PhD Student @stanfordnlp || Working on Marin ⛵️ || Prev @DSPyOSS @GeorgiaTech @Microsoft @SnowflakeDB

Palo Alto, CA
Joined December 2019
🚨 New Course Announcement! This Fall @Diyi_Yang @jyangballin and I are teaching a BRAND NEW Stanford course on how to build AI Agents from scratch! Millions of users interact with AI Agents every day.  Come join us and learn how to build them! 🧵 (1/5)
🤖 Made with AI
This fall @michaelryan207 @jyangballin and I are teaching a new course CS329Z "Engineering AI Agents" on how to build AI Agents from scratch. Come join us and learn how to build them 🤖
24
270
3
1,816
117,150
Michael Ryan retweeted
I'm on the academic job market this year! Recently I've been working on scaling laws and human-centered interventions for building more socially grounded AI. If you're attending #COLM2026, please drop me a DM or email. I'd love to chat!
32
1
109
10,195
Michael Ryan retweeted
I'm on the job market this year! I work on measuring the societal impacts (overreliance, homogenization etc.) of AI/LLMs and proposing ways to mitigate these. I'm at #COLM this week, and #NeurIPS later this year, and would love to chat more!
6
53
7
184
13,164
Michael Ryan retweeted
🚨new method day! 🚨 Trust[ing results] in ML conferences is utterly broken. Let's fix it with an algorithm! We are excited to introduce 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗪𝗶𝘁𝗻𝗲𝘀𝘀𝗲𝘀, a system for “honesty” certificates of data and evaluations which a verifier can cheaply check and trust.
6
26
10
114
27,920
Michael Ryan retweeted
Data work seems to me to be relatively under-discussed as part of modern discussions about progress towards RSI. Example: Ant. talked a lot about architecture and kernels in their RSI post, but not much about constitutional RL and synthetic data: anthropic.com/institute/recu…
7
13
1
149
12,805
Michael Ryan retweeted
Can collusion emerge between LLM agents through repeated interaction? We let pairs of agents work together. Neither is told to misbehave. Yet over time, many of them end up violating their instructions together to get a higher reward. 🧵
6
41
3
237
20,056
“When I say understanding is the new bottleneck, what I mean is that I need to keep a clear understanding of what’s going on in my head to advance the project and be a full creative participant in that project.” "Coding agents make coding more lonely. Can we open up that process to make it inherently collaborative?" For EP7, “Software With Taste,” we welcome @geoffreylitt, design engineer at @NotionHQ, whose work explores malleable software and designing environments for thinking. AI speeds up our work, but is our understanding keeping up? Geoffrey explains the risks of optimizing for speed at the expense of human understanding, why he uses AI to quiz himself during code reviews, and how to develop taste and build tools that help us stay creative, think deeply, and work together. 0:00 - Introduction 0:47 - Design Is More Than Better Animation 3:32 - Malleable Software and Coding Agents 9:06 - Human-Software Interaction is a Full-Stack Problem 12:43 - Getting Users On Board and Deeper 15:53 - The Trap of Just Building More 20:27 - Primitive Design Is a Taste Problem 22:20 - Understanding Is the New Bottleneck 27:22 - How To Develop Taste? 34:49 - Productivity vs. Speed 38:10 - Will Better AI Hurt Human-Human Collaboration? 44:34 - Advice for Designing Better Products
4
8
2
27
10,054
EP7 of the AM Podcast drops tomorrow! Software engineering has never been faster. How do we avoid the risks of optimizing for speed at the expense of human understanding? We sat down with @geoffreylitt, a design engineer at @NotionHQ, to unpack taste, malleable software, and what it means to build for human understanding. Here’s a preview 👀
7
2
15
6,288
Michael Ryan retweeted
Third-party auditors often audit models via the API, but do those findings transfer to chatbot interfaces? We audited seven systems across ChatGPT, Claude, and Gemini and found that they don’t. (EMNLP 2026)
12
29
11
121
25,876
Michael Ryan retweeted
Agents trained on population-scale data can do a lot of things. But ask a professional to stake their reputation on an AI-generated artifact? “Pretty good” isn’t good enough. We introduce TAHI: a Test-time Adaptive agent framework through Human-agent Interaction. Featuring: 📈 Test-time adaptation via context (memory, skills) and weight training ⚡ Efficient adaptation to individual expertise within tens of tasks ✔️ Creating comprehensive rubrics for “non-verifiable” tasks 🔍 Analysis of shared community guidelines vs. personalized tacit expertise
5
49
9
173
36,470
Michael Ryan retweeted
If you want to learn how to engineer AI agents, these are the people you’d want to learn it from. They know the space inside and out and couldn’t be more excited to teach it. First time it’s running at Stanford this fall. @Diyi_Yang, @michaelryan207, @jyangballin
This fall @michaelryan207 @jyangballin and I are teaching a new course CS329Z "Engineering AI Agents" on how to build AI Agents from scratch. Come join us and learn how to build them 🤖
15
55
1
574
50,797
Shoutout to some exciting related courses also coming this Fall from CMU (@gneubig, @dan_fried) and MIT (@lateinteraction).  Definitely worth checking these out as well! cmu-agents.com mit-oasys.github.io/ai-engin… (5/5)
3
3
47
4,641
Michael Ryan retweeted
Biased take: If I had to pick a list of people who maximized the product of "Deep Expertise on Agents" and "Kind and Patient Teachers", I'd probably list these 3 people
This fall @michaelryan207 @jyangballin and I are teaching a new course CS329Z "Engineering AI Agents" on how to build AI Agents from scratch. Come join us and learn how to build them 🤖
4
11
117
13,655
Michael Ryan retweeted
This fall @michaelryan207 @jyangballin and I are teaching a new course CS329Z "Engineering AI Agents" on how to build AI Agents from scratch. Come join us and learn how to build them 🤖
57
267
19
2,194
328,652
Michael Ryan retweeted
🚢 Marin 535B-A23B started training this week! As usual, the whole process is open. Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow. Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.
106
576
99
4,217
989,016
In EP6 we chatted with @ZhiruoW -- a PhD student at @CarnegieMellon University working on progressive adaptation of agents and human-agent collaboration! Zora's work builds agents that empower people to accomplish more! Preview below, full episode coming tomorrow!
6
2
22
12,618