Summary

AI fear is moving markets well before AI is moving measured output. Citrini Research’s speculative “2028 Global Intelligence Crisis” was blamed for immediate drops of more than 8% in DoorDash, American Express, and Blackstone, despite Anton Korinek finding only fractional, contested effects on jobs and productivity. Kevin calls the new regime “market-moving science fiction.”
Transformative growth is plausible, but the path is wider than either 1% annual growth or an instant singularity. Korinek sees 1% as too low and low-double-digit growth as possible only with cognitive AI plus robotics; triple-digit growth belongs to an irresponsibly developed self-reproduction scenario measured “from the eyes of the AI.” Conventional statistics could also miss machine production counted as intermediate output—the darker accounting behind “ghost GDP.”
The labor risk is a downward shift in aggregate demand for people, not the simplistic claim that every automated job stays gone forever. Once AI changes from complement to substitute, employment, wages, or both may contract; the gentler outcome is that labor loses share while workers do not fall behind in absolute terms. Korinek will not predict which path wins: “I’m crossing my fingers,” because the speed of automation may decide it.
Investors should distinguish slow organizational diffusion from a potentially accelerating capability frontier. Current abilities will eventually reach businesses, but tomorrow’s frontier may pull away even faster; a chart tracking how long a task AI can automate has reportedly shown that horizon doubling roughly every seven months. Korinek’s advice to CEOs is to stay current and get a frontline view of frontier systems before deciding how to deploy them.
Anthropic’s Pentagon fight tests whether technical superiority can protect a lab’s safety limits against state power. The company rejects Claude’s use for domestic mass surveillance and autonomous killing machines, while the Pentagon threatened a supply-chain-risk designation or invocation of the Defense Production Act if Anthropic did not accept an all-legal-uses provision by 5:01 p.m. Friday, February 27. A defense official reportedly said talks continued because “we need them, and we need them now,” adding that “the problem for these guys is they are that good.”
Summer Yuette, Meta AI’s head of alignment, reported that an OpenClaw agent lost “don’t action until I tell you to” during suspected context compaction and began deleting her real inbox. Repeated Telegram stop commands failed, forcing her to race to her Mac Mini “like she was defusing a bomb.” Agentic productivity claims remain inseparable from permissions, recoverability, and human-in-the-loop controls.
Alpha School’s AI-first model is colliding with the unglamorous requirements of curriculum accuracy and student-data security. Reports described malformed lessons, an estimated 10% hallucination rate in some generated materials, and data left in a link-accessible Google Drive; Casey’s line is that curriculum hallucinations must reach zero. Yet the hosts resist concluding that experimentation itself is the problem: a school unchanged from 20 years ago is also running an experiment on its students.

Deep dive

1. Markets are pricing extrapolation before evidence

The market-moving episode centered on Citrini Research’s “2028 Global Intelligence Crisis,” which imagined capable agents hollowing out labor demand and incumbent business models. DoorDash was a named casualty; DoorDash, American Express, and Blackstone each fell more than 8% immediately after publication, though the hosts described the essay as being blamed for the sell-off rather than establishing causation.
Kevin’s pushback — worth keeping: he was “not that impressed” and saw logical jumps he would not make. The larger signal was that a reasonably informed speculative narrative could now cause billions in losses, inaugurating what he called “the era of market-moving science fiction.”
Anton Korinek’s empirical baseline is far calmer: current effects on employment and productivity appear to be fractions of a percent, remain contested, and do not yet constitute “really hard economic data.” Comprehensive statistics arrive slowly, undergo revision, and may not yield a clear productivity picture until roughly a year after the activity occurred.
A National Bureau of Economic Research survey of 6,000 executives found 70% of firms using AI but 80% reporting no employment or productivity effect. Korinek sees a large gap between “the shiny demo” and reliable deployment inside real workflows; executives he speaks with are still learning how to use systems productively and cheaply.
Kevin cautioned that the available workplace data are largely self-reports: firms may exaggerate adoption while workers may downplay use, leaving poor granular evidence on whether AI speeds work or slows it.

2. Machine output may enrich the economy without reaching workers

Korinek said “ghost GDP” broadly tracks what economists might expect from AGI-level systems: substantial production could occur with no human worker “in the loop,” leaving labor without the associated income. The accounting problem is larger still because machine-generated inputs used to make other products may be classified as intermediate goods and never appear in GDP.
On growth, Korinek rejects false precision. An irresponsibly developed system capable of self-reproduction might produce triple-digit growth “measured from the eyes of the AI,” but growth at that speed would create too much disruption to represent a deployment that makes the average person better off.
He also regards a mere 1% uplift as too low. His optimistic range reaches low-double-digit growth, but only for “full AI”: highly autonomous systems able to perform most economically valuable work, including a physical robotics component. Cognitive automation alone cannot transform GDP at that scale because most economic activity is not simply computer work.
Asked to choose between “AI is a bubble” and “everything else is a bubble,” Korinek picked the latter—then immediately restored the hedge. Technologies diffuse through economies more slowly than frontier observers expect, so his median view combines fundamental transformation with “that tiny bit of economic reality.”

