PastAGI

PastAGI is an AI engineering publication for builders who want the math, not the hype. Recent pieces include a full audit of GPT-6 Astra pricing at $6 an hour with the serving math shown line by line, an analysis of custom AI inference chips eating into the GPU market, a review of TontaubeV1 - a 2.9B-parameter char-level TTS model - with real serving numbers, and a guide to prompt dependency graphs that shrink your retest set. Security coverage explains why AI agents leak sensitive data and concrete ways to stop it, plus what happens to chain-of-thought transparency when reasoning traces get summarized away. Sections cover guides, tools, engineering deep dives, use cases, and news. Everything is written to be auditable: prices are benchmarked, benchmark claims are stress-tested, and tool reviews ship with the receipts. If you build with LLMs and are tired of launch-week blog posts, this is the practical layer underneath - serving math, eval methodology, and honest tradeoffs.