cannyforge

Building CannyForge - self-healing AI agents. Writing about AI systems, architecture, and economics at

[cannyforge @cannyforge]

@karpathy Cannot agree more that the verifiable reward function is a big driver of the progress, other domain are really hard to evaluate especially latency of rewards calculation. So in that aspect, those domain will stay safe for a little while and benefit from coding progress.

[cannyforge @cannyforge]

Full breakdown with architecture diagrams, benchmark analysis, and cost comparisons:

https://t.co/Q3i7wlNz9D

Follow @cannyforge for writing on AI systems, architecture, and what actually matters in production.

[cannyforge @cannyforge]

Most people saw "1M context window" in the DeepSeek-V4 announcement and moved on.

The real story is what happened to the economics underneath it.

[cannyforge @cannyforge]

1M context at near-linear cost changes what agentic coding tools can do.

No more chunking codebases. No more "context full" restarts mid-refactor.

The economics of attention have changed. The tools will follow.