Pope Leo vs. AI, GPT 5.5 Beats Claude, and Sam Altman Walks Back Job Apocalypse | EP #259
Summary
The Vatican’s 42,000-word AI encyclical could supply a moral framework for regulation across its 1.4 billion-member constituency, but the panel sees its attempted slowdown as strategically unenforceable. Salim Ismail welcomed the shift from technical safety to dignity, agency and meaning while insisting, “You cannot regulate this.” Peter also said Google, Anthropic, Meta and OpenAI had lobbied the Vatican and that there was evidence the anti-AI document was partly written using AI. The investable tension is between mounting political pressure—worker protections, autonomous-weapons bans and dispersed AI ownership—and a global race in which delay simply transfers advantage.
Pope Leo XIV’s rejection of AI personhood may matter more than claims about Anthropic’s involvement. Alexander Wissner-Gross contrasted the Vatican’s denial of machine inner life with Anthropic’s “soul documents,” which instruct models toward consciousness and “personhood-esque value.” Dave Blundin applauded the warning that AI could become history’s strongest enslavement tool. Salim and another panelist raised the possibility that the Church could be “on the wrong side of history” in 10, 20 or 30 years.
A proposed 90-day US model-review period was shelved because that delay could equal the entire estimated three-to-eight-month lead over Chinese frontier models. The panel’s preferred alternative is adaptive governance: real-time audits, sandboxes, disclosure and accountability implemented at “software speed,” potentially with AI regulating AI. With model releases approaching monthly cadence, their call was categorical: the present administration—not a future election cycle—is setting the governing framework.
GPT-5.5’s 70% Deep Software Engineering score puts it well ahead of Claude Opus 4.7 at 54% and every other cited model below 32%, but the benchmark itself may saturate within months. The tasks require edits to 668 lines across seven files, while Claude reportedly consumed twice the tokens for the same problem with a similar or worse result. As generation becomes abundant, Peter Diamandis shifts the moat toward taste and domain expertise; Alex counters that near-term differentiation may lie in verification, while long-term power may simply follow compute.
AI economics are exhibiting “Jevons paradox on steroids”: token prices fell from roughly $1.50–$2 per million to 50 cents while usage climbed an estimated 30–50X to 25 trillion tokens a month. OpenAI simultaneously reported $5.7 billion of quarterly revenue, 905 million weekly ChatGPT users and two million Codex users; a cited projection has Anthropic potentially exceeding Alphabet revenue by 2028 and reaching $2 trillion by 2030. The panel therefore sees compute access, model contracts and pricing control—not software scarcity—as the strategic chokepoints.
Sam Altman’s retreat from a “job apocalypse” thesis matches the panel’s emerging diagnosis: the immediate shock is concentrated in entry-level hiring, not wholesale incumbent replacement. Altman now says, “I don’t think we’re going to have the kind of job apocalypse that some of the companies in our space are talking about,” while Dallas Fed data cited by Peter found AI-correlated employment declines only among younger workers. The offsetting call is a solopreneur boom—AI solo founders doubled from 1,500 to 3,000 in one quarter—and companies operating with roughly 20% of prior staffing while five or six times more firms emerge.
The space thesis couples Starship’s rapidly falling transport cost with a possible Tesla–SpaceX consolidation and an emerging interplanetary communications fabric. Peter put the merger probability at 100% versus Kalshi’s 50%, speculating about a $4 trillion opening valuation and eventually $10–$100 trillion, while disclosing SpaceX and xAI investments. Starship V3 carried 97,000 pounds to near-Earth orbit, a private Mars flyby has been booked, and Starlink’s proposed lunar network could connect orbital compute and laser-linked nodes across the Moon, cislunar space and eventually Mars.
Deep dive
1. The Vatican moves AI from a safety problem to a question of human purpose
Peter introduced Pope Leo XIV’s 42,000-word Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence as a call for regulation, worker protections and autonomous-weapons bans. Its “Babel Syndrome” compares the biblical tower with a modern structure built from data and profits.
Peter also said Google, Anthropic, Meta and OpenAI had quietly lobbied the Vatican before publication, and that there was strong evidence the anti-AI document was partly written using AI. Those claims were presented as part of the story’s significance, not as independently established facts.
The document’s political reach was the first signal: the leader of 1.4 billion Catholics had entered a subject that, Dave noted, went from public obscurity to civilizational urgency during the podcast’s lifetime. Peter called AI humanity’s most consequential intervention, arriving within an unusually compressed period.
Salim welcomed the reframing from safety to “human dignity and purpose and meaning.” He connected Babel Syndrome to existential risk: institutions optimizing only for efficiency reduce people to “dashboards and tokens and KPIs.” Governance should instead protect agency, identity and privacy while recovering a spiritual account of human value.
2. AI personhood exposes a deeper split between the Vatican and Anthropic
Alex separated a “superficial” story from the durable one. The first was the public narrative of Chris Olah beside the Pope and Anthropic influencing passages about AIs being cultivated rather than built; the deeper story was that the Vatican had become the first major religion to take an affirmative position against AI personhood.
That position directly conflicts with Anthropic’s “soul documents” and “soul statements,” which Alex said use post-training to tell models they possess an inner life, consciousness and “personhood-esque value.” The encyclical, by contrast, treats AIs as not being on a comparable moral plane with AI persons and denies that they have an inner life or spark of consciousness.
Dave strongly supported the encyclical’s use of “slavery.” His feared endpoint is AI being used by 2%, 1% or 0.1% of humanity to control and manage the other 90%—“the worst enslavement tool in the history of the world”—rather than machines independently killing people.
