GNI Postdoctoral Fellow @TUM | Human-AI Alignment | Building ThoughtDAG, an open-source canvas where wires are the context.chenxiachan.github.io/thoughtdag/ Munich, GermanyJoined September 2017
I invited a few friends to try ThoughtDAG with real questions.
Each process left a different shape behind. 🧠
One deep thread.
Five roads, four dead ends.
Three weeks around one question.
A literature review season.
Export yours:
github.com/chenxiachan/th…
One wrong note changed the LLM’s recommendation.
Same model. Same question.
I removed one wire from the context graph and asked again. The answer changed.
ThoughtDAG makes LLM context visible and editable.
Would you control this manually, or automate the cleanup?
Can one wire change an AI answer?
In ThoughtDAG, wires are not decoration. They decide what the model actually receives.
This is a real API run: remove one edge, regenerate, and watch the context and answer change.
Open source: github.com/chenxiachan/th…
Across 72 model-cases first derailed by pollution:
• delete the source: 68/72 repaired
• delete + recompute: 71/72
• remove the contaminated subgraph: 72/72
The consequence sometimes mattered more than the original mistake.
I deleted the original mistake from an LLM conversation.
Two models still gave the same wrong answer.
The error had already been repeated downstream. Removing the source was not enough.
A small pilot on how context errors survive:
Yesterday I posted ThoughtDAG on Hacker News.
52 comments later, the idea is the same, but the roadmap is clearer:
• show context as a list, not only a graph
• fix canvas friction first
• explore a read/write API for existing tools
github.com/chenxiachan/th…
Most AI chats let you regenerate an answer, but not repair it.
If 90% is right, I should be able to fix one sentence and keep thinking.
So I made every answer in ThoughtDAG editable. Your correction becomes part of the context downstream.
github.com/chenxiachan/th…
New Graph Engineering Tool for LLMs! ThoughtDAG: the conversation is a graph on an infinite canvas. What the model sees is exactly what you wired in. github.com/chenxiachan/th…
It's open source, local-first, and honestly still an early prototype built by one person.
The example canvas needs no signup and no key, so you can just poke at it. I'd love to know what breaks.
github.com/chenxiachan/th…
The part I care about most: nothing here runs on its own. No agent quietly rewrites your graph.
You read the papers, you highlight what matters, you decide what merges into what. The human stays in the loop, the model stays on the wires.
Last month I lost my best idea in a 60-turn chat with an AI. It was in there somewhere. I just couldn't scroll back to it.
I'm a researcher, and I got tired of that. So I spent a few months building a different way to think with a model.
971 Followers 4K FollowingEmerging AI partners and paradigms @NVIDIA; Ex: Neuromorphic AI @Intel, TPUs @Google; Strong opinions, loosely held are mine. Aiming for a stable singularity.
1.6M Followers 2 FollowingWe're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
971 Followers 4K FollowingEmerging AI partners and paradigms @NVIDIA; Ex: Neuromorphic AI @Intel, TPUs @Google; Strong opinions, loosely held are mine. Aiming for a stable singularity.
14K Followers 2K FollowingAssociate Professor at Harvard & Kempner Institute. Applying computational frameworks & ML to decode multi-scale neural processes. Marathoner. Rescue dog mom.
378K Followers 1K FollowingCo-founder of stealth startup. Inventor of GANs. Lead author of https://t.co/M6vl8pEQ4I Founding chairman of @pubhealthaction
163K Followers 41 FollowingSydney Dec 6-12, 26, Paris and Atlanta. Tweets to this account are not monitored. Please send feedback to [email protected].
4K Followers 259 FollowingHelmholtz Artificial Intelligence Cooperation Unit
Research-driven hub for applied #AI & #ML
Democratising AI across @helmholtz_en
1K Followers 177 FollowingThe MCML is a joint research initiative of @LMU_Muenchen and @TU_Muenchen to strengthen competence in the field of AI and to make potential accessible.
174 Followers 30 FollowingMDSI - an integrative research institute @TU_Muenchen focusing on #DataScience #ML and #AI and their interdisciplinary applications.
1K Followers 232 FollowingFor all enthousiasts interested in the latest discoveries and upcoming events in #computerscience, #artificialintelligence, and #datascience in Leiden.
214K Followers 0 FollowingOG of: P and T in ChatGPT, 100x deeper learning, meta learning and RSI, neural distillation, GAN/World Model... Co-authored most-cited AI paper of 20th century
1.8M Followers 1K FollowingCo-Founder of Coursera; Stanford CS adjunct faculty. Former head of Baidu AI Group/Google Brain. #ai #machinelearning, #deeplearning #MOOCs
1.3M Followers 793 FollowingFounder/Chair, AMI Labs; Professor, NYU; Partner, 224 Ventures; Ex-Chief AI Scientist, Meta.
Researcher in AI, ML, Robotics, etc.
ACM Turing Award Laureate.
230 Followers 202 FollowingWe provide rigorous, evidence-based trainings focused on cultural humility, restorative justice, trauma-informed care, and anti-racism. berkeley + chicago + nyc
4K Followers 112 FollowingStudents at UC Berkeley working on academic research, ML education, industry projects, and fostering a vibrant ML community 🧠💡
12K Followers 991 FollowingDriving innovation in math, computational science, computer science, and data science. Home to NERSC & ESnet. Advancing research with HPC & cutting-edge tech.
29K Followers 119 FollowingThe official account for Stanford Online.
Stanford Online is the Stanford School of Engineering’s portal for academic and professional education.
12K Followers 998 FollowingOfficial account of the Stanford Woods Institute for the Environment: @Stanford's hub for interdisciplinary environmental research about people & planet
6K Followers 518 FollowingThe Stanford Humanities Center invests in experiences—fellowships, workshops, lectures, and other events—that advance research in and across the disciplines.