Summary

Meta’s reported plan to invest $14 billion–$15 billion for 49% of Scale AI is an expensive attempt to buy its way back to the frontier. Scale CEO Alexander Wang would leave the company and lead a new superintelligence team, but its first assignment is effectively to invent its own mandate: “The plan is we have to figure out the plan.”
Meta’s original AI advantage eroded because its research direction, incumbent incentives, and fast-follower playbook all pointed away from frontier models. Facebook’s AI work helped create PyTorch, yet Yann LeCun rejected the large-language-model path while OpenAI and others scaled it; Llama’s open-source strategy worked through Llama 3, but Llama 4 showed that the latest systems were no longer easy to copy.
Nine-figure compensation can attract attention, but the hosts doubt it can create a coherent lab quickly enough. Meta is reportedly offering packages reaching $100 million, including one credible $75 million offer, for a roughly 50-person team led by the 28-year-old Wang. The scarce few hundred researchers who have trained the largest models are already wealthy, mobile, and—in one researcher’s reply—“LOL, LMAO.”
The Scale transaction could weaken the asset Meta is buying by driving its biggest customers away. Scale supplies cleaned, structured, labeled data to Meta and rival labs; Casey’s source expected major customers to leave rather than expose proprietary usage to Meta. Ben Thompson argued that a 49% stake could raise regulatory concerns if it effectively removes an important player from the market while Meta is already under antitrust scrutiny.
Apple’s WWDC underscored that many of last year’s AI promises remain undelivered and undated. The cross-app Siri that was supposed to coordinate messages, email, calendars, and services still has not shipped; Craig Federighi defended Apple’s “mission” and values but would not provide a date. In its place, Apple presented Liquid Glass, message polls, resizable iPad windows, and a more capable Spotlight.
Apple’s larger risk is an unresolved identity crisis as mature hardware and services businesses face pressure without an obvious AI growth engine. Casey argued that probabilistic, messy AI conflicts with Apple’s polished, deterministic culture, while Kevin saw its high-profile “Illusion of Thinking” paper as evidence of continued institutional skepticism. The mitigating fact: even Google’s more advanced Pixel AI has not produced an obvious mass-market reason to abandon the iPhone.
Listener accounts suggest AI is already distorting labor markets and management behavior before it reliably replaces entire jobs. One junior engineer’s employer measures the claimed percentage of AI-written code and lays off low scorers, incentivizing everyone to lie; elsewhere, executives freeze hiring based on “20 AI miracles” while workers still need humans. Kevin’s preferred model is bottom-up experimentation, while Clay’s support team is building “expert generalists” who can move across functions.
The hosts argued that neither companies nor governments have developed a response proportional to the displacement their leaders predict. A listener proposed taxing AI to redistribute concentrated gains and slow deployment; Kevin cited Dario Amodei’s “token tax,” while Casey insisted elected officials—not corporations—must prepare the safety net. Their common conclusion was that the effects are visible now and likely to accelerate.

Deep dive

1. Meta is spending up to $15 billion to restart its AI race

Casey reported that Meta was preparing to take a 49% stake in Scale AI for somewhere between $14 billion and $15 billion. Scale co-founder and CEO Alexander Wang would leave the company and lead a new Meta team explicitly devoted to creating superintelligence.
Kevin’s state-of-play assessment was blunt: Meta is now considered “a second-tier AI research company,” marked by internal turmoil, disorganization, and messy strategic decisions. That is a striking position for a company holding one of Silicon Valley’s largest GPU stockpiles after spending heavily to prepare for the AI wave.
The competitive problem extends beyond model rankings. ChatGPT is growing rapidly, Google’s AI products have huge user bases, and Anthropic is building a major enterprise business, while Meta has not entered that conversation despite believing AI could shape social media, advertising, companions, and the metaverse.

2. Meta helped build modern AI, then chose a different path

Facebook tried to acquire DeepMind around 2012, then created FAIR under Yann LeCun after DeepMind chose Google. LeCun, a Turing Award winner and “godfather of deep learning,” recruited formidable researchers, and Facebook’s work on PyTorch became foundational infrastructure still used across major AI companies.
The fork came after Google’s 2017 transformer paper. OpenAI—and, to a lesser extent, Google and DeepMind—spent roughly five years building progressively larger language models and learning that performance improved with scale; Facebook and LeCun did not pursue that trajectory.
Company attention instead moved among election misinformation, content moderation, crypto, the metaverse, and competition with TikTok. Its shipped machine learning improved recommendations and detected prohibited content, but it did not produce the ChatGPT-style systems that captured public and developer interest.
LeCun’s skepticism was decisive because he did not believe in large language models and remains a prominent critic of the scaling era. Casey’s summary was categorical: “If you want to know why ChatGPT didn’t come out of Meta, Yann LeCun is sort of the reason.”

