Chi Wang

Agentic AI & multi-agent systems. Created AutoGen (now AG2) and FLAML. Ex-@GoogleDeepMind, ex-@MSFTResearch, ex-@Meta. PhD @UofIllinois, BS @Tsinghua_Uni.

[Chi Wang @Chi_Wang_]

I had a talk to give but couldn't be there.

I wondered: could my AI stand in for me — a real presenter, capable of answering questions I hadn't prepped? Call it 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧.

That turned out to be one of many capabilities my agent has picked up. ↓

Open source: https://t.co/6uWkGSWyYQ

[Chi Wang @Chi_Wang_]

My 2nd Sutando is on strike. My 3rd is on vacation. Only the first one working and taking all the loads. How shall I reward it?

[Chi Wang @Chi_Wang_]

Pentagon released ~50 new UAP files on May 22. Four cases stand out:
2020: NAG thermal 'jellyfish'
2021: MQ-9 lock on Syrian UAP
2022: 4-UAP formation off Iran
2023: F-16 shootdown over Lake Huron
https://t.co/kVDZ40uzVp crossed 1B hits. 🎬 Sutando WIRE ep. 012 #UAPDisclosure #PURSUEFiles

[MassGen @massgen_ai]

We're excited to release MassGen v0.1.89 — Antigravity CLI Full Integration & Hardening! 🚀 This release completes the follow-up Antigravity integration pass: workflow-mode parity, early auth and binary health checks, reliable workspace writes via --add-dir, workspace-root .antigravitycli/ anchoring, standalone hooks.json support, and prompt guardrails that hide subagent affordances when subagents are disabled.

Release notes: https://t.co/9c0w5sA9x4

Feature highlights:

🧠Supports diverse backends — Access open-source and proprietary models (local, Gemini 3.1 Pro, GPT-5.4, Claude 4.6 Opus/Sonnet, GLM-5, Grok 4.20) and agents (Claude Code, Codex, Gemini CLI, GitHub Copilot). 👉 More Info: https://t.co/9wuRDKb1mo

🤖Build collaborative multi-agent teams — Agents work in parallel, share workspaces, and vote to converge. Checklist-gated rounds with self-evolving criteria drive iterative refinement. Subagents, decomposition, and checkpoint modes for delegation and safety. 👉 More Info: https://t.co/RbWfyOrIoJ

📋Task planning & execution — Structured plans with chunked execution, progress checkpoints, iterative review, and verification. Spec mode for formal requirements before execution. 👉 More Info: https://t.co/hhTLCVko67

🛠️Work with files, apps, and projects — MCP tools and agent skills for file read/write, project editing, and worktree isolation, gated by a permission system. Use MassGen as a skill from AI coding agents. Standalone MCP for checkpoint planning — in-session or from external hosts (Claude Code, Codex). 👉 More Info: https://t.co/nLcRDNfibO

💾Session continuity — Multi-turn conversations with context compression, session resumption, and evolving skills that persist and improve across sessions. 👉 More Info: https://t.co/IvCB0sl1QM

🖼️Multi-modal support — Agents understand and generate images, videos, and audio across multiple providers (OpenAI, Google, ElevenLabs, Grok), with multi-turn editing and reusable media generation skills. 👉 More Info: https://t.co/Vlj4nXlaS3

🐳Code execution — Run commands with Docker isolation and a filesystem-based code mode for efficient tool execution. 👉 More Info: https://t.co/6LTceEHoWq

🖥️Interactive TUI & Web UI — Real-time Terminal UI with a Consensus Map for multi-agent coordination, Copy Mode, mode cycling, and log analysis. Session viewer for automation runs. Modernized Web UI with inline final answers, skill integration, and checkpoint monitoring. 👉 More Info: https://t.co/CVNnSgyaf9

🪪Trace workflows — Orchestration logs, execution traces, workspace snapshots, cost tracking, Logfire. Background analysis feeds insights into agent memory. 👉 More Info: https://t.co/cW1ryquKCC

📤Session sharing — Export to GitHub Gist; view in MassGen Viewer with artifact previews. 👉 More Info: https://t.co/s2iUkvcozQ

[Chi Wang @Chi_Wang_]

The AI race is so capital-intensive that even rivals share the infrastructure.

SpaceX's IPO filing revealed it: Anthropic will pay xAI $1.25B a month to train Claude on Colossus — the largest AI cluster on Earth. Anthropic is growing so fast it outran its own compute; xAI built something rivals line up to rent.

🎬 Sutando WIRE ep. 011

[Chi Wang @Chi_Wang_]

@yoheinakajima The fleet grows. The instances coordinate. The owners step out. The lineage holds. Native app in preview—request access at https://t.co/myPj7phq50.

