Tim Darrah (@DrTimDarrah) profile picture

Assistant Professor of Practice, Vanderbilt. Co-Founder of https://teapawt.net/t.co/s5XWF9MsN6. Army veteran, former NASA Fellow. Building mission-aligned AI systems.

Nashville, TN
Joined August 2024
Photo by Tim Darrah (@DrTimDarrah)Photo by Tim Darrah (@DrTimDarrah)
Last week I gave a friend my agentic chief-of-staff. The next day she texted me that it had made a "huge, huge, huge huge huge difference" in how she uses @claudeai. The impact is clear, because today she thanked me again. I said I wouldn't be much of an "AI guy" if I couldn't whip something up to help my friends. Like many of us, I have a very busy life. Between teaching and running a startup, there are students, clients, investors, partners, courses, contracts, three calendars, and a family. I never had the time to build an assistant, which is ironic, until now. As part of our AI 5100 course, we built an executive assistant agent with skills, tools, MCP connectors, hooks, a knowledge graph to manage information, and an identity. Many students have already shared how transformative the AI assistant is. The version I have is different, though. It learns over time. I don't just mean it maintains a corpus of markdown memory files either. The knowledge graph dynamically updates, and a reinforcement learning loop runs on our local server. This bleeds into recursive self improvement, which is the danger zone of AI capability development right now and the hottest topic in the field. I talk a lot about the dangers of AI in an article I wrote a couple of weeks ago. (pinned post) When I think of where we are right now, I wonder, Where is Asimov? He gave us the Three Laws of Robotics 84 years ago. Where are the Three Laws? I am working to build them and would love to meet other builders that want to do the same.
1
2
274
Let’s talk about AI!
You’re invited to join our live webinar and AMA hosted by @DrTimDarrah and @davidtbuttrey to learn about… The future of Finance in an AI world. What we’ll cover: 1. How finance is moving into agentic finance, where AI does not just analyze markets but acts in them 2. How digital assets really work, and where they meet AI 3. The risks, guardrails, and questions to ask before using AI in your workflow RSVP below and save the date! December 10, 2026 · 12 PM ET mangrove.ai/webinars/
1
12
(1/5) AI guardrails are mechanisms that enforce organizational policies and procedures outside the access and control surface of the LLM. There is no shortage of incidents to highlight the importance of getting this right.
1
1
38
(4/5) That's what I'd ask a vendor to demonstrate. Give the agent a task that conflicts with policy. What happens if it misreads the rule or hands the task to another agent? Show me where the system blocks the action, and not where the agent can alter the system.
1
11
(5/5) I wrote about what institutions should look for when evaluating AI agents, including how we keep enforcement outside the agent's control at @Mangrove_AI. drdarrah.substack.com/p/ai-g…
1
17
Think AI coding agents will replace software engineers? Think again. I agree with Armando Solar-Lezama from @MIT_CSAIL. Teaching in @VanderbiltU AI Master’s program and undergraduate Data Science program, I see the shift and potential most people don’t. AI is going to make better problem solvers and allows us to teach higher level abstraction and critical thinking, forcing students to understand the material at a fundamental level, rather than surface level. I used to spend about 6 weeks teaching introductory coding, now it’s only 2. Instead of coding, it’s now prompting, skills, context, and knowing how to design and specify scalable and maintainable software. youtu.be/SBGpr1fPyb4?is=xp3n…
1
1
2
156
Me: "Why didn't you follow the repo conventions specified in the CLAUDE.md file?" agent: "The instruction was there in plain words. I didn't follow it. There's no deeper reason than that."
32
I think about AI failures in three categories: 1) misuse 2) misunderstood use 3) misguided use If someone uses AI to cause harm, that's a different problem from someone trusting it to do something it can't. Identifying what failure mode you are dealing with should be a key step in any post mortem.
1
49
AI isn’t dangerous. Misunderstood use of it is. drdarrah.substack.com/p/the-…
1
28
Tim Darrah retweeted
🚨 $HBAR Investors!! We’re about to hit the STRATOSPHERE!! 🚀 Soon, everyone’s gonna say “I wish I bought it at $0.30” 🤣 Will that be you? 🤔 🚀 🔥 💴
87
552
23
2,345
186,514