I’ve been listening to a lot of friends discuss the amount of money that’s being spent to build data centers and related infrastructure for AI. Some of them are investors and others are entrepreneurs, so I hear varying perspectives on whether the current and proposed levels of spending make sense. And after listening to Kevin Koharki explain how stock-based compensation hits shareholders twice and reduces cash flow per share significantly (see here), I began thinking about it from that perspective as well.

I’m not sure what to make of this boom in spending on AI, but I want to understand it better because I think its impact on the economy and society will be massive. Luckily, this isn’t the first time we’ve had a massive spending boom driven by new technology. That being the case, I think the best thing for me to do is study history to understand what happened so I can better understand today’s reality. I have some historical books on this topic that I’m going to crack open.

I watched a few interviews where founders mentioned that physical AI and robotics are the next frontier in AI. I’m not very familiar with either, so I started digging. It’s early, but the more I learn, the more I can see why they believe that incorporating AI into robots and other devices that execute complex actions in the real world is a big deal. Especially when you consider repetitive, dangerous, or labor-intensive tasks being handled by machines or robots. That’s an enormous market.

I’m still learning, but so far I’m intrigued. I don’t have a deep enough understanding to share insights at this point, but anyone interested in finding out what physical AI is or why it’s important should check out a great article from Nvidia about generative physical AI here.

Yesterday I finished my second AI workshop at Georgia Tech. This session was focused on learning how to use the AI agent and workflow tool n8n. The person leading the session has built a sophisticated e-commerce company that’s automated on the n8n platform, so he was very knowledgeable.

n8n reminds me of workflow and conditional logic tools I’ve used in the past, but it’s way more powerful. The ability to connect it to foundational models like those from OpenAI (ChatGPT’s parent company) is huge. Last night, we linked our n8n workflow to OpenAI models via API, connected our n8n workflow to Telegram so we can give it written instructions, and built an OpenClaw-type setup in n8n.

Working with these tools hands-on with other people who are motivated to learn them makes for a wonderful environment. I met several great people last night, and I’m looking forward to the next session.

Last week, I shared that I wanted to start attending hands-on AI sessions to accelerate my learning and be around others trying to do the same. This week, I attended my first evening session at Georgia Tech. It was free and open to the public. This session was focused on using AI tools to create high-caliber images and videos.

My takeaway is that the session was exactly what I needed. It was hands-on, with everyone helping each other figure out how to use the tools and sharing what they created. I learned a ton and met some good people. I definitely want to do more of this.

Continuing with my new protocol, here I’m going to share content I consumed and learned from.

What I struggled with:

• No material struggles this week

What I consumed this week and what I learned from it:

• Claude Managed Agents = office space of AI agents – YouTube interview detailing how to think about managed agents and when to use them, with detailed explanations and a demo of the Claude Console.
• Claude Managed Agents explanation – YouTube video that explains how Claude Managed Agents works and demos them. Quick, short video.
• $6M to $32B acquisition by Alphabet – YouTube interview with Gili Raanan, Founding Partner at VC firm Cyberstarts. Interesting interview where Raanan discusses his strategy for investing in early-stage cyber companies and why his unicorn hit rate has been so high. Raanan doesn’t ask about the founder’s idea at all. He invests in the person. And he invests only in people who have shown the ability to recover from setbacks.
• The Roblox economy turning devs into millionaires – YouTube interview with the Founder and CEO of Roblox, David Baszucki. Interesting interview where Baszucki describes how the Roblox platform works and how it’s really a small economy. His points about the value of the 3D data Roblox has are interesting, especially for robotics training. He explains how building AI-generated worlds is different from building interactive games. He makes a few references to parallels between YouTube and Roblox and how Roblox has a technical advantage over other game platforms.
• OpenAI’s 2029 projected growth, unique in history – YouTube interview with Michael J. Mauboussin, adjunct professor and investor. Fascinating interview where he discusses the importance of using base rates to evaluate probabilities. He shows how in 18,900 companies examined, no company the size of OpenAI has ever grown 108% compounded annually for five years, and how OpenAI’s 2029 $145B revenue projection would be a first in history. His ideas about intangible assets, AI, and expectations got me thinking.
• Secrets of investor athletes, and fear distorts reality – YouTube interview with Dr. Gio Valiante, world-renowned performance coach for professional athletes and investors. Founders of smaller funds and small investors can pivot and adapt to change faster than institutional investors, which is an edge. His thoughts on fear, motivation, and confidence were interesting. Confidence is inversely related to fear. Fear distorts our interpretation of objective reality. When you approach things from a place of fear, you see only the dangers, not the opportunities. Higher-performing investors operate at a level of optimization just like professional athletes’. I like his term “investor athletes.” Also, his point that high performers fall to the levels of their system (i.e., habits) or the culture and environment of their organization stuck with me.
• Thinking like an M&A buyer – YouTube Interview with Javier Enrile, Managing Director of M&A at TIAA. Interesting interview to understand how corporate buyers think about M&A. His thoughts on valuation and finding companies to buy caught my attention.

