I’ll share which AI tools I’m exploring, some experiments I’m conducting, and insightful information about what I’m observing in the world every week.

🔧 Three Tools I’m Testing

👨‍🎨 Picsart Logo Generation - I used this to create the inspiration for the logo for this newsletter. I then took that and recreated a version in Adobe Illustrator that turned into the final version. Overall, it’s a cool product that expands well beyond logo generation.

🧩 Wordware Lego Image - A tool that provides a wide variety of model agents that can be used for business and fun. I used it to recreate some of my photos at Lego People. It’s worth seeing what other people are using it for.

🗒️ Fathom - I recently purchased this tool to record and automate my note-taking in business meetings. The output has been fantastic! I am very impressed with the meeting summary and the action items. If it’s acceptable in your business, you should try this out.

🧪 AI Experiment of The Week

I decided to build a web application using Replit this week. Replit uses AI code generation and a chat interface to build apps. It all starts with a prompt. My prompt was:

Then, Replit spends a few minutes building web application scaffolding, including creating a Postgres database and React code for the core application. Seriously, this took about four minutes. Then, it has a working prototype of the most basic features (dashboard and create campaign). I went back and forth with Replit for about fifteen more minutes, testing the application and asking it to build more features (campaign recommendations). The application is about 15% of what I envision, but it’s getting there quickly. I decided to deploy it to test the process. You choose the autoscaling rules, and it’s off and running (reminds me of my Xervo days). The application is live (but still private) a couple more minutes later.

Overall, I’m very impressed with the capabilities. I could easily build real production applications in Replit (and will). Even if you’re not a developer, give it a try - I’d love to hear your feedback on it.

📰 Article of The Week

DeepSeek R1 Research Paper - research paper that outlines how DeepSeek trained its model.

DeepSeek is the talk of the world this week. The introduction of DeepSeek R1 has shocked AI companies and shaken the public tech market (NVDA ⬇ 6.4% in the last 30 days). Its impact on the market has been dramatic because the claimed cost of training the model was roughly $6 million. This is a small percentage of what it costs to train a model from OpenAI. The reason it can be trained so cheaply is because it’s relying heavily on distillation for training. Distillation uses other models to “teach” the new model; there are heavy claims that DeepSeek even used OpenAI models to train its model. I think that DeepSeek is an excellent innovation in LLM training and will end up helping reduce overall costs. We’ll see their techniques being used in a variety of manners. Scott Galloway has a solid take I agree with on his Prof G podcast (first ~10 minutes).

I’ve tried the model using generic prompts and found it decent but not nearly as good as the top models from Anthropic and OpenAI.

I would strongly advise against anyone using the online DeepSeek chat at this moment. Wiz security researchers have revealed that private data has been compromised. If you’re interested in testing this out, I suggest using Perplexity, which hosts the model separately.

🌎 Where the World is Going

The AI revolution reminds me of the early internet days—not just in its potential scale but also in how we're all struggling to grasp its full implications. When I talk to colleagues about AI agents, there's often a laser focus on large language models and chatbots. But watching developments in IoT and edge computing, I'm convinced we're only seeing the tip of the technological iceberg.

What fascinates me is how AI is quietly transforming industries far from the spotlight. In my work with IoT systems, I'm seeing small, specialized machine learning models¹ revolutionizing everything from predictive maintenance to resource optimization. These models, running on limited compute at the edge, solve real-world problems that massive language models can't touch. This is a reminder that AI's future isn't just about generating text or code—it's about embedding intelligence everywhere it can make a difference.

The real magic, though, will come from combining these technologies. Just as the internet became transformative when we learned to layer protocols and applications, AI's true potential will emerge as we blend different types of models and approaches. Those who learn to orchestrate AI agents - combining the linguistic capabilities of LLMs with specialized models for specific tasks - will have an advantage similar to early internet adopters. The question isn't whether AI will transform your industry but whether you'll be ready when it does.

👨‍💻 About Me

🔗 Extra Links to Check Out

• NFX Article on AI Agent Revolution 💞

• Darmesh from Hubspot’s Agent.ai 🤖

• Block’s Goose 🪿

👣 Footnotes

1. TinyML - ML for resource-constrained devices