Avinash

1/ I’m excited to share recent work on a scalable and generalizable causal discovery framework from @mingZedh, me and @sophokevl, supervised by @elhamazizi. Open the thread 🧵 for a brief tweetorial of our method. arXiv link: arxiv.org/abs/2603.03411….

Our latest iteration is now live at withcascade.com A continued collaboration with @johnathanchewy

We still have prior commits, but now as a side panel where users can drag and prior version to compare visually on the canvas. The core of Cascade is now freeform visual exploration with coding agents.

We rebuilt Cascade from the ground up. Think @figma meets @conductor_build: a visual canvas for coding. Preview your latest commit → copy and paste to spin up new worktrees → edit visually → commit.

Reliving my recent japan trip through MinecraftLM (github.com/mattzh72/minec…), our world building agent. Some examples in 🧵 Prompt: "A traditional Japanese onsen as an isometric diorama. A small wooden bathhouse building with tiled roof, noren curtains at the entrance, and shoe lockers visible inside. Behind it, an outdoor hot spring pool surrounded by natural rocks, steam rising from the water. Wooden deck connecting the bathhouse to the pool with a small changing area, towel racks, and wooden buckets. Bamboo fence enclosing the space, stone lanterns along a path, maple trees with red leaves overhead. Snow on the ground and rocks for the classic winter onsen look. The bathhouse glows warm from inside, contrasting with the cold outside."

holiday feature: selectors to easily preview mobile, tablet and desktop views @withcascade

@ahall_research @karpathy Would love to see this kind of study applied to prediction markets - multi agent strategies could be next on @ProphetArena

Recently @karpathy built an "LLM council" to provide advice to users. This got me thinking: what are the best governance rules for the council? So I built a little experiment testing four different procedures: (1) The LLMs do a simple majority vote without communicating to each other (2) LLMs vote, then deliberate, then can update their vote. (3) LLMs vote, deliberate, update their vote, and then a Chairman makes the decision. (4) LLMs vote, evaluate each other's answers, and then the Chairman makes the decision (like Karpathy's procedure). I evaluated how well these procedures do compared to individual models on two very basic eval sets with known answers (GSM8K and TruthfulQA). I focused on very cheap models to spare my wallet. For the very particular setup I created---and I don't claim this generalizes---it looks like all four approaches do a little bit better than the best model on its own, but option 2, voting with deliberation, does the best. (I didn't run the study long enough to get "statistical significance" but if someone wants to they probably could). It turns out, unbeknownst to me, there's a cool literature in cs already exploring some of these procedural questions, so I don't think I've discovered anything new here (one main example: a Du et al 2023 paper showing that councils outperform individual models). That being said, this is a cool opportunity for people in the social sciences who study collective decision-making mechanisms and governance. In the future there will be tons of situations where multiple agents have to interact and make decisions together, and we should develop a science of how to make optimal collective decisions with AI agents.

Repo: github.com/mattzh72/minec… Made in collaboration with @Mattzh1314 @johnathanchewy and @prestonb0urne

Open-source Minecraft world builder. Introducing MinecraftLM.

excited to share our explorations with gemini 3 flash check out our repo in the thread!

1/ All the structures you see here were generated by an LLM from a single text prompt (notably, Gemini-3 👀). Over the past week, we’ve been researching the spatial understanding capabilities of agents and how they can build large scale, playable environments (particularly Minecraft to start). We found they do extremely well despite what many think. We're open-sourcing our work for others to build on top of github.com/mattzh72/minec…

Coming soon to @withcascade I think you'd love this @rauchg @ryolu_

Prototyping has never been this fun. Spin up multiple explorations from a single prompt, each with a live preview and watch claude code bring them to life.

branch to explore → merge → repeat light mode vs. dark mode explored @withcascade

Agents don't just do tasks. They help you move across timelines. Excited to share our take on a new kind of environment that enables creativity for web development.

We're heavily bug bashing this new iteration but it's live now at withcascade.com! @johnathanchewy and I can't wait to hear what you think.

The last few weeks of nights and weekends have been a labor of love. v0.3 of @withcascade is the closest we've come to a tool that feels truly explorative. Create n branches → visualize changes → merge an exploration back in → repeat.

@mattarderne @johnathanchewy Hey Matt! We just pushed an update to fix this. Once you restart the app to trigger an update, let us know if it works. DM me if you run into any issues :)

@johnathanchewy @avinashj_ All good, keen to try this!

We built a visual canvas for tracking changes and experimenting with claude code. Describe your change → auto-commit with git → view a preview of that change → drag and drop a new branch to keep experimenting. It's called Cascade and here's how it works...