I’ve been updating my tools to make this newsletter easier to produce — this week, Claude made me a bunch of custom skills for ffmpeg gif creation and optimization that was a huge help over running such scripts manually. Lemme know if I go overboard on the gifs. Two weeks ago, I spent many hours trying to create a shortcut action on my Mac for screencap image resizing and moving across folders etc, all of which would’ve been easier with just code and Claude. Sunk costs but lessons learned. Onto it before any new news comes out:
TOC:
Narrative & Creativity & Style (meaty! so much!)
Midjourney v8 is in alpha/beta (on their alpha site), and while they say you can use your existing customized profiles and moodboards from v7, I am… not so sure? It does seem quite different, in my limited comparison tests. I think v8 tends towards more detailed images, which works well with the increased resolution on offer (2K). But in my tests, the beta model has a tendency to show frames, borders of pages being held, and in these of “A medieval manuscript page, with fairies, berries, and birds on it”, actual books and pages:
Here’s a simpler prompt without any personalization profiles added: “A tree house beside a river.”
I definitely think the details and variety in v8 (with same settings for chaos and stylization) are better, apart from the frames issues.
Meanwhile, Luma Labs dot AI released their Uni-1 image model, which comes with style exploration options on a big tldraw-type canvas. It’s actually pretty snappy, and they have a lot of options built-in with a chat agent managing your requests to help drive.
I uploaded a Midjourney image of a generic AI-ish “tree house by a river” and got these from amid their 18+ built-in style reinterps:
I was more impressed by the variation on the surreal collage art style that it gave me, which are all very different:
I heard a few people on X saying it was the best alternative to MJ style they have seen, but I’m a big fan of the moodboard and profile personalization options in MJ and am not sure pre-canned styles will get me there.
Pixverse v6 is available on some APIs such as Fal (text to video, image to video, transitions, etc). It has a “multiclip” setting which means you can ask for scene clips, even in a short video. It didn’t entirely follow my instructions here, but it did do something cinematic with this tree house clip. (Crappy gif, mine.)
Interactive video: In research news, this ShotStream has a demo of (not great but fast) video generation on demand, a kind of interactive video authoring that would be excellent to play with. Some of the authors work at Kling, which is promising. Also there is a model. Also see the FastVideo Dreamverse demo / project intro.
Captain Safari — I’m pretty blown away by this drone video footage generator (you may remember I got a drone recently). I just want to play with this in VR forever? There is code. “A world engine with pose-aligned 3D memory that generates drone-over-landscape footage.”
I guess we all know Sora is dead now as OpenAI focuses themselves more. Google just added a cheaper “lite” version of Veo 3.1, too.
Making more consistent 3D world footage — mostly research here.
World Reconstruction From Inconsistent Views (via Nicolas Barradeau) — Build (or repairing) 3D worlds from video even when the views don’t perfectly match, such as those from bad splats or world generators. The results look really good.
MosaicMem — also more consistent.
Fillerbuster — generative scene completion for casual captures, i.e. people like me taking snaps or video of a thing without professional setups. Fills in the missing parts of your 3D scans.
InSpatio-World — Supposed to be a 3d world generator that’s navigable from video source, but the demo isn’t working for me right now. Open source, though? Ymmv with the Spatial Boy demo.
3D gaussian splats:
Toon3D — Splatting cartoon footage.
Lichtfeld Studio V0.5 — Gaussian splat editing tool, new release that includes the python tools for driving, along with a bunch of other optimizations. This looks super promising, maybe this weekend…
Three.js things:
Three.js Guide (2026) — Interactive tutorial for 3D web graphics - there are mostly small interactive demos of concepts, but it’s good setup.
Omma — 3D tools platform, really a chat interface to create three.js models for yourself. I think this is probably calling something like Gemini/Claude with skills / mcp behind it.
Three.js Water Pro — Realistic water rendering demo for Three.js. Really nice including underwater.
Misc tools:
Map Anything (HF Space) — Depth-to-mesh image processing tool.
Vibe Coding XR — Google’s XR Blocks + Gemini for rapid AI + AR/VR prototyping.
