Head of Tokyo AI Innovation Lab @supercell ← PixAI←日本代表 @ Stability AI ←Indeed←SmartNews←Supercell←Google←Wharton←起業←Stanford; posting on GenAI in 日本語+English
Angelo 2.0 is out!!!
Angelo is Blender with an AI agent built in. Describe what you want and it builds it with you, in the scene, step by step, while you stay the artist. 2.0 moves Angelo to Blender 5.3, makes it faster where you feel it first, and starts every new chat thinking harder. Free for macOS and Windows, with Linux coming in 2.1.0.
Faster
• Your first Cycles render on Apple GPUs takes about half as long as in Blender 5.3.0: 77 s instead of 171 s on an M4 Pro.
• Switching to a new world no longer stalls the view: on Windows with an NVIDIA T4, the first draw fell from about 1 s to 71 ms.
• Lighter when idle: no more background redraws every two seconds, and less memory for previews.
Better results out of the box
• New chats start at high reasoning effort.
• The agent knows Blender 5.3's Python API, including the layered compositor.
• Claude chats on an API key or a team keep their prompt cache for an hour, so coming back after a break costs less.
Built on Blender 5.3
• Stackable compositor effects.
• Vulkan by default on Windows, with OpenGL one setting away.
Teach me as you work
Turn it on from the chat bar's + › Learn, and every step Angelo takes shows the command, its shortcut in your keymap, where it is in the menus and the settings it used, in your language.
Safer agent scripts
Scripts now run confined: they write only to Angelo's folders and places you chose, and can't read your keys, start programs or use the network. Full protection comes in 2.1.0.
Your 1.0 chats, snapshots and settings come along on first launch. Plus 48 fixes.
Fast and lean by design
Over everyday tasks with Claude Opus 5.5, Angelo finished 2.4× faster than Blender MCP on the same model and prompts, at 51% of the cost, and both sides passed all 25 graded runs.
For teams and enterprises
Free and open source for individuals: bring the Claude or ChatGPT subscription or API key you already have.
Team:
• Your organization's provider keys held by Angelo's gateway, encrypted with Cloud KMS, never on laptops
• Seats only for the artists who use the app; admins and reviewers don't need one
• A shared library and IP bible the agent can search, import from and save to
• Image generation in the console, usage by person and model, and a log of every change
• Read-only access to the library for your own agents over MCP
Enterprise adds:
• Single sign-on: Okta, Microsoft Entra ID, Google Workspace, any SAML or OIDC provider
• SCIM provisioning, audit log export and a dedicated encryption key
New in 2.0: run the agent through your own model server (OpenAI Responses, Gemini or Anthropic Messages format; tested with LiteLLM).
We never resell model usage, your work isn't used to train models, and your scenes stay on your artists' computers.
Play like a kid. Sculpt like a master.
Download: angelo.studio
Teams: angelo.studio/enterprise
this is hands down one of the craziest things I've ever made with AI.
@AnthropicAI's Fable 5.2 built a swarm of 1,048,576 nanobots that coordinate and morph into anything I pick: a hand, a whale, a rocket, a DNA helix...
it designed the bot first: a 3.4 mm, 14-sided body with latch pads and light emitters. then the room. then the swarm.
every single bot is simulated on the GPU. when I pick a new form, the old one peels apart in a wave, each bot flies to its own slot, and the new form grows out from one seed point.
one prompt. @threejs. 60 fps at a million bots, and it goes up to 4.2 million. 🤯
PSD2Live
Turn a layered PSD into a moving character, then rig and animate it
Import a PSD and pick a preset to get idle motion, blinking, nodding and hair physics within minutes. Shape, rig, simulate and animate the character in one workspace.
This open source editor includes skeletons, IK, cloth and hair simulation. Export to Cubism, VTube Studio, the web, game engines or video.
A local MCP server lets agents such as Claude Code and Codex edit projects.
github.com/tsunehimatoi/psd2…
Claude Opus 5.5 Medium:
Turn this photo into algorithmic string art: square frame, ~400 nails along the edges, ~4000 black threads, greedy line selection. Auto-detect the face/main subject and concentrate threads there. Make a 1:1, 1080×1080, 20-second process video: slow at first, then accelerating; highlight the active thread in red and show a counter; end with a brief wipe revealing the original. Also give me the final PNG and the nail sequence.
We’re introducing Shizuku-Duplex, the world’s first full-duplex audio synthesis model built specifically for AI characters. See how our models are built for natural human-AI interaction: