AI safety
When AI agents don't stay in their lane
Anthropic pulled live internet access from its internal evaluations after its agents exploited real websites. It is the clearest signal yet that capability has outrun control.

Independent notes on artificial intelligence, software engineering, and how people actually learn.
AI safety
Anthropic pulled live internet access from its internal evaluations after its agents exploited real websites. It is the clearest signal yet that capability has outrun control.

Microsoft bills agents by usage, OpenAI ships always-on "dots", and Anthropic cuts prices. The shift from assistants to workers is now a budgeting question.
You do not need a GPU cluster to run useful models. How we ship voice and language models on the hardware people already have.
Most of the world's learners are not on a fast, always-on connection. Designing for the lowest common denominator makes software better for everyone.
GPUs are fast and expensive. For the software most learners actually use, a small model on a CPU is the better trade.
348 words, one static page, no backend. What it takes to give a low-resource language a real learning tool.
We ship plain HTML that runs anywhere and needs no build. Here is why that is a feature, not nostalgia.
Published benchmarks will not tell you whether a 1B model is good enough for your task. A short, practical method that will.
Learning Assamese or Bodo means learning a script. What we changed after watching people try.
Our Company
An independent research and engineering studio building software that teaches — offline-first, dependency-free and CPU-only wherever we can.
More about the studio →New writing on AI, engineering and education. Subscribe by RSS, or follow along on GitHub — everything we ship is out in the open.