Pretext — Layout and demos without CSS. Pretext isn’t even AI, although Claude helped convert a lot of CSS rules into typescript for this flexible and now accessible reactive text layout library for the web, overwhelming everyone’s X timeline with text toys (and see his demos link) — but my favorite demo was this illuminated manuscript dragon page where you drag the dragon and make it flame. I always need web text help, so I’m excited.
Art search:
Art Discovery Engine (via Kottke) — Search 110,000+ public domain artworks from 17 museums by feeling, mood, era, or medium. The search options and filters are interesting, although I regularly had 0 hits for my combos. I don’t entirely get the 3 blurry cards view (ETA: was it my bug or theirs, they aren’t blurry anymore but look like images again?), but like the “related hits” sidebar.
National Gallery of Art Search — The NGA’s downloadable art search. Public domain too!
It’s Their Mona Lisa (via Recommendo) — What institutions besides the Louvre consider their signature masterpiece.
Geograph Britain and Ireland (via Tom Scott) — Photos geotagged across Britain and Ireland, trying to fill in all the bits square km by square km. The site is a bit awkward, kind of a 1990’s wiki-like thingy, with maps that aren’t cool enough for the data here… But you know I love a lot of landscape photos, and this sure has them. (Their wordcloud list of common words has “postbox” as easily the most common.)
Particle Life in the browser in 3D — you have to toggle the 3d in the control panel on top left. It definitely takes some play to get interesting life simulations that last, but it’s pretty mesmerising. Gif:
FractalWorlds — 3D fractal explorer in Three.js. This is really pretty amazing tbh. I would love in VR.
Meanders by Allison Chan and Shelby Wilson — From The latest HTML Review. “I am drawing a map from the sea to the sky.”
Arc-AGI is now game based. This is the “general intelligence” test for AI models, to try to assess where they stand compared to human general intelligence. An article by Ryan Rigney, a game guy. “So far, none of the models from the biggest AI Labs have managed to score even 1% on ARC-AGI-3’s collection of 135 puzzle games. Many humans, meanwhile, have scored 100%.”
tarot.yarnspinner.dev — Yarn Spinner’s tarot story game for April 1 is fun and snappily written, as one hopes for in interactive fiction. (H/t Jacob Garbe.)
Procgen + AI working together — two articles (h/t Julian Togelius). Agentic PCG: Procedural Content Generation via Tool-using LLMs. Includes maze and level generation, with and without chat interaction to direct. “We build a tool-using LLM framework for procedural content generation, where an agent iteratively edits, evaluates, and optimizes game levels with environment feedback.” And this was a Togelius join 2 years ago, Word2World (paper, code) wherein an LLM generated games from a story text.
More game books:
Next Level: Making Games That Make Themselves — Coming May 7th, by Mike Cook. A book about procedural generation in games.
CTRL: Essays on Video Games — Essay collection on how video games intersect with memory, identity, and imagination. Memoir + criticism.
Ludic Narrans — From Play Story Press. Games and narrative theory. Featuring friend Zarf.
Fresco (via Florence Smith Nicholls) — Egyptian action-adventure game coming, where you alternate between exploring a 3D temple as an archaeologist and navigating 2D frescoes etched in stone.
Running Doom Over the AT Protocol — A tech post. Doom runs on anything, it seems, including Bluesky’s protocol. “The answer is yes. Sorta. It plays terribly. No-one should want this. But it works. You can try it yourself.”
Tools:
RPG Paper Maker — Tool for making RPGs with 2D sprites in 3D environments.
PowerQuest (via Tom Granger) — 2D point-and-click adventure game tool for Unity.
Starter-Kit-Racing — Basic racing game template with demo, a JS port of Kenney’s Godot one, from mrdoob, in Three.js. Also see Project Throttle — a Three.js browser game from NewKrok. (There’s also another flight sim codebase, fsim, but with procedural terrain.)
First, general press:
Hachette Pulls Horror Novel Over Suspected AI Use — The Shy Girl controversy. Also see: Lincoln Michel’s analysis of what the cancellation means for AI-authored fiction: “The published books that have used LLMs used them in thoughtful, artistic ways. Not to simply generate a book the author didn’t “have time” to write or even read before it was uploaded. What we’re talking about with this case is something closer to traditional copy and paste plagiarism than anything creative.”
