Thoughts on AI, development, and navigating emerging tech
Anthropic's new legal plugin for Claude isn't just a product for lawyers. It's a concrete example of something much broader: the ability to describe complex workflows in plain language, connect them to real systems, and let an agent handle the coordination while humans handle the judgment calls. Law, accounting, HR, compliance, any white-collar function that runs on documents and defined processes is a candidate.
Open-source AI tooling is outpacing proprietary alternatives in browser automation, voice processing, and inference optimization. A look at what's driving the shift (model commoditization, hardware accessibility, and trust premiums) and what it means for teams making build-vs-buy decisions.
Two approaches to AI-powered browser automation are diverging. Vision-based tools like Operator use screenshots; text-based tools like Browser Use parse the DOM directly. The benchmarks point in a surprising direction, and the reasons why tell us something broader about practical AI tooling.
AI agents are moving beyond software sandboxes into phone calls, edge hardware, and on-device voice processing. Projects like ClawdTalk and PicoClaw represent early experiments in giving agents a physical footprint. Here's what that shift looks like and where the practical limits are.
A straightforward technique for making AI agents improve over time: add a reflection step after each task, store the compressed lessons, and inject them into future context. One customer support deployment saw first-contact resolution improve from 62% to 79% over four weeks without changing the model.
Chinese AI labs, constrained by hardware access and budgets, have pursued efficiency and deployment speed over model scale. Labs like DeepSeek are producing competitive models at a fraction of the training cost, and the results are complicating the narrative that bigger is always better.