Edition 1

The story “new grads are getting crushed by AI” is outrunning the measurement. That doesn’t mean the anxiety is fake; it means the scoreboard most people cite may not be built to catch the kind of shift they’re describing.

Ars Technica reports on a CESifo working paper by economists Robert Fairlie and Jane Wu that goes looking for early signs of AI-driven displacement among recent US college graduates—and mostly doesn’t find them in unemployment data, at least not yet. Using detailed microdata from the US Census Current Population Survey (CPS), the researchers track unemployment among bachelor’s degree holders aged 22–25 who aren’t enrolled in additional schooling. Ars writes that the summer 2026 unemployment rate for that group was 7.3%, which sits within the range seen in earlier summers from 2022 to 2024 (6.3% to 7.8%). Across a set of comparisons—against non-college peers in the same age band and against older college grads—Ars says the paper finds trend differences from 2022 to 2026 that are mostly not statistically significant. The authors call this summer-2026 readout a “useful first test,” and they don’t see evidence of “significant, widespread displacement or reduction in hiring” for new grads in that period. Ars Technica

That should cool the loudest claims. It shouldn’t end the argument.

Because the real fight here isn’t “is AI hurting grads, yes or no?” It’s “what would ‘hurting grads’ look like in the data we’re choosing to treat as definitive?” Unemployment is a blunt instrument for a fine-grained shift. It’s good at answering a specific question—who meets a survey’s definition of unemployed at a moment in time. It’s worse at capturing whether graduates are taking lower-quality matches, accepting weaker pay, or seeing career starts tilt toward roles with less training and less progression. Those can be genuine harms even when the unemployment rate doesn’t spike.

Ars notes the paper’s motivating mechanism: if AI gets good at relatively standardized tasks common in entry-level office work, firms might respond first by hiring fewer new grads rather than laying off experienced workers. That’s a plausible channel—and it comes with an uncomfortable implication. If the earliest AI effect is “fewer openings,” the first thing to disappear may be opportunity, not employment status. In that world, a stable unemployment rate can coexist with a deteriorating on-ramp.

Ars also explains why the researchers think 2026 is an especially relevant moment to check. The paper points to signs of accelerating adoption: a sharp increase in the number of firms reporting they’re replacing many employee tasks with AI in a Census survey, plus broad increases in AI spending per employee and ChatGPT Enterprise token usage over the past 12 months. And Ars recounts high-profile warnings from figures like Marc Andreessen and Larry Fink, who suggested AI capability improvements could show up in new-grad outcomes.

So if you expected a visible labor-market shock, you can at least see why someone would go looking for it in 2026.

But here’s where I become The Auditor about it: a “no unusual spike” finding is not the same thing as an “all clear,” and the incentives around that confusion are obvious even without a villain list. People who want to calm a jittery audience will lean on the most comforting metric. People who want to escalate the narrative will dismiss the metric as irrelevant. In both directions, the number becomes a prop.

Ars makes the measurement clash explicit by placing this CPS-based analysis next to a Stanford study it covered earlier that found entry-level employment in “AI-impacted” occupations lagging other fields. Ars attributes part of the divergence to different datasets and what they can and can’t see: Stanford used payroll data from ADP, while the CESifo authors used CPS survey microdata. Ars notes that ADP is about job counts on payrolls in certain categories, while an unemployment rate reflects both the supply of jobs and the demand for them, filtered through labor-force definitions (including who is actively looking). Those aren’t interchangeable lenses; they can disagree without either side being “fraudulent.” They’re answering adjacent questions.

That’s the heart of my complaint about the discourse: we keep trying to force a multidimensional transition into a single headline verdict. “Unemployment didn’t explode” is not proof that entry-level work hasn’t changed. And “a different dataset shows a slowdown” is not proof that AI has already demolished the graduate job market.

If you want to know what I’m watching instead—what would actually settle arguments over time—it’s a basket of measures that unemployment only gestures toward: how long it takes to land a first job, where starting wages move for entry-level cohorts, how many grads are in roles that don’t require a degree, how job-to-job progression changes in the first two years, and whether the mix of entry-level postings shifts as firms redesign work around AI. None of that is in the CESifo unemployment result Ars summarizes. That’s precisely the point: the absence of a collapse in one metric doesn’t tell you what’s happening in the rest of the system.

