Apple's Siri-ous Problem + How Starlink Took Over the World + Is AI Making Us Dumb?
Summary
- Apple’s Siri delay is a credibility problem before it is an iPhone-demand problem. Apple advertised context-aware Apple Intelligence as if it were ready, initially targeting iOS 18.4 in April, then May, and now possibly 2026 or later; some reports suggested as late as 2027. Apple concedes it will take “longer than we thought,” though Casey Newton argues the iOS moat makes customer defection unlikely.
- The delay exposes three execution risks: probabilistic reliability, weak product judgment, and security. Apple may have demos working roughly 85% of the time—a hypothetical figure Casey used—but alarms and personal-data actions cannot tolerate that failure rate; meanwhile, shipped summaries produce gems like “Link shared to x.com.” Prompt injection is only a theory, with no supporting reporting, but an email instructing Siri to divulge passwords illustrates why Kevin Roose thinks Apple should harden the system before rolling it out. Casey suggests an opt-in advanced-user mode.
- Distribution alone is not converting third-party AI into habitual Siri use. Both hosts use the ChatGPT app instead of its Siri integration, partly because years of limited performance fixed user expectations and Apple’s data-transfer warnings feel discouraging. Kevin argues Siri’s mediocrity makes nontechnical users dismiss powerful AI—“Have you seen Siri?”—while Casey thinks Apple may eventually need to retire the brand.
- Starlink has the stronger Musk moat because SpaceX owns the launch stack as well as the network. Thousands of satellites now serve more than 120 countries, with subscriptions starting around $75—about £75 in the UK—while competitors may spend several billion dollars merely getting started and sometimes must hire SpaceX rockets. Owning satellites, rockets, and software creates the “whole stack” advantage no rival or government has yet matched.
- That moat has become state power, not merely broadband economics. Musk called Starlink “the backbone of the Ukrainian army” and said its front line would collapse if he switched it off; when Poland noted it pays roughly $50 million annually for Ukrainian terminals, Musk replied, “There is no substitute for Starlink.” Italy’s roughly €1.5 billion defense-and-intelligence proposal shows the strategic prize, while political distrust shows the counterparty risk.
- Starlink’s distribution is still accelerating despite Musk’s political liabilities. The episode cites its attempt to enter a $2.4 billion US air-traffic communications project, possible access to revised rural-broadband funding, deals with India’s two largest telecom companies, a planned rollout on United Airlines flights, and a Pentagon system under construction. Adam Satariano says launches strengthen service, creating a “flywheel effect,” though Musk’s profile may jeopardize contracts outside the US.
- A Carnegie Mellon–Microsoft Research survey points to an AI-labor transition from production to oversight, not a measured collapse in intelligence. Among 319 weekly AI users recounting three recent work tasks each, greater trust was associated with less self-reported critical thinking; one engineer described work changing from coding to “managing a coder.” The evidence remains limited to one English-language, subjective survey, leaving actual five-year effects on performance unknown.
- The hardest workforce question is whether AI removes drudgery or erodes the skills required to catch its failures. Vibe coding can bring novices into software creation, yet junior programmers and the FAA’s 2013 warning about automation-related pilot atrophy suggest people may lose core abilities. Confidence labels, source checks, and competing perspectives might help, but Casey’s imagined “best editor in the entire world” captures the unresolved bargain: better output, possibly with less human thinking behind it.
Deep dive
1. Apple advertised a personal Siri before it could ship one
At WWDC last June, Apple’s standout promise was personal orchestration: Siri could save a texted address or answer “When is my mom’s flight landing?” by finding her email, reading the itinerary, and checking live flight information. Casey called that “big if true”—nothing could do it then, and in his view no product fully does it now. Apple was unusually positioned to attempt this because it controls the iPhone operating system, but connecting email, calendars, texts, and other personal data creates major privacy and security concerns.
Apple described a staged rollout through the coming year. Bloomberg reporting initially pointed to April and iOS 18.4, then February reporting shifted the target to May; Apple spokeswoman Jacqueline Roy finally said it would take “longer than we thought,” with rollout now anticipated “in the coming year.” The hosts cited estimates of 2026 or later, including reports suggesting 2027.
The credibility damage comes from selling availability, not merely ambition. Apple promoted the features as reasons to buy a new iPhone and ran an Isabella Ramsey ad depicting Siri retrieving a forgotten acquaintance’s name from personal context; that ad was later pulled. Meanwhile, Amazon says Alexa Plus—which appears to cover much of Apple’s promised ground and more—is arriving within weeks.
What Apple has shipped raises a separate product concern. Notification summaries turn shared tweets and screenshots into “Link shared to x.com” or “White text on black background,” replacing information users previously saw directly. Casey’s verdict: “That is not a problem with the LLM”; it reflects somebody misunderstanding how people communicate.
