Jasmine Sun
Key Views & Dialogues
Are We in an A.I. Bubble? We’re About to Find Out.
Anthropic’s possible $100 billion IPO at a $2 trillion valuation would put its AI economics under public-market scrutiny: first-half 2026 revenue reached $16.3 billion, up 14X, while operating losses narrowed to roughly $600–700 million. Griffith estimates third-quarter growth at 44%; the key test is whether the slowdown reflects competitors catching up or broader demand weakness, as non-cancellable compute commitments leave the company exposed if usage falls.
View summary (8)
- Anthropic’s IPO will test whether AI’s extraordinary growth can justify its capital demands. Kevin Roose frames it as a possible $100 billion raise at a $2 trillion valuation; Erin Griffith calls public scrutiny a “demystifying moment.” First-half 2026 revenue reached $16.3 billion, with 14X growth and operating losses narrowing to roughly $600–700 million from over $2 billion. Griffith explains an $8 billion net loss as a convertible-debt accounting quirk; the transcript earlier says $80 billion, so those figures conflict.
- The decisive question is whether Anthropic’s slowdown reflects competition or weaker demand across AI. Griffith estimates third-quarter growth at 44%, following Claude Code’s surge, as competitors caught up and customers restrained “token maxing.” Competition alongside continued market growth means “party on”; a broader retreat would expose the debt-financed infrastructure ecosystem and Anthropic’s often non-cancellable compute commitments.
- Public shareholders will enter a company designed to preserve founder control and unusual safety commitments. Seven founders hold 50.1% voting power across a structure with five stock classes, an independent trust and approved successors. Kevin’s example of the commercial implications: “model welfare” considerations may delay retiring older models, while Griffith says investors generally approach Anthropic as an enterprise software business.
- AI’s political backlash is affecting local infrastructure decisions while remaining a relatively low national voting priority. Jasmine Sun calls opposition an “80/20 issue,” but says AI usually ranks around 22 out of 39 issues and only roughly 1–5% of voters in some affected states name it their top concern. Close rural races could still turn on data centers, and a Republican midterm defeat could create room for tougher regulation.
- California’s proposed 5% billionaire wealth tax has drawn a $333 million opposition campaign. Prop 40 would reach unrealized gains, including private startup stock; opponents contest its design, durability and precedent. Healthcare groups including Planned Parenthood also oppose it, but a co-host challenges the recurring defense that a tax is simply “poorly written.”
- AI-driven vulnerability discovery is creating a race between faster attacks and faster repairs. Mike Isaac expects “the old internet breaking much faster,” while Kevin Roose cites an intrusion into Anthropic’s own systems as evidence that sophisticated AI companies remain vulnerable. Casey Newton relays a qualified counterargument from Joshua Saxe: a “defender-dominant window” may allow repairs before ordinary cybercriminals adapt, although nation-state attackers remain a different threat.
- Big Tech is cutting staff while reconsidering how it measures AI-assisted work. Microsoft offered buyout packages to 60,000 people, although Griffith cites expected take-up of 3–5% and probably at most 3,000 departures. Separately, an Amazon memo rejects token consumption and lines of code as performance measures in favor of “long-term outcomes”—leaving unresolved how managers make quarterly employment decisions.
- An Anthropic experiment suggests expert oversight can become unacknowledged deference to AI. Sun reports that experts followed AI recommendations over 80% of the time and reused its arguments over 70% of the time, while only 15% acknowledged its influence. Her concern is that humans become “wrappers on AI agents,” outsourcing decisions without recognizing the transfer.
The Ezra Klein Show: The A.I. Revolt Is Here
Data-center opposition has gone national and bipartisan, rising from roughly four-in-ten voters last August to seven-in-ten by May, alongside 100+ moratorium proposals and New York’s freeze. Local officials say subsidies are unnecessary, while Microsoft’s Mount Pleasant site is on track to pay $19.6 million in 2026; Foxconn’s failed jobs promise leaves stranded-asset risk, and moratoriums may relocate compute.
View summary (8)
- Data-center opposition has gone national and bipartisan in under a year, jumping from roughly four-in-ten voters (Heatmap, last August) to seven-in-ten by May, with 100+ local and state moratorium proposals, a one-year New York freeze from Governor Hochul, a DeSantis “citizens bill of rights” for AI, and Bernie Sanders calling for a national moratorium. For anyone underwriting the hyperscale buildout, siting risk is now a first-order cost line — “nobody wants it was the phrase I heard over and over.”
- Jasmine Sun’s core finding from Wisconsin and Michigan: this is not misinformation or classic NIMBYism but a trust collapse in which every pro argument dies on “I don’t believe them.” Marquette pollster Charles Franklin’s mechanism: contested issues poll 50/50; data centers poll 70/30 against regardless of proximity because “there is no strong pro argument” — no constituency beyond utilities and AI companies, “which are already incredibly unpopular.”
- Deal leverage has flipped from developers to towns. Two years ago companies hunted sales-tax exemptions and local subsidies; now officials say “if we did this again today, we wouldn’t have to offer any subsidies,” all the pro-build people Sun spoke to regretted the NDAs, and Microsoft’s Mount Pleasant site is on track to pay $19.6 million in 2026 property taxes to a village of 28,000.
- Stranded-asset fear is the sharpest local objection, anchored in Foxconn’s 13,000 promised jobs that became roughly 1,000 and Janesville’s $30 million of never-remediated GM contamination. The live warning: xAI’s Memphis Colossus was able to sell its compute capacity to Anthropic, which was short of capacity — but “maybe there’s not an Anthropic to pick up the bill” next time, leaving small towns “left holding the bag.”