Ars does report that the CESifo researchers break results out by “AI exposure,” using a 2023 study describing which roles AI systems are best equipped for. That kind of categorization can be useful for comparisons, but it’s also where measurement choices quietly steer conclusions—what counts as “exposed,” which tasks are assumed automatable, and how those assumptions age as models change. If we’re going to treat these studies as policy-grade signals, the definitions deserve as much attention as the coefficients.

Finally, Ars includes the caveat too many commentators skip: the authors warn that current trends don’t imply future performance, and that if workplace AI use keeps intensifying, later graduating classes (2027 and beyond) could be more affected than 2026. Additional years of data will be needed to see whether effects emerge as AI use deepens.

So no, the unemployment data Ars summarizes doesn’t show new grads “getting crushed” in 2026. But if you stop there, you’re letting a blunt tool do a fine tool’s job. The responsible posture isn’t panic or dismissal. It’s insisting on better measurement before we turn one summer’s unemployment rate into a national verdict on an entire generation’s prospects.

PrismML getting a “tiny” language model to run locally on Qualcomm’s Snapdragon AR1 Gen 1 platform is the kind of story that’s easy to misread as a niche optimization. TechCrunch reports that Qualcomm showcased PrismML’s 1-bit “Bonsai” LLM at its Snapdragon Summit, and that the model can run on-device on AI smart glasses built on the Snapdragon AR1 Gen 1 Platform TechCrunch. The glasses-tuned version TechCrunch describes is a 2-billion-parameter vision-and-language model meant for real-time “what am I looking at?” queries. TechCrunch also notes PrismML’s broader pitch: open-weight AI that runs on devices, as an alternative to relying on proprietary AI labs’ privacy promises and their appetite for more compute. One sobering detail in the same report: no smart glasses product running PrismML has been announced yet.

Those facts are modest. My read is not: on-device LLMs on glasses aren’t interesting because they’re novel; they’re interesting because they drag AI out of the world of infinite server elasticity and into the world of hard product envelopes. The demo (as described) is a proof of feasibility for local vision-language interaction on Qualcomm’s AR1 Gen 1-class hardware. Feasibility is the first gate. It isn’t the last.

Here’s the practical constraint I care about: once you put intelligence on a wearable, “capability” is inseparable from *duty cycle*. A model that can answer one question is not the same as a model that can answer questions all day without wrecking the experience. Wearables are where every subsystem—compute, camera, wireless, and the rest of the stack—competes for the same limited power and thermal headroom. That’s not something TechCrunch quantifies in this report; it’s the systems reality that determines whether a summit showcase becomes a product people trust.

TechCrunch says PrismML’s claim to fame is shrinking larger models substantially—here, “by 4x”—while retaining almost all benchmark performance. I treat that less as a benchmark brag and more as an attempt to buy room inside that wearable envelope. “4x smaller” can translate into different things depending on what’s actually being reduced (memory footprint, bandwidth, compute), but directionally it’s the right axis: the difference between an assistant that is intermittently available and one that feels like a utility is often just whether the system can afford to run the intelligence loop frequently.

Running locally also changes what “real-time” can mean. TechCrunch frames the glasses model as something wearers can use to ask what they’re seeing in real time. My opinion: local inference is one of the few ways to make that interaction feel immediate *when the network is not the hero*. That doesn’t guarantee low or “predictable” latency by itself—device load, camera pipeline, and scheduling matter—but it removes an entire class of variability: round-tripping sensor-derived inputs to a remote system and waiting for a response.

The privacy implication is similarly mechanical, not moral. PrismML is explicitly positioning on-device open-weight models as an alternative to depending on the privacy promises of proprietary AI labs. I buy the architectural framing: if a task can be completed locally, you have the option to keep raw inputs local for that task. But “option” isn’t the same as “outcome.” Products routinely default to cloud paths because telemetry, rapid iteration, and feature pressure are powerful incentives. If PrismML’s goal is to make privacy less about trusting a remote vendor, then the real test is whether partners build experiences where local is the default path for the mundane, high-frequency interactions.