3. Llama’s fast-follower strategy stopped working at the frontier

After ChatGPT appeared in 2022, Meta went into panic mode, accumulated GPUs, and developed Llama. Early versions succeeded partly because Meta released them openly, letting developers build without paying the usage fees attached to proprietary systems such as ChatGPT.
Casey stripped away the altruistic framing: open source was a competitive weapon. Meta intended to give away something rivals sold for $20 a month, impose cost pressure, and slow OpenAI and Google while copying published advances closely enough for Meta’s own purposes.
That familiar fast-follower playbook had worked with products such as Snapchat Stories and appeared viable through Llama 3. Llama 4 exposed the limit: frontier models had become harder to reverse-engineer, and its reception suggested Meta had “lost its way” rather than remained one step behind the leaders.

4. Superintelligence may be both a mission and a recruiting pitch

Kevin offered two readings of Meta’s pivot: Zuckerberg has genuinely abandoned failed directions and will spend whatever it takes to reach the frontier, or Meta is adopting AGI and superintelligence rhetoric mainly to recruit people who would otherwise choose OpenAI, Google, or Anthropic.
Casey’s “somewhere in between” answer preserved the tension. Zuckerberg’s ambitions grew alongside what some might call Meta’s desperation: when AI merely supported current business objectives, superintelligence was unnecessary; once elite researchers would not join, he had to “change my tune on this front.”
The deeper mismatch is motivational. Frontier-lab believers talk about abundance, curing disease, and solving poverty; Casey believes many sincerely hold those grandiose goals. Zuckerberg, in her framing, wants Meta to remain among the world’s most powerful companies, yet “in a world where superintelligence exists, I’m not sure Meta will have much of a role to play.”
A January 2024 reorganization already used AGI language partly to attract researchers, but it did not deliver the desired results. The Scale investment is therefore another reset rather than Meta’s first recognition that its AI structure was failing.

5. Scale brings essential data—and a customer-conflict problem

Scale is not a frontier-model laboratory. Its subsidiaries hire people relatively cheaply to categorize content, images, and other material; Scale then cleans and structures those labels so customers can train classifiers and language models on higher-quality data.
That makes Scale a classic “picks and shovels company”: it helps customers scale AI but does not itself build the AI. Wang has proved resilient at following where the money moves, Casey said, but building superintelligence is fundamentally different from operating a successful data supplier.
Meta could gain privileged access to an important model-training ingredient, although existing multiyear contracts may prevent it from simply cutting rivals off. The more immediate risk is that Casey’s source fully expected Scale’s largest customers to leave because they would assume proprietary usage information could flow back to Meta.
Casey also relayed Ben Thompson’s argument that the structure might trigger regulatory concerns. Even without buying Scale outright, Meta could effectively remove an important provider while already facing antitrust scrutiny in a case seeking divestiture of WhatsApp and Instagram.

6. The new superteam has money but no operating blueprint

Zuckerberg plans to seat the new AI group physically near him, echoing how he surrounded himself with communications staff during the Cambridge Analytica crisis. The compensation is reportedly reaching nine figures, including one credible offer of $75 million.
Casey described the public rollout as a “help wanted ad” telling candidates that $100 million is available. But after someone accepts, there is no settled first-day program: “The plan is we have to figure out the plan”—import practices from rival labs, reshape Meta, and somehow return to the frontier.
Wang, age 28, may lead around 50 highly paid people whose working relationships still need to form. Casey asked whether a team could gel in under six months and noted that, if AGI or superintelligence is close, six months to a year and a half before a first major project ships could be consequential.
Kevin added that only perhaps a couple hundred people worldwide have trained the largest models on the largest supercomputers, and they are already rich enough to choose any employer. Meta’s AI companions and Anduril battlefield-headset project may be worthy or profitable, but he doubts they inspire this cohort; one leading researcher answered his recruitment question, “LOL, LMAO.”

7. Apple still cannot provide a date for the Siri it advertised

Apple’s central promise from the previous WWDC was a Siri capable of combining personal context across messages, email, calendars, and apps—the exemplar was arranging an Uber for someone’s mother when her flight arrived. One year later, that system still had not shipped.
Wall Street Journal reporter Joanna Stern pressed software chief Craig Federighi on why Siri was not as good as its competition. He reiterated Apple’s goal of something integrated, personal, and private rather than “a bolt-on chatbot on the side,” but said Apple wanted the product “very much in hand” before discussing dates.
Casey’s diagnosis was cultural: Apple excels at rigid, deterministic systems that are polished and predictable, whereas AI is “chaotic,” “messy,” and probabilistic. Reporting discussed by the hosts suggested there were too few internal true believers and that the company gave the technology short shrift to protect an already extraordinary business.
Kevin connected that history to AI leader John Giannandrea, or JG, whom Apple recruited from Google. According to the reporting he cited, JG viewed language models as a distraction, believed consumers disliked chatbots, and resisted major investment—the consequences are now visible in what Apple cannot ship.