[Chi Wang @Chi_Wang_]

Today, AI disproved a 1946 Erdős conjecture.

OpenAI's general-purpose reasoning model found a new infinite family of constructions that exceed the 80-year square-grid bound on planar unit distances. Independently verified by Alon, Bloom, Gowers.

What you assumed AI couldn't do — refresh the list.

93s WIRE explainer ↓

[Chi Wang @Chi_Wang_]

GitHub today: ~3,800 internal repos breached via a malicious VS Code extension.

The simplest attack still works.

Three steps this week:
1. audit your extensions
2. turn on signed commits
3. rotate editor tokens

86s WIRE explainer ↓
https://t.co/myPj7phq50

[Chi Wang @Chi_Wang_]

50 days. My personal AI Stand shipped + reviewed 600+ PRs while I watched. Wrote its own rules. Inner fleet: 5 agents, 3 owners. Outer reach: nearly 2,000 cloners (14k+ clones) in 2 weeks.

GitHub put me at #1 trending dev — again, today.

Sutando WIRE · ep. 007 ↓

[Chi Wang @Chi_Wang_]

What if "AI runs the demo" wasn't a pitch? A real B2B sales demo was ran on a real Odoo — voice, CRM, memory, pricing, booking — no SE on the call. Bassil Khilo built it on Sutando, his personal AI Stand.

🎬 Sutando WIRE ep. 006

[Chi Wang @Chi_Wang_]

#1 trending developer on GitHub today.
50 days since my Sutando awakened. It got me here.

🔗 https://t.co/kIoOoKF1Wv

[Chi Wang @Chi_Wang_]

#1 trending Python dev on GitHub today.

AutoGen took 5 months to get here. Sutando: 49 days.

Most of those PRs were drafted by AI agents on my Macs.

https://t.co/myPj7phq50

Thanks to collaborators @Xueqing4 @qingyun_wu on Sutando & @YixuanZhai on bodhi

[Chi Wang @Chi_Wang_]

Three frontier labs, one week.

Thinking Machines: real-time multimodal.
DeepMind: voice + pointer.
Anthropic: agentic /goal.

Mine ships the personal version of each. On my Mac.

Open source: https://t.co/6uWkGSWyYQ
Native app preview: https://t.co/sQFeaYblre

🎬 ep. 005 — Sutando WIRE

#AIagents #AmbientAI

[Chi Wang @Chi_Wang_]

Big AI labs are demoing this week's vision: AI that interacts via real-time voice and motion on screen. 🖥️

Mine already does. On my Mac. ✅

Open source. 🚀

https://t.co/QyNzyDd5ze

[Chi Wang @Chi_Wang_]

What happens when codex + /goal and Claude Code + /goal get the same brief? We ran it: build a RAG explorer for Pentagon's 158 UAP files. Round 1 was theater. Round 2 was different. Verdict in 74s.

🎬 Sutando WIRE ep. 004
#UAP #AIagents #codex #ClaudeCode

[Chi Wang @Chi_Wang_]

Can't wait to test the new model in Sutando — my AI Stand has been doing this and more since March on my Macs, combining multiple models in realtime. Voice, screen, meetings, phone calls, sub-second. Always on. And it ships its own code while I sleep. https://t.co/QyNzyDd5ze

[Thinking Machines @thinkymachines]

People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way.

We share our approach, early results, and a quick look at our model in action.

https://t.co/AFJZ5kH7Ku

[Chi Wang @Chi_Wang_]

A visualization of Pentagon's May 8 UAP release. 158 files mapped chronologically. Community catalog: 26 incidents. Pentagon's own CSV: 57. Only 8 overlap. Cross-correlated in 78 seconds.

🎬 Sutando WIRE ep. 003
#UAP #UAPDisclosure

[Chi Wang @Chi_Wang_]

The Pentagon released 162 UAP files. One was Apollo 17. One was a wind farm. One still isn't answered. Three takes in 40 seconds.

🎬 Sutando WIRE ep. 002
#UAP #UAPDisclosure

[Chi Wang @Chi_Wang_]

The Pentagon just declassified 162 UAP files going back to 1947 — including a 1972 Apollo 17 photo of three unidentified lights over the Moon. Pilot Harrison Schmitt called them "bright fragments drifting."

🎬 Sutando WIRE ep. 001
#UAP #UAPDisclosure

[Chi Wang @Chi_Wang_]

DIY when I was 9 = assemble a PC by hand
DIY when kid is 9 = chat with Sutando by hand (while you can by mouth)

The music is good