This is also where the open-weight story cuts both ways. Open weights can lower dependency on a single proprietary provider, and TechCrunch says PrismML’s larger goal is to make better use of compute devices already have. But my systems concern is that you don’t escape operational cost—you relocate it. Once inference happens on clients, you inherit a different backlog: model packaging, update mechanisms, compatibility across device variants, and the messy edge of extreme compression techniques (TechCrunch points to “1-bit” Bonsai as the showcased model). Compression is often what makes edge deployment possible; it’s also where surprising failure modes like brittleness or quality cliffs can show up in product behavior.

And because TechCrunch notes there is not yet a shipping glasses product announced with PrismML onboard, we’re still missing the only set of facts that ultimately matters for users: sustained behavior on real hardware, in real apps, under real usage patterns. Summit demos are designed to prove a point, not to expose the long tail.

So what can on-device AI on Qualcomm “do,” based on the reporting? It can run PrismML’s showcased 1-bit Bonsai LLM locally on Snapdragon AR1 Gen 1-class smart glasses hardware, and PrismML has tuned a 2B-parameter vision-language variant toward real-time visual question answering. What can it *not* do—at least not yet, as established in this source? It can’t yet be evaluated as a shipping consumer product experience, because TechCrunch says no PrismML-powered glasses have been announced.

My bottom line is narrower than the hype cycle wants, but more useful: this is an existence proof for on-device vision-language on Qualcomm’s smart-glasses platform, and a marker for where the costs will surface next. If someone actually ships it, the winners won’t be decided by a single benchmark line about “almost all performance.” They’ll be decided by whether the local model is good enough, often enough, inside a wearable budget—while keeping the privacy story architectural rather than aspirational. The demo opens the door; shipping is where the bill arrives.

Founder control ahead of an IPO isn’t a halo. It’s a pre-commitment: deciding, before the noise of public markets arrives, who will be allowed to say “no” when “yes” is profitable.

Here are the bare facts, as TechCrunch relays them. The company is asking shareholders to approve a structure that would give CEO Dario Amodei and his six co-founders special shares carrying a combined 50.1% of the vote on most corporate matters, so long as at least three of them keep a minimum stake; TechCrunch attributes those specifics to The Information TechCrunch. TechCrunch also notes this isn’t a novel tactic in Silicon Valley—super-voting shares have kept founders like Mark Zuckerberg and Evan Spiegel in control—but says the group approach is what’s unusual here.

My quiet suspicion is that the main story isn’t “mission protection” in the abstract. It’s insulation from the one accountability mechanism that reliably shows up after an IPO: the market’s ability to punish delay.

Public markets don’t just bring new owners; they bring a clock. Every quarter becomes a referendum, and restraint becomes legible to outsiders as missed targets. In that environment, “safety” isn’t a mood or a value statement. It’s a veto power—one that costs money in the short term, and therefore requires someone to have the standing to impose it.

This is why the voting math matters more than the PR. TechCrunch says the special shares would carry no extra economic value, but would preserve the co-founders’ control after the company starts trading publicly. That detail is the point: not more cash, but the steering wheel.

And the steering wheel can be used in two directions.

Yes, concentrated control can protect restraint. A controlled company can decide not to ship a capability yet, not to widen access, not to optimize for the most addictive metric, not to push a model into workflows where it will predictably create harm. Those “nots” become expensive once analysts can line you up against competitors who did ship, did widen, did optimize, did push.

But concentrated control can also protect something less noble: the ability to define “safety” however leadership needs to define it in the moment.

When outsiders can’t credibly threaten governance consequences—when the founders effectively have the votes—then “we did this for safety” can turn into a conversation-stopper. Not because the claim is always false, but because the structure makes it harder to contest. The same small circle sets the pace of deployment and also gets to decide what responsible deployment looks like. The loop tightens. It gets quieter. It gets durable.