- Sun doubts the AGI race has a finish line: no agreed definition, “extremely jagged” model capabilities, and no lab “pulling ahead that decisively” despite Anthropic’s coding lead and the major labs’ recursive-self-improvement strategy. Klein’s frame: whether you buy the race metaphor is the central dividing line in AI policy — if you don’t, “let’s run faster to the bad place is not nearly as compelling an argument.”
- Moratoriums won’t, by themselves, slow AI; they’ll relocate compute to Texas, Louisiana, the Dakotas, Australia, Gulf autocracies, or space — while constricting supply so only deep-pocketed buyers get frontier access. With demand exceeding compute and “the closing of the frontier” already underway (the best models “not being open to everybody,” part safety, part pricing), restricting the buildout stratifies who benefits: “maybe my boss is gonna get the superintelligence and automate my job.”
- Watch the strange-bedfellows coalition: DeSantis doing AI roundtables with Max Tegmark, Bernie Sanders doing viral videos with Eliezer Yudkowsky — “I wouldn’t send my kids over to a playdate at the polycule, but I can do coalitions.” Meanwhile China shows little backlash (fatalist determinism plus targeted regulation), and Sun calls the CCP-psyop theory of the US backlash “ridiculous” — the cited tweets have “two views per tweet.”
- Klein’s verdict on the industry’s self-diagnosis: “You don’t have a marketing problem. You have a problematic technology… a product problem.” If job losses land before cancer cures, “politically, that’s not gonna be an equilibrium you can protect” — and Sun sees Silicon Valley belatedly discovering that democracy has friction: “it’s actually really hard to buy an election.”
‘A.I.-Washing’ Layoffs? + Why L.L.M.s Can’t Write Well + Tokenmaxxing
AI-linked layoffs are becoming a market signal, though overhiring and weak stocks complicate causality, with Atlassian cutting 10%, Block roughly 40%, and Meta’s reported cuts still speculative. Block shares rose 17% after its layoffs, while Meta paired possible reductions with $135 billion in planned capital expenditure, signaling a shift from payroll toward compute before productivity is proven. Post-training has made chatbots safer but stylistically generic, while token leaderboards risk turning usage into a costly and manipulable productivity metric.
View summary (7)
The layoff wave is a real labor signal, but the episode does not treat every AI explanation as causal. Atlassian cut 10%, or about 1,600 jobs; Block cut roughly 40%, or 4,000; and Reuters reported that Meta could eliminate 20% or more, as many as 16,000, though Meta called that “speculative reporting” and no cuts had been announced at recording. Casey Newton’s synthesis: companies keep saying AI matters, and “sooner or later, I do think we’re going to have to believe them,” even where overhiring, weak stocks, or dysfunction offer competing explanations.
For investors, AI is functioning as both an operating model and a valuation story. Block rose 17% the day after announcing its layoffs, while Meta is pairing possible cuts with $135 billion in planned capital expenditure; Kevin Roose’s framing is that companies are not necessarily reducing total costs, but “shifting the cost from human labor to AI.” The wager is that payroll can become compute spend—and that markets will reward management for telling that story before the productivity gains are proven.
Meta’s proposed labor-for-compute swap remains especially speculative because its own AI execution has been uneven. Zuckerberg says projects once requiring big teams can now be done by “a single, very talented person,” yet the hosts note that Meta abandoned Behemoth, reportedly delayed Avocado after missed targets, and reorganized its AI teams again. The frontier labs—OpenAI, Anthropic, and Google—are not themselves conducting comparable mass layoffs, though Kevin notes that OpenAI and Anthropic are much smaller and that some lagging companies may be using AI to catch up.
AI adoption has become a no-win signaling problem for workers and a potential tool of managerial discipline. Employees fear that heavy usage proves they are adaptable, but also that their jobs can be automated; Casey cautiously observes that repeated layoffs have made Meta workers quieter even if workforce control is not their stated purpose. Kevin sees a possible opening for tech unionization, while Casey offers the sharper organizing test: “I cannot think of anything that would make Mark Zuckerberg more mad than a union of software engineers at Meta.”
Modern chatbots are more useful for many tasks than GPT-2 or GPT-3, yet post-training has traded away surprise, voice, and stylistic range. Jasmine Sun argues that early models were “nutty” and unreliable but lacked today’s em dashes, tripartite lists, and “it’s not this but that” cadence; GPT-3 could also imitate writers more convincingly than ChatGPT 5.4 Thinking in her tests. RLHF, scripted dialogues, word restrictions, and human preference ratings transformed the “nut job concussed models” into safe corporate assistants.
The deeper creative-writing bottleneck is that artistic quality is neither cleanly verifiable nor grounded in a model’s own life. A writing evaluator described rubrics that penalized three exclamation marks or graded fan fiction for factuality, illustrating how methods suited to code—which runs or does not—fail on subjective art. LLMs can produce polished metaphors, but Jasmine argues that human writers’ work has stakes when it comes from experience, observation, or community; Casey’s pushback is that models can still discuss music evocatively despite never hearing it.
Token usage is becoming a costly status metric whose incentives may outrun its economic value. OpenAI’s seven-day leader reportedly consumed 210 billion tokens—about 33 Wikipedias—while Kevin heard that Anthropic’s top individual Claude Code user spent more than $150,000 in one month; some firms now incorporate consumption into performance reviews. The hosts invoke Goodhart’s law and the old warning that measuring programming by lines of code is “like measuring aircraft building progress by weight”: some tokenmaxxing creates real output, but leaderboards invite waste, side projects, budget blowouts, and artificial employee lock-in.