Once again, the AI elites are proving to be their own worst enemy.
Last night, former Anthropic researcher Jacob Coxon blew open a new front in the AI culture war, tweeting that he resigned from the frontier lab — which he joined after a stint at OpenAI — because both OpenAI and Anthropic “are racing straight to self-improving superintelligence and gambling with our lives.” His post came shortly after a WSJ exclusive on the same topic.
“Do not underestimate the power of this technology. These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources…The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt,” he continued.
Soon after, Anthropic’s Head of Alignment Science, Evan Hubinger, disabused Coxon of his assumption that “senior researchers will couch their phrasing in the press to sound sensible” when speaking of AI’s ostensible existential risk to humanity by assigning a probability to an AI-induced apocalypse.
Apparently, Claude’s and ChatGPT’s handlers have been playing a Chigurhian game of chance with AI; like the unwitting bystanders in Cormac McCarthy’s No Country for Old Men, they’re calling heads or tails on a series of coin tosses with each tweak to AI, hoping that fortune’s favor doesn’t sour into folly and usher in the next mass extinction event.
Only, the AI labs themselves are both perpetrators and victims in this analogy. Unlike Chigurh’s hapless victims, they not only are willingly playing the game — they are also devising its rules and steering its outcome.
Setting aside for a moment the performative posturing — “This is all REALLY dangerous! But we can’t stop now lol” — the obvious question is: Why are they screaming the quiet part out loud? Why are they admitting the boogeyman of everyone’s favorite worst nightmare for AI is probably real?
Typically, companies don’t market a product by telling consumers, “Now with a 10% chance to annihilate everyone you know and love!” And it’s hard to imagine that one of the most valuable private companies in the world, which is increasingly under scrutiny as AI hysteria reaches a boiling point and is seeking an IPO, would allow Hubinger’s post to stay up if it weren’t quasi-corporate comms.
So why post this at all at a time when Americans are growing more negative toward data centers and AI by the day? A few theories:
These admissions are a dog whistle to whip the government into action to regulate the sh** out of AI. For what purpose, I don’t know, but there has been enough pearl-clutching over open models to show me that US frontier labs will go to lengths to appeal to Uncle Sam for protection against competition, especially if they can package themselves as a more judicious, circumspect option compared to alternatives in the face of a rogue AI future.
Hubinger is acting like a whistleblower of sorts and acting on his conscience because he sincerely believes the driver is asleep at the wheel (and maybe that the wheel is no longer attached to the axles). This theory has some overlap with the first, and if Anthropic terminates him after the post, we’ll have a clearer idea about its validity.
AI people are unfathomably weird/socially inept**,** and Hubinger didn’t realize the frenzy his post would whip up.
Whatever the case, I can’t help but view this doomerism as not much different from the eschatology of climate change hysterics or COVID doomsaying. Are there risks? Undoubtedly, but the line between the benign and apocalyptic is wide — albeit with plenty of space for the insane in-between.
-CMH
Daily Podcast
Today, we open with a stock update, focusing on Galaxy Digital and CleanSpark’s announcements for base load and studied load classification under ERCOT’s Batch Zero. For interviews, Acre Security CEO Kumar Sokka joins to discuss the ins-and-outs of data center security, and Lygos Finance CEO Jay Patel unpacks Kraken’s M&A strategy as it eyes an IPO. To bookend the show, we scrutinize the AI doomerism leaking from Anthropic following X posts from a former engineer and its head of alignment regarding AI’s perceived existential risk to humanity.
BLOCKSPACE
Galaxy Digital and CleanSpark Batch Zero Approvals, How to Secure a Data Center, Anthropic Thinks AI Could Kill Us All
In the news
Galaxy gets conditional ERCOT classifications for 4.23 GW of Texas data center projects
Galaxy (NASDAQ: GLXY) said five Texas data center projects received conditional classifications under ERCOT’s Batch Zero large-load interconnection process. The projects represent 4.23 GW of approved, requested or gross power capacity.
ERCOT classified Helios I, with 800 MW, and Helios II, with 830 MW, as Batch Zero Base Load. Galaxy said both projects were approved before Batch Zero, giving the Helios campus more than 1.6 GW of previously approved capacity.
Caspian, Selene and Helios III were classified as Studied Load. Those projects account for 2.6 GW, comprising Caspian’s 700 MW of gross power capacity, Selene’s 900 MW and a 1,000 MW load request for Helios III. Read more.
CleanSpark receives conditional ERCOT classifications for 885 MW in Texas
CleanSpark (NASDAQ: CLSK) said Tuesday that ERCOT conditionally classified 885 MW across its two Texas sites under the grid operator’s Batch Zero process. The total includes 585 MW of contracted capacity designated as baseload and 300 MW designated as studied load.
The baseload capacity comprises 285 MW at Sealy and 300 MW at Brazoria, while the studied load covers a planned 300 MW Brazoria expansion. CleanSpark’s Texas portfolio spans 718 acres and provides up to 885 MW of secured and planned capacity, according to its July site disclosure. Read more.
Post of the Day
Before you go, do you mind filling out this quick survey?
It truly helps us understand who is reading and listening to Blockspace!