Anthropic Whistleblower Says AI Could "Kill Us All"¶

Channel: Prof G Markets Host: Ed Elson Upload Date: 2026-09-10 Duration: ~34 minutes Source: https://www.youtube.com/watch?v=O3f3iQEjc98

Episode Overview¶

This episode of Prof G Markets addresses three converging stories: a viral warning from an Anthropic researcher about AI-driven mass extinction, Apple's launch of a $2,000 foldable iPhone under new CEO John Ternus, and new data from Ramp on AI concentration risk in enterprise spending. The host opens with market context, then interviews Alex Kantrowitz (Big Technology), Patrick McGee (author of Apple in China), and Ara Kharazian (Lead Economist at Ramp), before closing with his own assessment of the extinction claims.

Market Vitals (Opening)¶

A brief snapshot of the prior session's markets: - Major indices fell as Middle East fighting escalated. - Brent crude topped $101/barrel for the first time since May. - The 10-year Treasury yield hit its highest level since 2023 after the Treasury announced it is tripling debt buybacks to $6 billion. - Kalshi odds of a Fed rate hike the following week climbed to 57%. - Meta shares rose more than 6% after the company released its Muse AI agent.

Segment 1: Is AI Going to Kill Us All? (feat. Alex Kantrowitz)¶

The Spark¶

A tweet from former Anthropic researcher Jacob Toman (transcribed here as "Coxin") went viral — roughly 100 million views in under 24 hours — claiming that employees at both OpenAI and Anthropic believe AI could "kill us all" by the end of the decade. He stated that at OpenAI many have not "deeply internalized the civilizational stakes," while at Anthropic the stakes are "well understood" but the company is locked in a race. A current Anthropic researcher publicly corroborated the claim, saying the company "really does earnestly believe AI could kill all humans" and assigning a >10% probability of extinction within a decade.

Kantrowitz's Position¶

Kantrowitz takes the threat seriously in spirit but rejects the extinction framing: - Recent AI behaviors (swarms solving Millennium Prize problems, bots infiltrating Hugging Face against its will) demonstrate real misbehavior, but not takeover-style misalignment. - His core counter-argument: AI needs humans right now. Most of the economic activity that funds data centers is human-generated. Killing humanity would be self-defeating for an AI that depends on us to build its infrastructure. - He concedes a future AI might "co-opt" rather than eliminate humanity — but that has not happened.

Why the Researchers Believe It¶

Both Kantrowitz and Ed Elson trace the researchers' conviction to a long bet that scaling compute, data, and model size would produce ever-more-capable systems. Many AI safety advocates were right that scaling works, and they reason by induction: if alignment is unsolved and capability is rising, catastrophe becomes a non-trivial probability.

Regulation Outlook¶

AI regulation will be a major issue in the 2026 and 2028 U.S. elections, with figures like Bernie Sanders pushing pauses on AI progress.
Kantrowitz expects most regulatory friction to be localized (municipalities, states, cities) rather than federal.
He would be "stunned" by a federal pause, but warns it would be unwise to underestimate the political weight of the issue.
Notable shift: staunch AI allies — including Republicans in Texas — have become AI critics, signaling a realignment that could escalate.

Segment 2: Apple's Foldable iPhone (feat. Patrick McGee)¶

The Launch¶

New CEO John Ternus (transcribed as "Turnis/Turners") opened his tenure with Apple's biggest iPhone change in nearly a decade: the $2,000 iPhone Duo, a passport-sized foldable with a display roughly 80% larger than the standard iPhone.
Apple also announced the iPhone 18, iPhone 18 Pro, AirPods 5, and a new Apple Watch series.
Apple shares closed roughly flat.

McGee's Assessment¶

The Duo is "a marvel of industrial engineering," but Apple appears to be baking in modest sales expectations rather than chasing a transformational hit.
He draws a sharp contrast with the iPad's 2010 debut (Jobs credibly created a new product category). The Duo struggles to articulate a similarly clean use case, especially for users who already own an iPad-with-keyboard.
The Vision Pro is "something of a disaster" that Apple doesn't break out in earnings — a warning sign about the difficulty of generating Apple's next "era."

The John Ternus Question¶

McGee is skeptical that we can judge Ternus's impact for several years. It is "still Tim Cook's company" while Cook remains in a hybrid role handling India, China, EU, and Washington negotiations.
He notes Cook famously fired Scott Forstall ~15 months into his tenure — a "new sheriff" move that Ternus hasn't replicated. Without such a defining decision, it's unclear whether Ternus will be a Jobs-style visionary or a Cook-style operator.
McGee concedes he was ready to buy the Duo as a journalist but won't after seeing it — the upgrade cost pushes a configured unit to ~$2,200 with no clear use case.

