In a small experiment, Doug found that (2) performed best at a nuanced spelling correction task: (1) no rules or examples (zero-shot) (2) clear rules without examples <-- 🏆 (3) rules with examples (4) examples without explanations (few-shot) (5) examples with explanations

AI assistance can massively improve the quality and security of the software we develop. It just depends how we use it. (And, over time, the tools are going to build more good practices in.)

Interesting reflections on tech stack choices from someone who's built a platform for building AI agents.

Agents building agents seems like a particularly interesting use case.

I've been testing Replit ⠕'s new Agent 3 to build a couple of agents. It's interesting to see the stack it's using. So far, @mastra_ai TypeScript agent framework for both. And traces go to Inngest.

Some great advice here for anyone trying to improve an AI agent.

AI Engineer Europe coming to London next April...

My article about how well LLMs understand different table formats is on the front page of Hacker News. news.ycombinator.com/item?id=454584…

Which Nested Data Format Do LLMs Understand Best? JSON vs. YAML vs. XML vs. Markdown improvingagents.com/blog/best-nest…

100% agree - if you want a taste of what's coming (and already here) with AI agents, try giving Claude Code some non-coding tasks.

Different perspectives...

This looks really interesting. I have a feeling we're going to be seeing more models optimised for agentic search, and not just for coding.

Interesting. As we shift to using AI agents more and more, it seems possible that whispering to our computers will become commonplace in offices everywhere.

Considering we humans have had relatively few serious options for searching the web for the last 20 years or so, AI agents are comparatively spoilt for choice. Parallel (Parallel Web Systems) looks like an interesting new option. I've added it to my comparison page here: mattcollins.net/web-search-api…