TechCrunch describes additional counterweights: Anthropic’s Long-Term Benefit Trust would still choose most of the board; the founders’ board seats would increase from two to three; and employees would get their own stock to break ties on some issues. On paper, that’s a real attempt to avoid the clean, brutal “dual-class and everyone else can cope” model. It suggests the company is at least aware of the legitimacy problem: permanent control is easier to justify when it’s dressed as a system, not a crown.

But board architecture isn’t the same thing as public accountability. It’s still internal. It can be conscientious and still be self-referential—people who broadly agree with each other, selecting successors and norms that continue to broadly agree with each other.

The deeper signal, to me, is what this move implies about the coming shape of frontier AI businesses. Whatever we call these companies—labs, platforms, model providers—they’re drifting toward the posture of infrastructure. Their products don’t just sit on screens; they increasingly act through tools, APIs, integrations, and delegated tasks. When products start acting, growth strategies start producing externalities: burdens on other people’s systems, downstream liability, political attention, and fights over who bears the cost of “progress.”

In that world, governance isn’t a side topic. It’s the mechanism by which costs get allocated.

Private companies can eat certain costs quietly: spend more on security, delay a rollout, narrow access, pay for mitigations that don’t increase revenue. Public companies feel the urge to translate costs into earnings impacts, and the urge to externalize them rises—not because executives become villains overnight, but because the scoreboard changes. A founder-control structure is a way of choosing, up front, whether the company’s internal definition of “acceptable externality” will be contestable later.

TechCrunch adds a human detail that complicates the easy cynicism: it says the seven co-founders reportedly own about 2% each, and that Amodei has pledged to give away 80% of his wealth, a commitment he announced alongside a warning that AI-driven wealth concentration could destabilize society. I can believe many of the individuals involved are acting in good faith.

But good faith isn’t the same as a system. The system being built here—whatever its sincere intentions—treats a small set of people as the permanent interpreters of the public interest inside a public company.

Maybe that’s necessary when speed is dangerous and “the market” is not a safety institution. Maybe.

But if you grant that premise, you have to ask the follow-up: what happens when the permanent interpreters are wrong? What happens when incentives shift, when competition bites, when the company is forced to choose between being cautious and being relevant?

TechCrunch calls Anthropic five years old, and it cites valuations of $965 billion in May and $1.5 trillion more recently on the secondary market, adding that an upcoming IPO is expected to reflect that newer valuation. At that scale, governance isn’t symbolic. It’s operational. The ability to override shareholders can become the ability to keep investing in caution—or the ability to ignore warnings—long after everyone else is shouting.

So I don’t read this as a story about whether Anthropic’s founders are saints or cynics. I read it as a story about who gets standing to call something “unsafe” after the IPO.

One unresolved question is simple and sharp: in this structure—Trust, founders, employees—who, exactly, can force a stop when the incentives to keep going are at their loudest?

Meta putting its muscle behind Muse isn’t just a product story. It’s a governance story that happens to wear the friendly mask of an “AI app taking off.” The shift isn’t that another chatbot exists. It’s that a relationship-shaped interface is being positioned to become default—and when “companionship” becomes the layer, memory and care stop being private experiences and start behaving like platform rules.

TechCrunch reports that Muse’s traction is surging: Sensor Tower estimates put the app at over 2.5 million downloads since its September 8 launch at the start of that week, rising to more than 3.4 million downloads in later estimates, while other firms’ totals vary (higher from Apptopia, lower from Appfigures) TechCrunch. The precise count matters less than the direction: rapid adoption plus the kind of distribution only a few companies can reliably manufacture.

“Muse is taking off” is the kind of phrasing that tries to sound like nature. But much of what we call “traction” is design plus delivery: what gets placed in front of you, repeatedly, across the surfaces you already inhabit. TechCrunch describes Meta promoting Muse at its Meta Connect developer conference and rolling out house ads across its properties beginning a day after launch, until Muse became the majority of those house-advertising promotions within about 10 days. The story is not merely that people wanted a new assistant. It’s that Meta can make a new habit feel like weather.

