Founder of Obvix Labs, an AI research lab building systems people can rely on. I write about what we're learning - AI safety, backend architecture, and what it takes to ship real software.
Everyone sees 'Zapier runs 800 agents' and thinks the secret is multi-agent orchestration. After reading every source — the full podcast, the Stanford paper, the Spotify engineering blog, the HN failure threads — I found something different. The companies winning at agents spent years building platforms before AI arrived. The model was the last piece.
You climbed on a stranger's bike yesterday. The system did not check his mind. A long look at the strangest blind spot in Indian transport, what the global evidence actually says about driver mental fitness screening, and why the obvious fix would make it worse.
Everyone found the Easter eggs. We mapped the engineering. 82 documents, 112K lines of original analysis, 16 architectural diagrams the complete architectural reconstruction of how Anthropic builds a production AI agent.
Does filling up a model's context window make it more likely to agree with you when you're wrong? I ran 80,433 trials across 6 models to find out. Context length matters less than you'd think - the conversational pattern is what really drives sycophancy.
A deep technical breakdown of Perplexity's pplx-embed-v1 and pplx-embed-context-v1 - diffusion-pretrained, INT8-native embedding models that fix RAG's oldest problem: context loss at chunk boundaries. Architecture, benchmarks, tradeoffs, and what changes for developers.