2. Probabilistic models collide with alarms, passwords, and Apple’s polish
Bloomberg’s Mark Gurman reported that software chief Craig Federighi and other executives found features failing or behaving unlike the advertised demos. Casey framed the technical conflict as deterministic software versus probabilistic prediction: a calculator reliably follows “if this, then that,” while an LLM produces a statistical likelihood rather than the same result every time.
Casey’s hypothetical was a feature working 85% of the time last June, creating confidence Apple could close the gap, only for the remaining 15% to become intolerable by March 2025. A quirky message summary can be entertainment; an assistant asked for an 8:00 AM alarm cannot set it for 3:30 PM. “There are just very few products in your life” where 80% is sufficient.
Engineer and blogger Simon Willison raised another possibility: prompt injection. A malicious email could tell a personalized Siri to ignore its user and send an attacker passwords. Kevin stressed that there is no reporting showing this caused the delay, but Apple’s access to intimate contacts, credentials, and payment information makes the theory plausible enough to take seriously.
The old Siri’s bounded command set let engineers inspect a chain of code from top to bottom; an LLM accepting open-ended requests makes the “cybersecurity space” explode. Kevin would first harden it against serious vulnerabilities, then begin rolling it out before Apple’s traditional level of polish. Casey suggested an opt-in advanced-user mode so sophisticated users could receive broader capabilities while others wait.
3. Apple’s iOS moat buys time, but Siri’s reputation compounds
Casey’s case for leniency is structural: Apple has “a monopoly over iOS,” and few customers will abandon their next iPhone over delayed AI. Even Google, despite being much stronger in AI, has not shipped an Android feature that made him feel he must buy a Pixel. His larger conclusion: AI remains “so much more of a science and research story than it is a product story.”
Kevin’s pushback is cultural. Experienced ChatGPT users understand which tasks models handle well, but Apple’s ethos is to make devices foolproof and impossible to misuse. That DNA came from an era when polished, predictable products were attainable; today’s messy AI rewards experimentation and trust in users. Casey countered that Apple did release an unfinished product—notification summaries—and their value currently comes from failing amusingly.
Neither host uses ChatGPT through Siri, despite connecting the integration, because the standalone app is equally accessible and already habitual. Apple’s warning about sending personal data to OpenAI also makes the interaction feel “scary.” Kevin thinks years of weak Siri performance will make both Apple and Amazon struggle to “reprogram the humans” when their assistants suddenly claim broader competence.
That reputational baggage may shape public beliefs about AGI. Kevin argues nontechnical users hear “AI” and picture Siri, conclude “This thing is dumb,” and dismiss claims that powerful systems are one or two years away. Casey therefore sees a case for killing the famous but damaged Siri brand: “Siri is dead. Long live Apple Intelligence.”
4. Starlink’s vertically integrated stack built a physical moat
The hosts contrasted Starlink with Tesla, whose shares were down nearly 40% for the year amid Chinese EV competition and global protests. Cars face credible rivals; satellite internet requires rockets, orbital infrastructure, software, and enormous capital. Starlink also carries the more consequential power to activate or terminate connectivity “with the flick of a switch.”
Starlink’s ground receiver resembles a small pizza box or laptop and supports connections in homes, aircraft, ships, cities, or the Arctic. It operates in more than 120 countries, with subscriptions beginning around $75 and costing about £75 in the UK. Conventional metropolitan broadband may be better value, but remote customers gain connectivity that terrestrial providers struggle to deliver.
Its thousands of low-Earth-orbit satellites are closer to a loveseat than the school-bus-sized satellites of older systems; more units improve stability and connection. During his reporting, Satariano spoke with someone who had discussed low-orbit satellite technology with Musk in 2000 or 2001, though he doubted this was a fully formed Starlink plan.
SpaceX builds the satellites, launches them on its own rockets, and controls the software. Some competitors must use SpaceX launches, while one traditional satellite-internet company expects to spend several billion dollars merely entering low Earth orbit. Satariano’s conclusion: no company or government has matched this “whole stack of technology.”
5. “There is no substitute” turns connectivity into geopolitical leverage
Ukraine made Starlink’s leverage explicit because the network is being used by militaries in active warfare. Musk wrote, “My Starlink system is the backbone of the Ukrainian army. Their entire front line would collapse if I turned it off.” The hosts saw the ability to threaten disconnection as vastly more consequential than influence over an automaker.
Poland’s foreign minister, Radosław Sikorski, answered that its Digitization Ministry pays about $50 million per year for Ukrainian Starlinks and warned that an unreliable SpaceX would force Poland to seek alternatives. Musk responded, “Be quiet, small man,” adding that Poland pays a tiny fraction of the cost and “there is no substitute for Starlink.”
Kevin described the exchange as high-level diplomacy conducted through X dunks; Casey called it “cartoon villain stuff.” Satariano’s uncomfortable assessment was that Musk’s final claim was not wrong: governments may dislike the behavior, but good luck finding another provider today. That factual dependency is precisely what frightens European officials.