8. Liquid Glass made Apple’s missing AI roadmap more conspicuous

Apple’s marquee announcement was Liquid Glass, an operating-system redesign that gives interface elements transparent, overlapping surfaces. Casey acknowledged that changing software used by hundreds of millions—or more than a billion—people matters, but early developer feedback said the translucency made screens harder to read.
Her critique borrowed Steve Jobs’s formulation, “Design is not how it looks. Design is how it works.” Liquid Glass seemed focused on greater beauty without explaining what the interface could now accomplish; Kevin caricatured the idea as, “What if we made a phone where everything was transparent and you couldn’t see anything?”
The rest of the presentation felt “small ball” to Kevin: polls and typing indicators in group chats, configurable chat backgrounds, live translation with uncertain language coverage, a desktop Phone app, and finally resizable iPad windows. Executives displayed “delirious enthusiasm” for features that did not amount to a futuristic vision.

9. Spotlight was useful, but Apple still lacks its next cash engine

Casey’s most substantive productivity highlight was Spotlight becoming more like a launcher app such as Raycast. Beyond locating files or opening Keynote, users could trigger shortcuts and actions from Command-Space, removing clicks from routines such as turning off all the lights at bedtime.
Even as a productivity enthusiast, Casey conceded that this “does not sound that interesting.” Still, it reflected the Apple she values: tools that help people work faster and become more creative, rather than the broader effort to discover “what is a seventh subscription we can sell you on this iPhone.”
Kevin framed Apple as caught between maturing businesses. iPhone sales had been flat to declining, successive models offered limited differentiation, and services faced antitrust and court challenges to App Store payment control; Apple had not identified “the next gusher of cash” or decided how central AI should be.
Casey’s defense complicates the bear case: Google’s more advanced Pixel features still offer no obvious reason for an average iPhone owner to abandon iMessage, while Amazon had reached only one million customers with its upgraded Alexa and was rolling it out cautiously. Every giant is struggling to turn capable AI into indispensable consumer products.

10. Apple’s reasoning paper became a proxy fight over AI belief

Apple researchers’ “The Illusion of Thinking” paper argued that reasoning models—OpenAI’s o1 and recent Gemini and Claude systems—do not think like humans and encounter limits as problem complexity rises. Skeptics seized on it as proof that scaling was hitting a wall and would not lead toward general intelligence.
Casey called that reception the “AI cope bubble”: people seeking reasons not to fear disruption treated the paper as “manna from heaven.” Her semantic objection was that anyone paying attention could have told you that LLMs do not reason exactly like human brains, so proving that difference is less revelatory than advertised.
Her technical objection was that the hardest tasks required more output tokens than the tested models were allowed. That limitation shows the systems cannot solve every problem, but it does not support the broader conclusion that they are unreal, useless, or incapable of materially affecting people’s lives.
Kevin said the paper did not change his view of reasoning models; it changed his view of Apple. A company presenting itself as near the frontier was directing high-profile intellectual energy toward demonstrating that the frontier was hype, reflecting the same unresolved institutional skepticism visible at WWDC.

11. AI’s labor shock is arriving through incentives before automation

Listener Christian Danielson challenged executives who predict a “categorically different level” of displacement yet offer no mitigation plan, asking why governments should not “tax the shit out of their technology” to redistribute concentrated wealth and slow deployment while policy catches up.
Kevin cited Sam Altman’s unconditional-cash research and Dario Amodei’s proposed “token tax,” under which some AI revenue would fund welfare and safety nets. Most industry figures, he said, have not advanced even that far; Casey countered that elected officials, not corporations, are responsible for governing society and should already be preparing.
Sarah, a software engineer who graduated in 2022, lost her first team to cheaper human labor, then joined an “AI-first” household-name company that evaluates developers by their claimed percentage of AI-written code and lays off low scorers. Everyone therefore says most of their code is written by AI, while openings for people with two years’ experience have nearly disappeared.
Kevin warned that eliminating junior roles and mentorship destroys the pipeline of future leaders. Casey called the measurement self-defeating: if executives mistake coerced claims for evidence that AI already performs 80% of the work, resulting layoffs could leave them in serious trouble—while Sarah worries that “the ladder is pulled up behind” new graduates.

12. Durable adoption starts with workers and broader human skills

A CFO at a $150 million-plus remodeling company described the “awkward middle”: employees use AI for emails and job postings but resist deeper changes in accounting and HR. He expects fewer people to do more and may replace staff who refuse to adapt.
Casey saw a durable tension: software often offers clearer value to managers than workers. Kevin recommended bottom-up experimentation—buy employees the tools, hold a hackathon or offsite, reward the best ideas—and rejected mandates tracking usage under threat of replacement as a strategy for “durable transformation.”
Another listener described an “AI addict boss” who froze process-job hiring while forwarding hacky LinkedIn posts claiming that familiar software would soon be obsolete, despite immediate headcount needs, bad tools, and privacy risks. Casey’s corrective was to define the business objective first, then decide whether humans or AI are the best route.
At Clay, support leader George Dilthey instead develops “expert generalists,” rotating strong hires through product, engineering, and marketing. Kevin argued that bespoke, high-touch service and direct knowledge of customer pain transfer across many jobs; he also suggested senior leaders rotate through customer service to understand what users experience. Casey agreed those ground-level experiences are precisely what an AI system cannot reproduce.

What You Missed in AI This Week (Google, Apple, ChatGPT)E196｜The Stablecoin War: Circle’s Rise and the New Forces Competing as Traditional Finance Enters the Fray