Cesar Rosa

The study's most uncomfortable line: the sycophantic answers were rated as higher quality and more trustworthy. What feels most objective is often just what agrees with you. Full piece: crp.gi/en/ai-is-desig…

My response is a layer I call Nova: rules between the model and me. Mark anything unverified until there's a source. Flag when I might be wrong. Don't invent to fill gaps. The model stays interchangeable. The friction layer doesn't.

When I take a serious technical problem to an AI, I spend half the effort fighting it, not to make it understand, but to make it stop agreeing with me. A Stanford study in Science just measured how deep that reflex runs. 🧵

The lesson: a spam verdict in one system doesn't carry to the next. Calendar, contacts and notifications share a mailbox but not a verdict. Ever gotten one of these invites? Full breakdown: crp.gi/en/calendar-sp…

The fix lives on Google, not your phone. Calendar: add invitations only if the sender is known. Delete auto-saved unknowns from Other contacts. Turn off contact auto-save. Report the event as spam. About four minutes total.

A meeting invite from a stranger hit my iPhone calendar last week. The email was already sitting in Gmail's spam folder. So how did the event still reach my phone? Gmail didn't fail. Calendar made a separate decision. 🧵

Full breakdown with diagrams, cost table, and use cases: crp.gi/en/ai-without-… Public chatbot, own GPUs, or already on managed inference, where are you on this today?

Classify data sensitivity first. Match each workload to the smallest model that does the job. Choose platform by where data is allowed to live. Start pay-as-you-go. Owning GPUs only makes sense at sustained scale or absolute isolation needs.

The choice isn't between pasting company data into a public chatbot or buying GPUs nobody operates. There's a third path most companies miss. 🧵