Taste in the Loop
A Note from the Editors
'Humans over clankers' is one of the values written into our manifesto. This week features a few articles that cover more human aspects like AI-induced psychosis and burnout. As we adapt to becoming managers of agents, it's important to take care of yourself, especially when the machine-priests of YouTube are constantly trying to guilt trip you for not running a team of 300 agents 24/7.
— Josh, Donnie, and Ben
The last six months in LLMs in five minutes ↗
— Simon Willison
LLMs have been improving at an incredible rate and it’s hard to keep up. If you’re feeling behind, checkout Simon’s overview of all the big changes that have occurred over the past six months. It feels absurd to say, but 6 months ago is “ancient history” in AI-land.
AI Psychosis ↗
— Mitchell Hashimoto
In a world that’s increasingly “AI-pilled”, Mitchell (of Hashicorp and Ghostty fame) offers a refreshing voice in a lenghty, but oh-so-spicy tweet:
I strongly believe there are entire companies right now under heavy AI psychosis and its impossible to have rational conversations about it with them…We already learned this lesson once in infrastructure: you can automate yourself into a very resilient catastrophe machine.
AI-assisted engineers are burning out, is this fine? ↗
— Irina Nazarova (Evil Martians)
Irina explores the impact heavy AI usage is having on our mental health and what to do about it in this article:
We’re more productive than ever. But there’s a dark side. AI-assisted code generation isn’t free; there’s a hidden cost that we as an industry are only beginning to realize: AI burnout. Are we dangerously ignorant to this problem? And how can we cope with it?
Don't Build Slop (4 Levels of AI Agent Maturity) ↗
— Ara Khan (Cline)
Ara gives a nice tech-talk at the AI Engineer Europe conference covering 4 levels of maturity when building AI agents. This talk carries a heavy emphasis on getting your architecture right, which requires a human in the loop.
Demystifying evals for AI agents ↗
— Anthropic
Outside of large AI labs and companies whose core business is building AI agents, most teams are heavily underinvested in evals. If you come from more of an engineering background rather than a ML/AI background, Anthropic’s blog post on demystifying evals is a great primer for teams who aren’t quite sure where to start.
Harness Curious?
If you built a TUI coding agent that only needs to render when a) the user types something, b) the LLM outputs text or c) an animated component needs to ‘tick’ — would you architect it as a “game engine” in TypeScript + React that renders scene graphs to the terminal?
Of course not. That’s exactly the type of over-engineered nonsense architecture a clanker would suggest. You aren’t a clanker. So you’d write that agent in Rust and use a simple event loop, like a reasonable human. Claude Code listened to the clankers (hence why it flickers like a drunk hummingbird who slammed too many Bee’s Knees on a Taco Tuesday).
Clanker Fail of the Week
Join our community Slack and send us screenshots of your favorite clanker fails.