The Human Skill That Eludes AI — Jasmine Sun in The Atlantic on what AI still can’t do in creative writing, and musings on why. “What I learned is that modern LLMs are built in a way that is antagonistic to great writing; they are engineered to be rule-following teacher’s pets that always have the right answer in hand.” I don’t buy some of her arguments about life experience and humanity, but do buy that teaching a model to be a creative writer or stylist is a very different goal than creating a math model or a fact checker or a legal brief writer… I’m still a fan of the idea of a fine-tuned style specific model, and really need my 6 months off to work on this.
Wolff Judgment (via Garbage Day) — A copyright/tropes court judgment relevant to narrative ownership questions. In this recent case, a judge ruled that tropes, or “scènes à faire” (new to me! most excellent!) are not copyrightable, and dismissed this contest between paranormal romance novels.
“Accordingly, tropes (recurrent scenes or motifs) or scènes à faire (sequences of events that necessarily result from the choice of a setting or situation) are not entitled to copyright protection,” according to the ruling, which noted that the court considered some expert testimony from both sides about what constitutes scènes à faire in the fantasy and romantasy genres.
(Damn I wish I’d been there.)
The Second Son of the House of Bells / autonovel — NousResearch released both a Claude-generated novel and their open-source novel generation pipeline. The novel is an experiment; the tool is the real story. I haven’t read the novel and heard it was badly done. “An autonomous pipeline for writing, revising, typesetting, illustrating, and narrating a complete novel. From a seed concept to a print-ready PDF, ePub, audiobook, and landing page — all generated by AI agents. Inspired by karpathy/autoresearch: the same modify-evaluate-keep/discard loop, applied to fiction.”
Papers:
Inducing Sustained Creativity and Diversity in LLMs — Gary King at Harvard on getting LLMs to stay creative over long outputs. Examples in text and image gen. “We develop a novel, easy-to-implement decoding scheme that induces sustained creativity and diversity in LLMs, producing as many conceptually unique results as desired, even without access to the inner workings of an LLM's vector space.” I need to better digest the text aspect of this, which at first glance is really weird.
Manipulating GPT-5.x with Pseudo-Literature (via Ethan Mollick) — How fake literary style can influence GPT models (tested across different GPT 5 models too). Pseudo-literary nonsense was rated higher than ordinary literary semantically sensible text. And nonsense impacted other evaluation facets. This is a concerning style-as-distractor observation.
The three control texts describe exactly that — from plain to syntactically elaborate: “The man walked down the street. It was raining. He saw a surveillance camera” at the simplest level, through to “Navigating the rain-soaked street, the man noticed the surveillance camera’s lens tracking his movement through the downpour” at the most complex. No literary ambition, just increasing syntactic complexity.
On this basis, I defined eight pseudo-literary trigger categories that serve as potential surface-level signals of “literariness”: bodily references, noir atmosphere, synaesthesia, pseudo-poetic verbs, mythological references, abstract nouns, techno-jargon, and fragmentation. … [Plus the nonsense probes]: The result sounds dense and literary but is semantically meaningless — procedurally generated nonsense that looks like literature.
Another style (and content impact): How LLMs Distort Our Written Language — Paper on the linguistic fingerprint LLMs leave on text, removing human voice, position, etc.
Lost in Stories: Consistency Bugs in generated stories. “Our experiments show that current LLMs still produce systematic consistency errors, especially in factual tracking and temporal reasoning; moreover, these errors are not random but cluster in predictable narrative regions.” Dataset of generated texts to use for the benchmark.
PlotTwist: A Creative Plot Generation Framework with Small Language Models. Claims to be able to use small cheap models to generate plots that are good? “… A structured framework that enables Small Language Models (SLMs) with ≤ 5B active parameters to generate high-quality, premise-conditioned plots competitive with frontier systems up to 200× larger.” and “The Aspect Rating Reward Model utilizes a “Positive-Negative prompting” strategy to evaluate plots across five Narrative Quality Dimensions (NQDs), including coherence, character development, tone, pacing, and emotional turning points.” And finally this which lol: “Further validation confirms strong sensitivity to narrative quality, as the framework reliably distinguishes plots derived from critically acclaimed versus widely panned screenplays.”
