PastAGI

PastAGI is an AI engineering publication for people who build with LLMs and want the math, not the marketing. Every piece starts from something verifiable: a serving cost we actually computed, a benchmark we actually reproduced, a claim we tried to break. Recent work includes an audit of GPT-6 Astra pricing at six dollars an hour, where we rebuilt the token math and showed exactly which usage profile makes that number real and which workloads it falls apart for. We looked at why chain-of-thought summaries are quietly replacing reasoning transparency and what that costs you when you debug an agent pipeline. We mapped the ways AI agents leak sensitive data, from prompt injection to over-broad tool permissions, and tested the controls that actually stop the bleeding rather than the ones that just look good in a slide. We ran the serving math for a 2.9B-parameter character-level TTS model, because someone has to publish the throughput numbers instead of the adjectives. And we built prompt dependency graphs that shrink your retest set when a system prompt changes, the kind of pattern you can apply the same afternoon you read it. The site is free, with no paywall and no email wall, and every article shows its assumptions so you can argue with the numbers instead of the conclusions. The audience is engineers, ML teams and technical founders who ship AI features and are tired of vendor benchmarks that never survive contact with real traffic. What makes PastAGI different is the discipline: we would rather publish one honest cost breakdown than ten hot takes, and we update pieces when the numbers move. If you want practical AI engineering grounded in audits, serving math and honest tool reviews, that is the entire point of the site. That discipline is what we publish every week.

What problem it solves

AI teams burn hours on vendor claims and benchmarks that never replicate. PastAGI replaces hype with auditable math and honest tradeoffs.