TIL: August 19, 2025 - Weekly Reading: AI and more AI

Artificial Intelligence

• Vibe code is legacy code – Steve Krouse

  This is an interesting take on vibe code. The idea is that vibe code that was written entirely by an LLM is the same as legacy code because neither are code that is well understood. Essentially, vibe code became legacy code the instant it’s written.

  I mostly agree with the idea, but I’d add that legacy code is a bit more nuanced in that it is code where maintaining it is not worth the cost. Even then I agree with his point because vibe code is basically not worth the cost of maintaining it the moment it’s created.

• Innovation, not Productivity: Why we Built Windsurf – Windsurf

  I’ve been thinking about AI IDEs a bit recently and this is Windsurf’s reasoning for building a full AI IDE rather than just an IDE extension. They mention that more powerful models are available now, but their reasoning seemed flimsy. The rest of the post basically says the reason we need AI IDEs is for the user experience. User experience is great and all but I’m not sure we need a full IDE fork just so that the UI can be snappier.

  I plan on writing a bit more about my thoughts on AI IDEs pretty soon.

• Why LLMs Can’t Really Build Software – Conrad Irwin

  Conrad writes that the reason LLMs can’t really build software is that they lack the ability to maintain mental models. I’ve long thought that this is the reason they can’t do a lot of reasoning tasks. They are just basically predicting the next word over and over again. They have a model but it’s built around the next word prediction task, not any kind of conceptual or spatial model.

• AIにはAI原則が必要 – Daisuke Maki

  Daisuke Maki writes about the need for AI principles. He observes that the AI will shirk responsibilities like avoiding writing complex code and just write stub code that passes some very basic stub tests.

  He argues we should have AI principles that are similar to the Isaac Asimov’s Three Laws of Robotics to guide AI do the right thing. He provides some links to OpenAI research and Anthropic research on this topic.

• Your MCP Doesn’t Need 30 Tools: It Needs Code – Armin Ronacher

  Armin writes about how connecting more MCP tools to AI coding agents often results in more confusion and not less. He notes that AI models aren’t good at using tools that can run in complex environments, like on different CPU architectures or operating systems.

  He narrows in on a solution for debugging with an AI agent using pexpect which is an older system with lots of existing code that the model has been trained on. This creates better results but means you need to have an MCP server that executes arbitrary pexpect code.

  Armin also notes that the security implications of this are pretty terrible. I hope to write a bit more about this soon, but I think that the only way to have anything sane is to sandbox the AI agent and all MCP servers. It’s a bit of a tall order to manage the sandboxing, on top of managing least privilege access for the MCP servers.