Summary

Apple’s trade-secrets suit puts OpenAI’s more-than-$6 billion IO hardware push under scrutiny, raising oversight and hardware-vision questions beyond a personnel dispute. Apple alleges job candidates were directed to bring prototypes and blueprints into recruiting “show-and-tell sessions,” that a former employee retained access to confidential storage, and that OpenAI tricked a supplier into revealing a metal-finishing process. Kevin Roose’s caveat was that only Apple’s filing is public, but his first impression was blunt: “These guys seem guilty as hell.”
GPT-5.6 Soul appears strong enough to reset frontier-model retention and pricing tactics. Casey Newton called it “a notable step up from 5.5,” while Anthropic repeatedly extended subscription access to Claude Fable rather than force users onto per-token API billing. Both hosts read that reversal as fear that customers would switch to ChatGPT or Codex, with Kevin expecting an eventual token-price war.
The OpenAI-Anthropic feud may benefit customers narrowly while making frontier-AI coordination more dangerous. Sam Altman mocked Anthropic’s doomy “Keep thinking” ad and its Fable access policies; OpenAI, Casey argued, enjoys playing the “scrappy challenger” while Anthropic’s mistakes receive leader-level scrutiny. But both hosts warned that labs motivated to “crush” one another may be unable to coordinate on model-release risks or a slowdown.
Nearly 200 economists and researchers are warning that transformative AI deserves an urgent economic research agenda. Their statement says AI may become radically more powerful within 10 years and drive change “larger than the Industrial Revolution, but unfolding over a vastly shorter timeframe.” Erik Brynjolfsson described the current period as the “suddenly” portion of an exponential curve, while stressing that disruption will unfold over years rather than instantly.
A measurable labor-market warning is concentrated among early-career workers, not headline unemployment. Early-career jobs fell 2.7% year over year while mid-career jobs grew 1.6%; headline unemployment remained near 4%, far below Kevin’s former 10%-20% expectation. Brynjolfsson said controls for interest rates, remote work, tech over-hiring, and education did not remove AI as an important factor, though AI is creating jobs as well as destroying them.
Whether AI complements or replaces labor depends partly on managerial scorecards, tax incentives, and product strategy—not technology alone. Brynjolfsson challenged a CFO who measured AI returns through headcount reductions, urging metrics such as new products, better service, and lower employee turnover. He also argued that lower marginal tax rates on capital than labor reward substitution, while defensible growth comes from new offerings and barriers to entry rather than cost-cutting alone.
Policy and trust risks are spreading across the AI stack, from power infrastructure to gambling, education, and consumer hardware. New York paused construction of hyperscale data centers using at least 50 megawatts for one year; ChatGPT displayed Kalshi odds that gave France a 60% chance before a 2-0 loss; and Brown students averaged 96% on a take-home midterm but 48.6% in person. Meta separately faced backlash over recording glasses and an opt-out Instagram likeness-remixing feature—evidence that deployment without trust remains a liability.

Deep dive

1. Apple alleges that OpenAI’s hardware foundation is “rotten to its core”

The hosts disclosed that The New York Times is suing OpenAI, Microsoft, and Perplexity, while Casey’s fiancé works at Anthropic—material context for an unusually contentious OpenAI segment.
Apple sued OpenAI, its Jony Ive-co-founded IO unit, and current OpenAI employees Qiang Lu and Tang Tan. Apple alleges theft “at every level” and says OpenAI’s hardware business rests “on the shakiest of foundations, rotten to its core.”
The filing claims Tan directed candidates to bring physical components, blueprints, and prototypes into recruiting “show-and-tell sessions.” Casey captured the absurdity: applicants supposedly needed “two forms of ID,” a résumé, and “an unreleased Apple product” spirited out of the lab.
Apple alleges Lu kept a work laptop, apparently exploited an unknown security vulnerability, accessed confidential network storage, and messaged, “LOL, I found out I can access the network storage.” It also alleges that OpenAI tricked an Apple manufacturing partner into demonstrating a secret metal-finishing process.

2. The suit compounds doubts about OpenAI’s hardware vision

Casey allowed that Apple treats virtually everything in Cupertino as a trade secret, so some claims may fail. But coaching candidates to bring unreleased hardware “does just seem transparently wrong”; the unresolved question is “how high up the chain did this go?” Neither Jony Ive nor Sam Altman was named.
Kevin emphasized that only Apple’s filing was available and OpenAI will answer, yet said the defendants looked “guilty as hell.” OpenAI’s statement—“We have no interest in other companies’ trade secrets”—struck Casey as a “non-denial denial” designed to buy time.
Bloomberg’s Mark Gurman described OpenAI’s first product as some sort of mobile, screen-free home smart speaker. Kevin heard little that clearly depended on Apple’s trade secrets and doubted OpenAI’s category-breaking vision; Casey heard a next-generation Echo or HomePod, recalling that Altman once insisted the project was “not just Alexa.” Her verdict now: “Sounds a lot like Alexa.”

