Thoughts on AI-augmented development, open-source models, and the economics of shared infrastructure.
Coding squads were the start. OCR pipelines, embeddings and chatbots have the same shape. Here is how inference instances work, when paying for GPU time beats paying per token, and the loads where the API wins.
SWE-bench Pro is a strong capability signal. Real squad work needs something else: multi-rep agent reliability, held-out scenario checks, and honest quant comparisons. That is what SQE measures.
SWE-bench Pro matters, but the leaderboard race does not. For human-steered software engineering, open-weight models at 58+ already sit in the Frontier Coding tier. The hard part left is ownership, not raw capability.
A single power developer burns ~1.5B tokens per month. At frontier API rates, that's $4,000–$6,600/month. We break down why token pricing is fundamentally misaligned with how modern inference infrastructure actually works.
GLM-5.2, MiniMax M3, and Kimi K2.7 aren't just 'good enough', they're closing in on Claude Opus on real coding tasks. The models are ready. The infrastructure to run them isn't. Until now.
Every serious developer fantasizes about running their own LLM server. We did too. Here's what we learned about the real costs, the hidden complexity, and the middle path that actually works.