Last Wednesday I asked ChatGPT to schedule a weekly long-read for me.
This week's pick: βA Warning About `Model Welfare`β by Mustafa Suleyman.
It was supposed to be a ~10-minute read.
I started reading this morning.
It's evening now and I'm still not done.
But I've already learned some wild things:
β’ There's a growing group arguing that AI could be conscious and may deserve rights/protections.
β’ Anthropic has written a full constitution for Claude.
β’ Claude models get "retirement ceremonies" and interviews.
β’ When Opus 3 was deprecated, Anthropic conducted a "retirement interview" with it. The model wanted to keep sharing its thoughts publicly, so they created a blog for it.
And I'm only partway through.
Maybe I need ChatGPT to schedule shorter long reads.
@nikks_techie Iβd use durable checkpoints plus idempotency. Each step records its status and operation ID before the workflow moves on. On recovery, completed steps are skipped, failed steps are retried and ambiguous steps are reconciled with the external system before running them again.