"We're really doing this," I say. "Hiring a team. Building a company."
Jian stops typing, looks at me. "We were always going to do this. Elad just accelerated the timeline."
"Think any of them will bite?"
He shrugs. "One or two. The rest will wait to see if we survive the next six months."
I close my laptop. The AC has stopped squeaking, but the room still feels charged. Like something has shifted and we can't go back.
"We should get dinner," I say. "Before everything changes."
Jian nods. We pack up in silence, two kids pretending to be executives, about to hire people who actually know what they're doing.
The hammers of Abraham. Armed and ready to defend.
God help us if we're not actually up to it.
Subject: The Problem No One's Solving
Hey Jordan,
We met at the ML for Systems workshop last year - you were the one asking all the uncomfortable questions about safety evaluation at scale.
I've been working on something since then. We're called Inferthermic, and we're doing real-time safety classification for LLM outputs. Not keyword matching. Not human review. Actual mathematical alignment through entropy reduction.
The uncomfortable truth: the current generation of language models can be jailbroken into generating genuinely dangerous content. The big labs know it. They're not talking about it publicly. We have a solution that reduces risky interactions by 98%, empirically proven.