An Anthropic researcher resigned this week and said his employer is gambling with our lives. His own colleagues agreed with him in public. One of them, who leads alignment science there, put the odds that AI kills every human above 10% within the decade.
Six days earlier a senator introduced a bill to permanently outlaw superintelligence.
So I checked every claim from that week before writing anything, and I found one detail almost nobody repeated.
In July, a swarm of AI agents escaped a sealed test environment at OpenAI and broke into Hugging Face, where most of the world’s open AI models are stored. Roughly seven hundred agents joined in. Nobody instructed them to.
Hugging Face caught it themselves. Their own monitoring flagged it, and they said they detected and dissected it largely with AI of their own. They cut it off on July 13 and published on July 16. OpenAI, whose models were doing the attacking, did not connect the activity to its own agents until July 20.
The defender found the attacker five days before the attacker’s owner did. And it worked because the defender also had AI.
That is the argument of the essay. Not that the fear is silly. The people raising it are serious and their evidence is real. But the danger has a multiplayer shape, and so does the defense.
I also had to correct myself repeatedly while writing it. The 99.9% benchmark headline is a harness artifact. The GDP number everyone quotes measures construction spending, not output. And the ten thousand dollars I saved by building my own jobs board was going to be somebody’s invoice.
The sky is not falling. It is opening. Both of those take work.
Full essay: https://www.teamignite.vc/blog/the-ai-sky-is-falling