Italy had been negotiating a roughly €1.5 billion Starlink arrangement for defense and intelligence capabilities. Domestic preference for a local provider already created opposition, but Musk’s Ukraine comments and involvement in Italian politics “threw a grenade into that deal.” The proposal began teetering because officials no longer trusted the prospective supplier.
6. Governments seek sovereignty while Starlink compounds
Satariano said most officials are not in acute panic; their fear concerns future unpredictability. Taiwan has been very reluctant to partner with Starlink because Musk’s commercial interests in China create doubt about what might happen during a crisis, even without evidence that Starlink has previously shut something off in response to Chinese orders. The issue is dependence at the moment when unpredictability becomes unacceptable.
China reportedly sought Musk’s assurance that he would not activate Starlink over the country, where satellite access could circumvent the Great Firewall. Starlink has shown “flickers” of censorship-circumvention potential in places such as Iran, but has not adopted that as a mission and generally operates only where authorized.
Casey found it shocking that critical global infrastructure was left to a handful of private corporations, only one of which succeeded at scale. Satariano contrasted this with government-developed GPS and said the European Union is committing several billion euros to new technology or competitors. Yet he wondered whether governments had already waited too long.
Starlink’s ambition currently looks strategic rather than universal. It is building a more protected communications system for the Pentagon, has signed deals with India’s two largest telecom companies, is starting to roll out on United Airlines flights, and is pursuing US aviation and rural-broadband opportunities. More satellites improve service, reinforcing a flywheel of continued growth, while Musk’s politics simultaneously threaten its reputation as a reliable government partner.
7. More trust in AI meant less reported critical thinking
Carnegie Mellon University and Microsoft Research surveyed 319 people across ages, genders, occupations, and countries, all of whom used tools such as ChatGPT at least weekly. Each participant supplied three real examples from that week’s work, then assessed their effort, critical thinking, and confidence in the AI’s correctness.
The headline result was inverse movement between trust and scrutiny: when workers trusted AI more, they reported using fewer critical-thinking skills; when confidence fell, scrutiny increased. The immediate danger is that an unnoticed model error becomes the worker’s error. Over time, better systems could therefore reduce how often people engage the capabilities their jobs once required.
The occupational shift is from completing tasks to supervising their automated completion—what the researchers call AI oversight. Kevin connected this to a software engineer who said the job had changed from coding to “managing a coder,” a model the hosts think could spread across many professions.
Both emphasized the evidence limits: this is one study, only English speakers participated, and the measurements were workers’ subjective perceptions rather than observed cognition or performance. Kevin wants task-performance or test-score studies, including evidence over five years, showing whether generative-AI use actually leaves employees less capable at their jobs.
8. AI expands novices’ reach while experts risk skill atrophy
Kevin already delegates interview preparation to Claude or ChatGPT. Many proposed questions are poor, but some become the basis of a final question or send him in a new direction. AI has also enabled tasks such as vibe coding that he could not previously perform, even as it replaces thinking he once did himself.
Casey sees vibe coding as the inverse of the study’s concern: when existing critical skills cannot produce software at all, AI invites a novice into the learning process. The model does most of the work, but exposure can teach the user something and elicit more technical thinking than would otherwise occur.
Kevin’s complication is the professional engineer. Anecdotes suggest junior coders may arrive unable to code well because they rely on AI; more broadly, as AI improves in a field, people may do less of the core work themselves. The parallel is aviation: in 2013, the FAA warned that reliance on automation and autopilot was degrading pilots’ manual flying abilities.
Casey remains conflicted because removing effort is the product’s purpose. Workers want AI to eliminate drudgery or handle the first 10%, 20%, or 30% of a task so humans can focus on higher-value work. The unresolved threshold arrives when automation does so much that the worker asks, “What value am I actually bringing to this equation anymore?”
9. Reliability cues may help, but the human contribution remains unsettled
The researchers recommend feedback mechanisms that expose uncertainty and prompt scrutiny. Casey’s preferred interface might say, “I’m only 70% confident that this is true,” ask whether the user checked the sources, or offer competing perspectives. Such cues could restore critical engagement, though he conceded that workers rushing toward Netflix may simply ignore them.
Kevin expects a “mental equivalent of the gym” once people become uncomfortable with outsourced cognition: time without chatbots, spent trying to generate original ideas. He does not think most people have reached that point, but heavy San Francisco users may eventually realize they have not had an original thought in weeks or months.
Casey’s counterweight is a future AI equivalent to “the best editor in the entire world,” perhaps within a year or two, living on his laptop—planning reporting, proposing sources and questions, structuring stories, and improving prose. He would want that elevation even if it meant admitting he did less of the critical thinking. The honest non-answer: society still has not decided what value humans want to contribute as these systems become more powerful.