September Update

I haven't been able to post many updates in a while, so here's what's been on my mind lately. First off, the state of AI startups. What I've noticed so far is that there are a lot of AI wrapper companies, for example in code review or search. I currently work at Waymo (part of Google/Alphabet), and we have internal equivalents that are pretty much indistinguishable from what startups are building outside. Features, especially on the software side, are very easily replicable. There is merit in go-to-market strategies and building a sticky customer base, but as far as product features go, as soon as one company builds something great, you can just ask an AI coding agent to copy it into your existing service. We often see internal versions of external tools, and while the quality might lag at first, it catches up quickly. At that point, you have to ask: is it even worth the switching cost? In most cases, I don't think it is.

I've been tinkering around with AI coding tools to build different SaaS or B2B SaaS ideas. Every time I get something working that feels pretty solid, I look at what's already out there and compare two or three existing products. Even if I have a feature or two that stands out, I realize that if I were to go all-in on it, other companies could easily copy what I built. At that point, it becomes a race to the bottom. As everyone copies each other's features, the only differentiators left are undercutting prices or offering hands-on deployment engineering to manually customize integrations for each customer. That might be a selling point, but it doesn't feel like a compelling business model. It makes me question where the SaaS and AI space is heading next.

Another topic that keeps popping up in my mind is AI personal agents. We talked about this when AI first blew up a couple of years ago, but now there's a new wave. For instance, Meta has their Muse personal agent on phones, and Instinct, a relatively new startup, raised a ton of money in a short period due to rapid growth. I haven't tried Instinct yet, but I did try Muse. That brings up a recurring question: if another company or I built a personal agent, would customers find it worth trying? What would they actually use it for? That was my exact reaction when using Muse. Personal agents rely on connectors. The more integrations you enable with your email, calendar, or messages, the better context the agent has to do tasks for you. One convincing use case I've heard is having an agent scan emails and messages to find tax refund opportunities or forgotten invoices. Beyond that, I haven't found a truly compelling use for myself. Part of the problem is that I manage multiple inboxes and email accounts. For the ones I do have connected, there isn't much for the agent to do except manage package tracking emails or marketing clutter that I periodically unsubscribe from. Sometimes the information is already outdated. For example, Luxer sends me a code whenever a package arrives at my apartment's locker. By the time the Muse agent sends me a reminder about it, I've already picked up the package. I'm not gonna bother telling the agent I already got it. There's a lot of onboarding friction, and unless you go all-in and grant the agent complete control and context over your life, I'm not sure how useful it really is.

I struggle to see a clear long-term direction for personal agents. There are legit use cases where people save money, and I've personally saved money using ChatGPT. But I don't know if the current interfaces are right. If I had a dedicated human executive assistant 24/7 with access to my computer, how would they make my day-to-day life significantly better? I don't have a strong use case for that right now, though it makes sense for senior corporate leadership. Honestly, I can't clearly see where this space goes. On the other hand, traditional SaaS has seen a lower barrier to entry thanks to AI prototyping, but that also makes solutions trivially easy to replicate.

I've also been thinking about the business side of running an AI startup. Some companies achieve incredible ARR and growth, yet most of their revenue goes straight to neoclouds to cover compute costs. They might save at the margins using open-source models, but from a product perspective, differentiation is tough. Some argue that companies like Anthropic and OpenAI are constrained by their own first-party models, whereas startups can flexibly route between various models. But when it comes to getting the job done, model performance differences are narrowing. Model providers have lower costs because they own their compute and weights, while startups can leverage model variety. Yet without distinct product positioning or a unique UI, users have little reason to pick one startup over another. Another interesting counter-argument I heard: AI model providers might avoid killing off AI wrappers because wrapper startups represent lucrative revenue margins. Replacing them directly risks driving those customers to competing model providers, which hurts their bottom line unless the provider's first-party app can generate significantly higher monetization. At the end of the day, I'm not sure this ecosystem excites me. The work ground-level founders need to do feels like grunt work. The AI frontier sounds sexy from the outside, but the day-to-day execution on the ground is anything but.

In short, you need to venture into domains that are harder for pure AI software to replicate. For me, that means physical AI. It's early, and the ground rules haven't been set yet. That's where I'm putting my next basket of eggs. There are so many real-world applications where physical AI can make an impact, and we're barely scratching the surface. LLMs alone won't trivially lower the barrier to entry there.