Sleeves-up AI

MCP All the Things

Like many, I initially dismissed MCP as merely "Anthropic-flavored function calling."

But using the Claude desktop app with MCP servers enabled has been a kind of revelation. In practice, it is still just function calling in most cases. But:

• Standardization is important for adoption.
• MCP's client / server model adds lots of interesting potential.

Both client and server can be full applications — complete with authentication, persistent state, and complex logic. It turns LLMChatMessage into a much more dynamic and interesting building block.

The possibilities are intriguing:

• Private tools and contexts, portable between different LLM providers?
• Distributed inference between local and cloud LLMs?
• Login with Claude / ChatGPT / OpenRouter?
• Web content that is gated only for AI consumption?
• 1-off, plain English Zapier workflows?
• etc.

AI Coding Interns

So far, my most successful AI-coded experiments all share a few similarities:

• Standalone, self-contained use case. No external integrations.
• Ephemeral state. No auth or database.
• Simple static or local deployment.

It's not that AI can't whip up a database... it's just that its habit of making compounding, spaghetti code errors means I can't trust it. And when I do review its code, I end up wishing I had done the foundations myself.

Using existing templates and libraries feels like a temporary but productive crutch for now. Using AI, even for greenfield projects, seems to benefit from strong conventions, strongly enforced.

How would we structure software differently to make it as easy as possible for AI?

Armchair Commentary

I cut the following section from a recent newsletter, but thought it was a sentiment snapshot worth preserving.

In one corner:

AI is too dangerous to release, but completely safe to deploy! It's transformative enough to justify investing billions, but definitely won't threaten anyone's job. Even if it does, we'll be so wealthy that we can simply share with everyone. Everyone except China — who must be stopped at all costs!

In the other:

AI is an existential threat! We need immediate global cooperation — from the same governments that can't agree on... anything. We must solve AI alignment — even if we currently don't know how to align small children or pets. Please sign our petition to ban research until we've solved philosophy!

In the messy middle:

AI is useless tech hype — can we please stop talking about it? ChatGPT has been out for over 2 years, yet all it managed to do is undermine homework, intellectual property, and internet public spaces. It confidently fabricates facts while failing at basic tasks. I don't own any $NVDA, and I'm both proud and angry about it!

Links and Listening

AI "biology" feels like a big deal.

Fun to think about: How to Create a New Town?

I recently heard the word "autarky" used to describe current global trends. It captures the zeitgeist perfectly.