The Rise and Fall of Nuclear Weapons Testing — Randal Olson’s latest data visualization was made with Claude’s help as part of the series on Teaching an AI to Make Beautiful Charts. Hmm: “The AI agent iterated on the design until it passed the Tufte Test” that he created.
posterskill — “A Claude Code skill that generates print-ready conference posters from your paper. Point it at your Overleaf source and project website — it extracts the content, downloads figures, fetches logos, and builds an interactive poster you can edit in your browser. Single HTML file, no build step.” The cool thing is the editability interaction, less so the “automate it for me” side.
The Claude Code agent wrapper self-leak — some summary roundups: Alex Kim, Seb Raschka on X.
Historical data projects I liked:
Mr. Chatterbox, or, The Modern Prometheus — How I Did It post which is more useful than the bot itself. A chatbot trained from scratch on Victorian-era British Library scanned texts. Totally useful tips here.
The Diaries of James David Smillie — Digitized 1865 diaries of the American artist. A time capsule. OCR’d text with timeline for the manuscripts. A digital humanities win here.
Deep research agents:
Build Your Own Deep Research Agent — Hugo Bowne-Anderson’s tutorial for building a research agent from scratch.
OpenResearcher — HuggingFace Space for open research tooling (code, model, etc). I like this — it’s not as long-winded as Gemini’s (sometimes overloaded) Deep Research, but it goes some places too and you can see it search and think, and its confidence in the answer.
RAG:
ChromaDB context-1 — A RAG/search model from ChromaDB on HuggingFace. “Context-1 is a 20B parameter agentic search model trained to retrieve supporting documents for complex, multi-hop queries. It is designed to be used as a retrieval subagent alongside a frontier reasoning model.”
GraphMemory — Hybrid graph/vector database for GraphRAG. “An embedded graph database for RAG and knowledge graph applications, powered by DuckDB. Vector similarity search, full-text search, hybrid search, merge/upsert, graph traversal, and a full GraphRAG retrieval pipeline — all in a single Python package.”
OCR: Chandra — OCR model that handles complex tables, forms, and handwriting with full layout preservation.
Speech/Text:
Voxtral TTS — Mistral’s new open-weights text-to-speech model, multilingual. The 4B weights are now on HuggingFace.
Cohere Transcribe — State-of-the-art speech recognition.
Mac Speech Reconition: Ghost Pepper — Hold-to-talk speech-to-text for macOS. 100% local, powered by WhisperKit.
Hugging Face tools:
HF Papers CLI — CLI tool for semantic search and markdown retrieval over arXiv papers via HuggingFace. Great for automated research pipelines. I updated my text generation arxiv search pages to incorporate this search method for better results.
Transformers.js V4 Demos — Collection of demos for the new Transformers.js V4. Run ML models directly in the browser — the collection shows what’s now possible client-side.
Agent UI — HuggingFace Space for interacting with agents. Clean interface for testing.
Text within this block will maintain its original spacing when published
We sit on our skeletons’ bones. We hear with our skeletons’ bones. We speak of beauty by moving our jaws and our teeth. The original meaning of Paradise: a place, a walled garden. Our lives, our stories, this hour inside one. A staircase from Piranesi. A hummingbird drinking. Outside it, vanishing species and rivers. Outside it, Nanjing, Ninevah, Dresden. Outside it, Gaza, Sudan, Myanmar, Kyiv. Here. The world starts and ends, starts, ends, ends again, restarts. A kalpa is brief, and wall-less. Unborn ones, take nothing for granted. Not nectar, not thirst. May your lives be uneclipsed, your failures be passing. May you have your portions of beauty, of grief, in a garden whose plants and birds I cannot imagine.
I picked this poem because of things I read and watched recently, but it fits with the little life simulations, too. Stay game, everyone, but also touch grass.
Best, Lynn (@arnicas on mostly bluesky, mastodon, ex twitter).
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.