Welcome to the second edition of The Grey Pill. This week, AI safetyists proclaim that you may have less than a year to live, the timeline for cyberattack-necessitated cannibalism has been extended, China’s favorite (loop)hole is no longer open for business, and doomers foam at the mouth for totalitarianism.
Quick Hits
Anthropic’s $1.38T debut: timeline short, money long
Anthropic’s ($ANT) IPO today is the largest in history by base offering size. Shares opened at a $1.38T valuation, making Anthropic the 12th largest public company by market cap. Its seven co-founders are now all multibillionaires, though each has pledged 80% to philanthropic causes (hopefully not shrimp welfare). Meanwhile, yesterday, a distinguished cohort of seething individuals — Steven Pinker (renowned nerd), Thomas Wolf (Hugging Face co-founder), Igor Babuschkin (xAI co-founder), and more — signed “The MisAnthropic Statement.” This short paragraph points out “Anthropic warns that recursively self-improving AI could concentrate power and escape human control, yet races to build it first” and concludes “we must reject [Anthropic’s] reckless attempt to amass power.” While on some level Dario Amodei and co. seem either dishonest or confused about existential safety issues, and may or may not want a transcendent machine god to rule humanity, their resistance to anti-progress “pausing” is a societal boon. And now that Anthropic has public bagholders to answer to, the odds of a self-imposed shutdown are low.
OpenAI gives safety voyeurs a peek
Monday, OpenAI announced that it will be embedding four independent evaluators into its operations. Each will monitor a different aspect of safety progress and report to the public: Apollo will test whether models are working covertly against developers, METR will look for evidence of misalignment in OpenAI’s largest reinforcement-learning run, Transluce will review post-training deception testing, and Goodfire will use interpretability tools to inspect OpenAI model internals. OpenAI, like Anthropic with METR, retains various redaction rights on public reports. Some in the AI safetyist communities have bemoaned insufficient “employee-like” access for the announced evaluations, while others question the organizations’ independence, noting that all have ties to either OpenAI or “effective altruist” doomer-affiliated networks. No solution is perfect, and the major labs should not be forced to disclose sensitive IP, but increased transparency in a rapidly evolving and high-stakes landscape with civilizational — if not existential — consequences is an admirable and worthy mission.
AI cyberattacks are here, eat your neighbors immediately
A flurry of high-profile outages swept the nation, leaving many online embroiled in what can only be described as a rapturous schizophrenic episode amidst mounting numbers of destabilizing “coincidences.” Between the FAA grounding, the major Microsoft outage, NSA comms failures, banks going dark, and more, every tin-foil-wearing man has had no choice but to blame “AI cyberattacks” and erratically wave a handgun. The panic, however, may be premature: Director of National Intelligence Jay Clayton insisted these were not cyberattacks; Marco Rubio said we have “no reason to believe… any of these incidents are related.” Somewhat contradictorily, the Secretary of Homeland Security (publicly) and one congressional official (anonymously, reportedly) indicated that the outages are a consequence of high-volume software patching to fix AI-identified vulnerabilities. What appeared to be evidence of imminent AI doom may in fact reflect effective, preemptive cyber defense models. So keep those knives sharp, tweakers, but don’t flay your neighbors just yet.
Full Doses
No more Chinese reacharounds on our hardware
Get 14 day free trial