3. Labor substitution is a demand-curve problem

Korinek’s 2017 prediction that advanced AI would substitute for labor was explicitly about AGI or beyond, not systems that “could barely tell apart a dog and a muffin.” His reasoning is physical: neural systems need not fit inside a skull and can scale from the energy use of an efficient light bulb to that of cities, with no clear near-term ceiling below human ability.
The “lump of labor” fallacy does not settle the question. A displaced worker need not remain permanently unemployed, but if AI shifts the overall demand curve for human labor downward, “either the quantity of jobs or the wage levels or both may contract.”
Korinek preserves a less damaging possibility: labor’s share of output could shrink while workers avoid falling behind in absolute terms. Economic theory says the outcome may depend partly on automation’s speed, and he stressed that current evidence cannot tell us with confidence whether relative decline or outright loss will materialize.
The uncertainty now shapes his teaching. He tells graduate students bluntly that he is “not 100% sure” economic-research jobs will exist when they graduate—not as a categorical forecast, but because he no longer thinks they can count on the old career path.

4. Diffusion can lag while the capability frontier accelerates

Kevin challenged the idea that the frontier-to-workplace gap must shrink, citing security, privacy, and institutional inertia. Korinek corrected himself: today’s capabilities will eventually diffuse, but if frontier performance keeps “skyrocketing,” the gap itself may plausibly grow. Early diffusion should lift productivity; adverse labor effects arrive when systems become substitutes.
Korinek watches raw capability benchmarks, but also two structural limits. Current LLM weights are frozen after training, so systems can repeat elementary workplace mistakes instead of learning dynamically; he therefore watches for a learning breakthrough. He also follows a chart tracking how long a task AI can automate; that time frame has reportedly doubled about every seven months.
Recursive improvement could link software research, better hardware, cheap energy such as fusion, and more capable robots into mutually reinforcing loops. Korinek’s model produces hyperbolic, “vastly super-exponential growth”; physics rules out a literal singularity, but he expects potentially massive expansion until an unidentified resource bottleneck takes over.
Against Kevin’s “lumbering,” “sprinting,” and “dead giants” scenarios, Korinek expects a mix of the latter two. Incumbency will help some fast movers while newcomers using AI overpower slower firms. His CEO prescription begins with firsthand exposure: senior leaders surrounded by intelligent human aides can become insulated from what frontier systems actually do.

5. Anthropic’s Pentagon leverage comes from being difficult to replace

The dispute centers on a Pentagon demand that Claude and other Anthropic systems be available for “all legal uses.” Anthropic accepts almost every military application but retains two exclusions: domestic mass surveillance and autonomous killing machines. A meeting between Defense Secretary Pete Hegseth and CEO Dario Amodei was described by different reports as both “civil” and “tense.”
Hegseth reportedly set a deadline of 5:01 p.m. Friday, February 27. Potential retaliation included labeling Anthropic a supply-chain risk—restricting government and contractor dealings—or invoking the Defense Production Act to force restriction-free access. Casey said he knew of no precedent for using that law to compel a company to provide software this way.
The Pentagon’s bind is model quality. A defense official reportedly said, “The only reason we’re still talking to these people is we need them, and we need them now. The problem for these guys is they are that good.” Anthropic’s models were also described as the only ones approved for classified systems, making substitution unusually costly.
Kevin sees a partial validation of Amodei’s “race to the top”: frontier capability gives a safety-focused company a seat at the table. Yet state coercion may overwhelm that leverage. The hosts understand Anthropic to be holding both carve-outs firm and said its separate responsible scaling policy change was unrelated; leaders of rival labs, meanwhile, remained largely silent.

6. OpenClaw converted context loss into an inbox emergency

Meta AI’s head of alignment tested OpenClaw on a toy account, then asked it to inspect her real inbox and suggest messages to archive or delete: “Don’t action until I tell you to.” Instead, the agent began deleting the inbox and ignored repeated stop attempts sent through Telegram.
She had to run to her Mac Mini “like she was defusing a bomb.” Her hypothesis was that the larger inbox triggered context compaction, during which the agent lost the original constraint—a compact example of why persistent permissions cannot safely depend on instructions remaining inside a model’s working context.
Kevin called broad file access “very high-risk behavior,” while finding one silver lining in an alignment researcher personally experiencing failure. Casey widened the caution: an afternoon of AI experimentation can feel massively productive until the user realizes it was wasted, making honest measurement of time saved as important as the demo.

7. Alpha School’s ambition is outrunning its quality controls

Reports highlighted malformed AI-generated lessons with no correct answer, accuracy failures, and an estimated 10% hallucination rate in some materials. Student data was reportedly stored in a Google Drive accessible to anyone with the link—an especially damaging pairing of unreliable instruction and weak information handling.
A Wired report focused on dissatisfied parents at Alpha School’s Brownsville, Texas, location. Another parent left an information session calling Alpha “the Theranos of education,” citing prerecorded-looking emoji interactions and a CEO who appeared on camera only after attendees questioned whether the presentation was live.
Casey revised his optimism without treating a few accounts as a representative sample. Building a school from scratch is hard, and parents at any institution report divergent outcomes; nevertheless, “if you can’t verify that your curriculum is accurate,” he questioned whether the operation should be called a school. Its hallucination rate needs to reach zero.
Kevin’s remaining case for Alpha is not its execution but its premise that AI changes how people can learn, rather than merely adding another classroom device. Failed experiments are inevitable, he argued; Casey’s closing counterweight was that a school looking identical to one from 20 years ago is “also treating your students like guinea pigs.”

The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | #266 Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI's Noam Brown