I build production AI systems for ecommerce. Multi-LLM ops, SP-API SaaS, MCP servers in Claude. Head of tech at a Shopify Plus brand. I post the builds.
Coinbase cut their AI bill roughly in half without giving up frontier models. Cheap models handle the routing and glue work. The expensive model only gets called when the reasoning actually matters. Most AI bills are paying frontier prices for grunt work.
Source: @a16z Show episode with Barbara Liskov, July 13. She co-authored PBFT. Worth the listen if you're building any multi-agent correctness gate. The 1990s distributed-systems literature already worked out what breaks and why.
Multi-agent review panels are rediscovering Byzantine fault tolerance. PBFT solved the lying-node problem in 1999 with 2f+1 majority certificates. Today's AI review gates are the same shape. The failure modes were mapped 30 years before the agents existed.
PwC analyzed over a billion job ads to answer the "will AI take jobs" question. The real finding is more useful than the headline: AI is splitting work into two tracks, and only one is winning.
Saw a small version in my own ops. A listing job that ate 1 to 2 hours now runs in under 5 minutes. The model types. Knowing what good looks like is the part that still pays. Expertise didn't get cheaper. It got concentrated.
@tferriss Government is always the most scary, but I went with one company owns it because at least with the government, we can hopefully vote them out. Unless, of course, it's China or some other dictatorship, I guess.
@thearslaniqbal One thousand percent it will, and it always was the case. Everything a programmer built was someone else's business knowledge turned into code. They didn't have it firsthand, so it took more rounds than if the person holding the knowledge could just do it themselves.
The most dangerous person with an AI coding agent isn't an engineer.
It's an operator who knows the business cold.
Anthropic studied 400k Claude Code sessions. The predictor of success wasn't coding skill. It was knowing the problem you're solving.