Keeping AI on a Leash

My Takeaways From Andrej’s Talk

I watched Andrej‘s whole talk so you don’t have to. The key point that resonated with me was @ 22:55: 1) “We have to keep AI on a leash” to increase the probability of successful verification. 2) Make the verification of what AI produces easy and fast.

There’s hype around agentic AI (which I posted about a few months ago) but it doesn’t even matter if you’re building a crazy multi-agentic system if YOU are the bottleneck for evaluation. If you’re trying to get things done, letting an overreacting agent loose on your project is not productive and you need to keep it on a leash. He uses Atharva‘s https://blog.nilenso.com/blog/2025/05/29/ai-assisted-coding/ as an example of how to do that, through the practice of metaprompting. When you keep AI on a leash, then auditing its output for accuracy and consistency is faster.

Later on @ 31:37 he shows off an app that he vibe-coded that allows you to take a photo of a menu (with mainly text) where it generates images of the menu items. Brilliant! menugen.app

The last chapter has practical tips on making docs and websites legible to LLMs and miscellaneous tools for that.

If you did watch the whole video, I’m curious what you found interesting about it. Here’s the link:

This post was originally on LinkedIn.