3. Talent mobility makes the legal boundary economically consequential

OpenAI has hired more than 400 people from Apple over the past couple of years, a recruiting campaign Casey said even a company Apple’s size would notice. Kevin said Apple is struggling in AI; Casey read the lawsuit partly as Apple “lashing back” after a substantial loss of talent.
Kevin’s broader framing was California’s porous AI labor market: non-competes are difficult to enforce, so researchers continuously move among OpenAI, Anthropic, and Google DeepMind. Employees cannot carry code, designs, prototypes, or protected IP, but “just knowing how to build something is not a protected class of IP.”
Executive instability adds to the overhang. Fiji Simo stepped back after an extended medical leave and less than a year at OpenAI; Casey credited her with streamlining offerings and cutting distractions such as Sora, but said her departure represented “yet more executive turmoil.”

4. GPT-5.6 Soul is already changing Anthropic’s commercial calculus

GPT-5.6 Soul became available after a period when access was withheld at the White House and federal government’s request. Kevin’s trusted testers called it highly capable; Casey, using it to critique vibe-coding projects, called it “a notable step up from 5.5.”
Anthropic had planned to remove Claude Fable—its most powerful model—from subscription bundles in early July, leaving per-token API access. It extended the deadline once and then again; Kevin’s read was that cutting Fable now would drive subscribers toward ChatGPT or Codex.
Casey inferred that Anthropic had expected several weeks for users to become “addicted to Fable” before competitors responded. He said a fight with the U.S. government meant Fable arrived only days before GPT-5.6, so Anthropic is now trying to get people hooked on Fable before pushing them toward paying per token.
Kevin consequently expects a token-price war: Anthropic cuts prices, OpenAI undercuts them, and frontier access becomes cheaper. Casey agreed that competition is good for customers, while warning that some dimensions of the underlying AI race remain frightening.

5. Brand warfare has become a proxy battle for frontier leadership

Anthropic’s campaign opened with a burning building and Arlington National Cemetery while asking whether AI can be trusted and who will “hit the brakes.” Casey called it “a disaster”: ordinary viewers were unlikely to leave with a positive view of either Anthropic or AI.
Kevin initially defended acknowledging real fears rather than selling sunny optimism. Casey’s pushback—worth keeping—was about format: leaders or a serious essay could address catastrophic risk, but brand marketing ending with “Keep thinking” sounded like advice to seniors to keep doing crossword puzzles while superintelligence does the thinking.
Altman reposted the ad, said he thought it was satire from “C1audAI,” and attacked Anthropic for restricting Fable, silently downgrading users, or denying access. Casey saw OpenAI exploiting its new underdog status: Anthropic is perceived as the leader, so its errors are magnified while OpenAI throws “scrappy challenger” punches.

6. Commercial hatred threatens the cooperation frontier safety requires

Kevin said the labs’ animosity is “more than a Coke and Pepsi thing”; researchers fear losing talent and the jokingly named “mandate of heaven” that comes with having the best model. When Dario and other Anthropic leaders left OpenAI in late 2020, they signed a non-disparagement agreement that Kevin said, to his knowledge, has not been rescinded. Casey quipped that their behavior suggests a disparagement agreement.
The consumer upside of rivalry did not erase Kevin’s larger warning: people at both companies want to “crush” the other side, even as increasingly powerful systems may require joint risk management. “You can’t do that if you’re constantly fighting each other.”
Casey agreed, citing Demis Hassabis’s proposal for a regulator governing frontier-model releases and his hope that labs could eventually collaborate on slowing development. Leaders need to call one another and consider “what is good for the rest of the world,” not merely what earns retweets.

7. Economists now see transformative AI as an urgent macroeconomic problem

The “We Must Act Now” statement says AI may become radically more powerful during the next 10 years, producing transformation larger than the Industrial Revolution on a much shorter timeline. It pairs large-scale displacement risk with possible major living-standard gains and calls for incentives, guardrails, and institutions that complement humans.
The signatories initially numbered nearly 200 and included Nobel laureates, Daron Acemoglu, David Autor, Anton Korinek, and economists previously skeptical of near-term job-loss claims. Brynjolfsson’s first objective was to catalyze a “sea change” from treating AI as a few tenths of a productivity point to treating it as a central economics research problem.
Brynjolfsson taught his first AI course in 1987. His signature framing: exponential curves move “slowly at first and then suddenly,” and “especially this year we’re in that suddenly part of the curve,” while economic understanding remains nowhere close to the technology.