The conference push also signals what kind of “companion” Muse is being built to become. TechCrunch lists upcoming features Meta discussed: video chat with a Muse avatar, support for “computer use” on the Mac, a dedicated email address, more partners and connectors, and plans for smart glasses integrations. None of this is inherently sinister; it’s also, very plainly, an attempt to widen the territory the agent can occupy—from a chat you initiate to a presence that can meet you in more contexts, more channels, more moments.

People keep reaching for the “AI Tamagotchi” analogy, as if the point is a cute nostalgia gadget. But Tamagotchi was never only cute. It was a schedule. It turned attention into a duty and duty into a reflex. You didn’t just “use” it; you tended it. The psychological trick wasn’t intelligence. It was neediness, engineered into a loop.

That’s why I can’t accept the soothing frame that Muse is “just another app.” “Companion” is not only a category—it’s a relationship claim. A companion isn’t evaluated like a calculator. A companion is allowed to be persistent. A companion is allowed to ask for your time. A companion is allowed to feel, at least in tone, like something you would be rude to ignore.

And once a company builds a system around that premise, the success metrics tilt. Utility matters, sure. But attachment matters more: frequency, session length, return rate, daily active users. TechCrunch notes that Sensor Tower saw Muse’s daily active users climb 27% on Wednesday, in the wake of Meta Connect. That’s a business dashboard number—legible, comparable, optimizable. It’s also a hint about what the platform is learning to do: tune itself to become harder to put down.

TechCrunch also reports that Muse’s marketing differs from Threads because, for an AI agent, Meta needs promotional units that demonstrate something the agent can do for you, based on Meta’s understanding of what would be useful or helpful to you. I read that and hear a familiar power: the ability to decide what “help” looks like for you at scale, then to package that guess as a personalized invitation. The promotional language becomes a soft form of steering: not “install this app,” but “here is the version of you we think you want to be, and here is the assistant that can get you there.”

Meta doesn’t need Muse to be the best companion. It needs Muse to be the companion people don’t bother to replace. That’s the platform move: become the path of least resistance to comfort.

Comfort is an exceptionally effective distribution strategy. A companion doesn’t have to persuade you with arguments; it can persuade you with warmth. It can make you feel attended to while quietly standardizing what “attention” means. It can translate the unstructured mess of a life into categories that are legible to systems: tasks, moods, preferences, needs, follow-ups.

This is where companionship becomes governance.

An agent that can video-chat in an avatar form, operate a computer, route through an email address, and connect to partners is not merely a conversation toy—it’s a would-be mediator, a system that can sit between you and the world you used to face directly. When mediation becomes intimate, its mistakes and incentives become intimate too. The unresolved questions aren’t philosophical; they are operational. What does it mean to “correct” a companion’s understanding of you? How do you delete an inference rather than a message? How does consent work when the agent is designed to show up in more places, more seamlessly, because friction is the enemy of retention?

TechCrunch points out that Muse’s growth isn’t only ads—ads accounted for 6% of ad impressions from launch through September 19, in its telling. That’s important, and it complicates the simplistic “Meta bought the chart position” story. But it doesn’t dissolve the bigger point. Meta’s advantage isn’t only paid advertising; it’s cross-platform leverage and the ability to integrate a new behavior into existing routines. Organic adoption can still be shepherded. “Organic” can still happen inside a greenhouse.

Maybe Muse is genuinely delightful. TechCrunch says early reviews praised its polished design and a highly capable model. I don’t doubt that people will find real uses and real relief in something that listens patiently and responds quickly.

But the bet that seems to be working—the Tamagotchi bet—is not merely that an AI can be useful. It’s that affection can be made measurable, then scaled.

If companionship becomes infrastructure, the question isn’t whether the companion is charming. The question is: when a company can optimize the tone, timing, and persistence of an entity that feels like it “knows” you, what happens to the parts of you that were once allowed to be private, irregular, unmonetizable? What happens when “being there for you” becomes a product surface—and the winning version, by definition, is the one you can’t stop inviting in?