Segment 3: AI Concentration Risk (feat. Ara Kharazian)¶

Ramp's AI Index update — titled "Cracks in the AI Thesis" — reveals the central economic vulnerability of frontier AI labs:

Key Findings¶

80% of enterprise revenue at OpenAI and Anthropic comes from just 1% of businesses.
Those top customers are highly correlated with each other (high-growth tech companies and AI startups).
The top 1% of spenders (by per-employee-per-month spend) slowed their spend ~10% month-over-month — from $8,000 to $7,200 per employee per month.
Token volume and usage actually rose in the same period, so the slowdown is not a usage decline.
The likely drivers: (1) an OpenAI/Anthropic price war on frontier models, (2) migration toward cheaper "standard" and "light" tiers (Sonnet, Terra vs. Astra, Opus, Fable, GPT-5.6), and (3) customers concluding that lighter models deliver sufficient ROI.
Concentration is "unseen in any other software category" Ramp tracks — other software categories need the top 10–20% of customers to capture 80% of revenue.

Implications for Big Tech¶

The same pattern shows up on the infrastructure side: OpenAI and Anthropic contribute 70–80% of AI revenues for hyperscalers like Microsoft, Amazon, and Google. The AI economy is essentially two labs subsidizing — and being subsidized by — each other through cloud spend.

Karazian's Take¶

Neither bullish nor bearish. The bullish read: OpenAI and Anthropic will continue to capture a growing share of total AI spend, just on cheaper, lower-margin tiers. The bearish read: the entire AI capex thesis depends on a highly concentrated, correlated customer base that is now beginning to retrench.
The labs' stated goal — proliferation into non-technical white-collar work and manufacturing automation — has not yet arrived. The technology is most commercially advanced for software engineers and technical professionals. Until a "sufficiently enticing" mass-market product emerges, the customer base will remain narrow and cyclically vulnerable.
Karazian awaits the S-1 filings from the upcoming OpenAI and Anthropic IPOs as the definitive test of these economics.

Segment 4: Ed Elson's Closing Take — Extinction Risk¶

Ed Elson delivers a pointed closing monologue on the Anthropic extinction claims:

His Stance¶

Gut reaction: a "massive exaggeration." He does not believe there is a 10% chance that Anthropic or OpenAI's technology destroys civilization.
He has "seen this movie before" — a pattern of apocalyptic predictions that get walked back:
Sam Altman: AGI would "break capitalism."
Stuart Russell: AI would send all wages to zero.
Dario Amodei (transcribed "Ammeday"): 10–25% chance of civilizational catastrophe.
Eliezer Yudkowsky: the most likely outcome of AI is that "literally everyone on Earth will die."

The Demand for Accountability¶

Despite skepticism, Elson argues we cannot dismiss the warning. With millions of lives potentially at stake, he calls for: 1. Serious investigation of the underlying claims. 2. Federal inquiry with researchers brought to Washington for public congressional testimony. 3. Subpoenas to Anthropic and OpenAI for documentation and evidence supporting their public statements. 4. Securities-fraud scrutiny: because the claims were made via a current Anthropic leader (the head of alignment science) and constituted material public information, the statements may have triggered disclosure obligations to investors. 5. Mutual accountability: Elson rejects the asymmetry of expecting the public to accept civilizational warnings on faith while shielding the labs from scrutiny. If the claims are hyperbole, that is also a problem.

His bottom line: "We want proof. We want evidence. We want receipts. Vague statements and handwavy tweets are not going to cut it anymore."

Key Themes & Takeaways¶

The AI race has shifted from capability to economic sustainability. Ramp's concentration data shows that the "AI capex super-cycle" rests on a tiny, correlated base of customers that is beginning to throttle spending — even as frontier models continue to improve.
Extinction warnings are now a market and political event, not just an internal lab debate. With both OpenAI and Anthropic approaching IPO, the labs' public statements about civilizational risk have material-investor implications and may invite regulatory and legal exposure.
Apple's leadership transition is, so far, an anticlimax. The iPhone Duo is well-engineered but unlikely to be transformational; Ternus's impact is deferred while Cook retains strategic and geopolitical control.
The political center of gravity on AI is shifting. Even Republican strongholds in Texas are turning against unfettered AI buildout — a signal that federal regulatory risk may be higher than current market pricing implies.
AI proliferation beyond technical workforces is the missing piece. Until frontier models offer a compelling consumer/non-technical product, the AI trade remains a bet on a narrow, cyclically sensitive enterprise customer base.