good at deception.” That, otoh, doesn’t surprise. I loled at the title on this example:
Roguelike Celebration - Event 2025: Online fun talks about games, procgen, narrative — tickets for sale now, program announced.
Doors in games, a game about doors — via Federico Pianzola. So cute, after my mid-month AI door exploration 2 weeks ago. They also have a bunch interesting other meta-games and think-piece games, plus articles, which all sound very intriguing to me.
Doors is our newest attempt at playable philosophy. It is designed to invite players to interactively examine existing theories about how objects are represented within games and virtual worlds more in general. Each door in the game raises a different question regarding their representation.
Narrative
“Playing God” — “Interactive Cinema Isn’t a Game,” by Claire Evans, via Critical Distance. “‘This is not a movie,’ Ebert, who would later name Mr. Payback the worst film of 1995, agreed. ‘A movie acts on you.’” A piece that ends up focusing on Sam Barlow’s film corpus oriented games (of course), and perhaps could go deeper. Still an interesting read.
“As stories let us record narratives,” Nguyen writes, “games let us record agencies,” encoding them into “artifactual vessels” that can be passed around and temporarily embodied by others.
Playful Narrative: A Toolbox for Story-Rich Mechanics (from Polaris Games Workshops, a working group writeup).
We became concerned with delving deeper into topic areas such as: story existing as a mechanic and not just a layer over top of interactivity, what story looks like in different contexts, how heavily system design interacts with narrative design, and understanding the friction points in these to know what pulls players deeper into story and what pulls them out of story.
Catch Me If You Can? Not Yet: LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday Authors. Also see a good post from Lincoln Michel about writers’ style, “Style is More Than Sentences.”
Beyond Plot: How Sentiment Analysis Reshapes Our Understanding of Narrative Structure, in Journal of Cultural Analytics. Some good thinking about the limits and affordances of sentiment in looking at narrative.
The Woolf case study exhibits what I call a “distributed heroine” (Elkins, Shapes 27). In other words, the emotional arc, with dramatic rises and falls, has peaks and valleys that highlight different characters’ perspectives.
The new Claude model Sonnet 4.5 has overtaken spot 1 in Sam Paech’s writing evals.
Data Science / NLP / Tools
Super quick dump of some interesting things I saved so I can get to relaxing…
mmBERT: A massively multilingual modern encoder language model. What it says, a recent release.
AutoIntent documentation: “AutoIntent is an open source tool for automatic configuration of text classification pipelines, with specialized support for intent prediction.”
Introducing Liquid Nanos — frontier‑grade performance on everyday devices | Liquid AI — small device models.
Large Language Model Hacking: Quantifying the Hidden Risks of Using LLMs for Text Annotation. “LLM outputs vary significantly depending on the implementation choices made by researchers (e.g., model selection, prompting strategy, or temperature settings).” Avoid misleading conclusions.
Knowledge distillation - Fireworks AI Docs.
Chrome DevTools (MCP) for your AI agent | Blog | Chrome for Developers ☑
mcp. There is also now a Claude Chrome extension in testing, which, interesting….
Observable Notebooks 2.0 Technology Preview — offering Python now, too.
MinishLab/vicinity: Lightweight Nearest Neighbors with Flexible Backends.
Train destinations from Lyon | Chronotrains - Europe Train Map via miska knapek.
hallucination_probes: Real-Time Detection of Hallucinated Entities in Long-Form Generation.
IBM Granite-Docling: End-to-end document understanding: IBM released Granite-Docling-258M, an ultra-compact, open-source vision-language model (VLM) for converting documents to machine-readable formats while fully preserving layout, tables, equations, code lists and more.
A Poem
Here the days don’t dissolve in air
they fall into the water
shaping their own shell
a sheen of separation.
A hawk flies over summer’s body
diving again, again
feeding and drunk from the fall.
There’s nothing here
but manic wind alone and stones
and sea
a senseless promise
sharpens our lust with the moon’s blade.
When I arrived here, in the landscape of endings,
the wind entered my mouth with so much rage
as if I were its only vessel
until all my words vanished.
Each tree receives the wind’s gust differently
some suffer, others—again—resist
(I’ve met a palm tree that birthed the wind,
then sent it in every direction)
others shiver all over and change colors.
I, of course, am not a tree
I sat down and wore the wind’s coat
I stooped my head and looked at the ground
through its cracks, thyme’s roots
& their hieroglyphics
struggled to enter the light.
Then the words came back.— Katerina Iliopoulou, Cape Tenaron
Best, Lynn (@arnicas on mostly bluesky, ex twitter, mastodon).
Marble looks very cool!