That’s what I consumed and learned from and struggled with last week.

Today a friend who’s also a former entrepreneur showed me something cool. He’s a technologist and loves to tinker. We talked about AI agents and how much running them continuously via Claude will cost, and he showed me his setup. He bought inexpensive (maybe $100) Raspberry Pi computers and set them up in his house on his local network. He installed the Claude app on them. He has Claude Cowork build software that solves a specific problem for him and then install the software to run locally on a Raspberry Pi.

My friend then uses Raspberry Pi Connect to monitor his agents and access the software that’s running locally on each Raspberry Pi machine. He likes this setup because it doesn’t require continuous work and token usage with Claude. Instead, Claude builds the software and the agent’s job is done. The software runs to do whatever task my friend wants. Because it’s running on a Raspberry Pi at his house, it doesn’t cost him anything to keep the software running 24/7. And he can access the software from anywhere, as long as he has internet access.

I thought this was an interesting setup to get around the increasing costs associated with using Claude and its agents.

The speed at which AI is moving and new products are being released is wild. People are still getting used to Claude Cowork, and yesterday Claude released its Managed Agents platform (see here). Like everyone, I find it hard to stay on top of it all. Actually getting around to playing with all these tools is even harder.

A friend told me about a group of AI engineers who work at companies like Meta during the day and do evening sessions on Wednesdays where they share the latest and explain it to the participants. On top of that, they do hands-on sessions where everyone practices setting up their own versions of the new tool(s) and building with them. The AI engineers help everyone, but participants also help each other.

I’ve decided I’m going to attend these sessions. Other people who want to stay on top of AI and learn the latest sound like a good group to be a part of. Plus, the idea of learning by doing suits me well, especially when I can do it with a group of like-minded people.

AI is real and will have a profound impact. The speed at which it’s moving is forcing me to take a different approach to learning about it.

Over the last few weeks, I’ve chatted independently with two entrepreneurs who sell physical products online. Both of them mentioned that TikTok is a core piece of their growth strategy. I wasn’t expecting that, so I asked a few questions. I knew that TikTok had launched TikTok Shop, where users could purchase products directly on TikTok. But what I didn’t know was how popular that service is. And I was unaware that TikTok has also launched Fulfilled by TikTok, a service that allows merchants to send items to one of 14 TikTok warehouses so that when an order is received, TikTok handles processing and shipping to the customer. It sounds just like Amazon’s Fulfilled by Amazon (FBA) service.

I’m curious to learn more about this. The idea of combining social media with commerce and fulfillment in a single platform intrigues me. Capturing the attention of a user and letting them complete a purchase in a few clicks without leaving the platform is a combination with huge potential.

Last week I listened to an interview with Annie Duke. She’s a former professional poker player who wrote two books on decision-making that I enjoyed (see here and here). During the interview, she explained why moving from implicit intuition to explicit decision-making, including expected-value calculations, leads to better outcomes in high-stakes situations. The gist of it was that instead of making decisions in our heads, we should write down our thinking and share the written document. The exercise of writing down your decision logic will help you spot errors (and avoid or fix them) and make it easier for others to spot flaws in your thinking.

I believe this to be true, but I don’t do it consistently enough with high-stakes decisions. I wanted to test it. Last week, I wrote, with the help of AI, an investment thesis for a publicly traded company. I then shared that thesis with two friends and asked them to discuss it with me after they’d reviewed it. Here are a few observations from the exercise:

• Writing down my logic and reading back through it helped me identify two points that contradicted a timeline that I’d proposed. When I did more digging to get more context around each point, I realized that my thesis was weaker than I’d thought, partly because of an update to my timeline.
• AI helped me create parts of this report, so going back and reading through it helped me catch errors in AI’s logic, too. AI made some irrational assumptions. When I pointed them out, it agreed and fixed them.
• Giving my decision logic to someone in writing, having them read it, and then discussing it afterward was a game changer. When we talked, they were totally up to speed on my thinking and were able to go straight to the areas where they thought my argument was strongest and weakest. The conversation was much deeper than if I’d tried to explain it to them orally.

This experiment showed me that Duke is spot on. When I’m making high-stakes decisions, it’s worth the effort to be explicit in my decision-making process by writing it down, reviewing it, and sharing it with others. The act of writing forces me to face holes in my thinking and allows others to easily see holes and point them out. All of this should reduce the chances of my making a bad decision.

If you want to listen to the section of the interview where Duke talks about explicit decision-making and calculating expected value, take a listen here.

I’ve been playing with the Claude for Mac app and using Cowork, its AI assistant. Overall, I like Cowork, but I can use it on my computer only when the Claude for Mac app is open. The issue with this is that it’s another app, one that isn’t in my normal workflow. Since I don’t naturally use it, I have to put extra effort into remembering to go to it.

The apps I use most are for email and text messages. So, I think it would be best to interact with Claude Cowork, or an equivalent product, via one of these; ideally, text messages. I need to do some digging on this, but I think that’s what I’m looking for: an autonomous AI task executor that I can interact with via text messages.