8. Labor disruption is lagging capability because organizations have not been rebuilt

Kevin offered a clear change of mind: shown Claude Fable or GPT-5.6 in 2020 and told they would exist in 2026, he would have predicted 10%-20% unemployment. Instead, headline unemployment hovered near 4%, roughly where it had been several years earlier.
Brynjolfsson was “not at all” surprised. Electricity initially produced little productivity or employment change because factories merely “paved the cow paths”; redesigning factories and the economy took 20-30 years, ultimately destroying and creating millions of jobs.
AI should move faster than electrification, but “it’s not the kind of thing that happens in 20 to 30 months either.” Brynjolfsson expects years of organizational reinvention, with substantial job creation and destruction still in their earliest stages.
Stanford’s Canaries dashboard shows early-career jobs down 2.7% year over year and mid-career jobs up 1.6%. Brynjolfsson would not claim profession-wide consensus, but said the pattern first published in August 2025 has strengthened; examining rates, remote work, tech over-hiring, and education did not remove AI’s importance.

9. Managers and governments can still steer AI toward complementing workers

Economics writer Noah Smith rejected the letter’s steering language: if inventors could not predict the last wave’s labor effects, why assume they can design complementary AI? Brynjolfsson welcomed the “right amount of friction” but answered categorically: “Noah’s wrong about this.”
His historical case was that incentives direct invention: carbon taxes are intended to influence technological direction, and research shows that engine technology evolved differently in countries with greater taxes. Computing has also long contained both Turing’s human-imitation path and Doug Engelbart’s amplification path.
One CFO told Brynjolfsson every department had to quantify AI through headcount reductions. He urged her to measure new products, better customer service, and lower employee turnover instead: changing managerial objectives would produce systems that “complement people more, rather than just replace them.”
Kevin remained skeptical that enlightened corporate leaders would resist replacing humans. Brynjolfsson’s answer combined ethics with self-interest: new products and barriers to entry are more defensible than cost-cutting, while lower marginal taxes on capital than labor currently reward replacing workers with machines.

10. Better data, broader insurance, and public trust are the missing guardrails

Brynjolfsson warned that during extraordinary economic turbulence, “we are destroying our instrument panel” by cutting statistical agencies. His lab is supplementing public data with Ramp, Revelio, and ADP, while updating O*NET’s task taxonomy to obtain more granular, real-time indicators.
Erik said he has resisted offers from frontier labs. He acknowledged that their researchers can do valuable work with remarkable data, but argued that independent academic voices remain important because even unbiased researchers at profitable labs may be perceived as speaking for their employers.
A sovereign wealth fund financed by AI companies was “not my first choice”: Brynjolfsson worried that recipients of a simple payout lack leverage over whoever allocates it. He favored exploring it alongside wage insurance—despite moral hazard—and government incentives to hire young workers, because “you don’t get mid-career workers unless they first start as young workers.”
In HatGPT, ChatGPT displayed Kalshi World Cup odds giving France a 60% chance before a 2-0 loss. Casey objected to gambling promotion inside a paid subscription; Kevin called AI capability and gambling addiction a “societal Voltron” likely to spread across every available surface absent regulation.
New York separately paused construction of hyperscale data centers using 50 megawatts or more for one year. Kevin wanted rigorous evidence rather than suspect environmental claims; Casey stressed that opposition is bipartisan, citing Utah, and said developers must prove they will not raise electricity prices or damage the environment.
Consumer trust broke elsewhere: Brown students averaged 96% on a take-home midterm, versus a usual 65%-80%, then 48.6% on the in-person final. The take-home exam followed a school shooting on Brown’s campus, which Kevin called a generous and compassionate response; he said the score gap left no other explanation than widespread AI cheating. Lorde attacked unnamed “effed-up” smart glasses as “not sexy,” while Meta also paused Instagram’s opt-out public-likeness remixing after backlash—another instance where deployment without consent followed by “We’ve heard the feedback” failed to restore trust. Users had also found ways to disable the recording indicator on Meta’s glasses, intensifying privacy concerns.

How Decagon Runs 90% of Its Agents on Open-Source Models Building an Autonomous Enterprise for Real-World Services with Netic Founder Melisa Tokmak