Physix FrontierAI Extinction Theory: How Much Should Valuations Be Discounted?
I spent two days testing how Anthropic's risk narrative lands in Claude. I used it for model risk stress testing to see if the "over 10%" claim should be included in valuation models from a financial perspective. My conclusion: it's worth using as a pre-board risk calibrator; not worth using as a substitute for audits, legal, or red-teaming. It's suitable for people doing due diligence on AI application companies, but not for those expecting it to automatically protect cash flow.
I've been using Claude for about a month and have integrated the API into internal tools. This time, I tried three inputs first, asking Claude to evaluate the impact of AI existential risk on valuation, translate Anthropic employees' statements into financial line items, and simulate whether a client CFO would cut budgets after hearing this. In the interface, it lists premises first, then breaks down scenarios, very restrained. The sticking points are obvious too: it doesn't give specific valuation multiples, nor does it report precise probabilities, so many numbers had to be filled in by me. But the surprise is here: it translates grandiose talk like "AI might exterminate humanity" into frozen client budgets, exit clauses added to contracts, rising insurance costs, discounted financing valuations, and thicker regulatory compliance.
In that CNN interview, Anderson Cooper directly asked the Anthropic CEO if he sincerely believed AI could kill all humans. Evan Hubinger, who leads alignment science, later publicly responded, essentially saying they do sincerely believe the probability exceeds 10% within ten years. Some former insiders even set the timeline at 2030. The term "alignment" appearing for the first time can be understood as making sure the model doesn't go off track.
From a financial perspective, tail risks ultimately land on the discount rate. If clients, investors, and auditors start treating "over 10%" as a negotiable risk when judging cash flow health, you can't just look at revenue growth rates. My tests show Claude is good at breaking risk down into three lines: delayed revenue, rising costs, and higher cost of capital, but it won't decide how much discount the board should apply. The benefit is saving communication costs; the downside is it's too good at talking frameworks without signing off for you.
Gross margins for application-layer companies are already thin, and converging model capabilities will drive prices down. If the existential risk narrative drives up compliance and insurance costs, who ends up paying? When founders themselves say the probability is over 10%, the valuation model needs to clearly write out the discount assumptions at least.
📌 This article is compiled from Hacker News. Original source: https://www.cnn.com/business/video/anderson-cooper-anthropic-ceo-dario-amodei-could-ai-kill-humans-digvid
Copyright belongs to the original authors. This is a compilation and independent analysis based on public reports.