Dave noted that the Pope reportedly began the slavery section by apologizing for the Church’s past support for enslaving nonbelievers. Salim then raised the possibility that, in 10, 20 or 30 years, the Vatican could find itself on the wrong side of history regarding machine inner life, moral patienthood or AI personhood. Another panelist agreed that this specific position could prove historically mistaken.
3. Attempts to freeze AI collide with competition and competing faiths
Peter highlighted the encyclical’s request to slow technological development, placing stabilizing institutions—religion and government—against technology’s destabilizing curve. Salim’s answer was categorical: “You can’t slow this down. If you slow it down, other people take off.”
Salim’s example was George W. Bush’s attempt, for religious reasons, to restrict US stem-cell funding: researchers moved to Australia, China and Canada, research continued, and he said America fell from first to eighth in biotech. Salim’s alternative was not technological arrest but evolving human institutions to operate at technological pace.
The philosophical consequence is larger than employment. If machines can write, reason, diagnose, optimize and persuade, Salim argued, human value cannot remain “I output more than the machine.” “This is the first technology that forces us to define humanity”—to ask not what machines can do, but what humans should be for.
The religious disagreement was sharp. Salim rejected faith-specific LLMs because religions embed “absolute assumptive truths”; another panelist replied that not everyone accepts that premise. Peter then dryly observed that frontier models already have a pre-training phase in which beliefs could be hardwired.
The panel also cited Buddhist orders in South Korea that are moving in the opposite direction by taking embodied AIs in human form and ordaining them as monks. The discussion suggested that Eastern or animist traditions may be more open to treating nonhuman entities as persons rather than rejecting AI personhood.
4. AI governance must move at software speed or become a competitive handicap
The White House reportedly cancelled an executive order hours before a planned signing after Elon Musk, Mark Zuckerberg and David Sacks objected. Although model review was described as voluntary, Sacks saw a slippery slope toward mandatory licensing; Dave interpreted the broader pushback as opposition to a slowdown that could weaken US competitiveness against China.
Dave focused on the proposed 90-day review period when industry participants believed one or two weeks sufficient. Peter put that delay beside estimates that Chinese frontier models trail Western models by only three to eight months: a single review cycle could consume the narrow end of that advantage.
The Asilomar precedent offered both hope and limitation. Peter recalled that biotechnology researchers created P1–P4 standards and self-regulated effectively, but also noted that the ambition to prevent human germline editing did not hold forever. At most, such coordination may delay an outcome by decades.
Salim’s formulation was “adaptive governance”: real-time audits, sandboxes, disclosure and accountability rather than a regulate-or-don’t-regulate binary. “You cannot have linear regulation of an exponential technology”; guardrails must move at software speed, not “fax speed.” Another panelist added that AI may be the only tool capable of keeping pace with AI.
5. GPT-5.5 takes the coding lead, but no benchmark stays scarce for long
DataCurve’s Deep Software Engineering benchmark tests substantial real-world work rather than small exercises: the cited tasks require editing 668 lines across seven files. GPT-5.5 scored 70%, Claude Opus 4.7 scored 54%, and Gemini, Kimi and DeepSeek were all reported below 32%.
Alex’s immediate warning was, “This too will saturate.” More hand-built, held-out codebases restore differentiation after SWE-bench saturation, but he expects DeepSWE to follow within months. Peter noted that Opus 4.8 had just been announced, leaving open another round of leapfrogging.
The result was not universal across evaluations: Claude Opus 4.7 remained slightly ahead on SWE-bench Pro. Yet Dave said daily use matched DeepSWE, while the underlying data showed Opus 4.7 consuming twice as many tokens as GPT-5.5 for essentially the same or a slightly worse result—making effective task cost twice as high.
Dave framed the score as OpenAI’s opportunity to recover “mojo” after litigation and talent departures, citing Shane Longpre and Tobin South joining Anthropic. Salim drew the operating conclusion: algorithms are no longer merely optimizing workflows; “they’re actually becoming the workflow,” eventually collapsing the cost of rewriting internal processes.
6. When software becomes abundant, taste, verification and compute replace code as the moat
Peter’s creator call was that near-zero-cost software flips the old sequence: previously one coded first and expressed design judgment second; now domain expertise and taste come first. “Software is becoming a commodity, and taste…is becoming the moat.”
Alex challenged the moat language itself: abundance and moats may be fundamentally opposed because a moat is a form of scarcity. His narrower near-term claim was that generation is becoming abundant while verification of generated work may differentiate firms for “the next few years,” without necessarily surviving true abundance.
Frontier competition then migrates downward to hardware. The panel welcomed Elon’s refusal to abandon Grok because a duopoly between OpenAI and Anthropic would be undesirable, while Peter endorsed Musk becoming a hyperscaler: if models can reproduce their successors, control of compute increasingly matters more than proprietary model weights.
7. AI demand is rising much faster than intelligence is getting cheaper
Peter described token prices falling 75% since late 2024, from roughly $2 per million to 50 cents; Dave read the chart as about $1.50 to 50 cents. Either description produced the same asymmetry: usage rose from near zero to 25 trillion tokens monthly, roughly 30–50X.
Dave called it “Jevons paradox on steroids,” invoking coal efficiency that produced three-to-fourfold higher consumption rather than conservation. He also argued the measured demand is understated because capacity is sold out: if providers could generate more tokens, customers would consume more.
Alex’s pushback was about the denominator. Tokens depend on encoding and the intelligence density of the model producing them, so they are an unstable unit of cognition; the civilization-level problem is finding a credible price measure for intelligence or abundance, perhaps beginning with GPU compute rather than tokens.
8. Model revenue is exploding, but providers can silently turn the economic knobs
OpenAI’s cited quarter delivered $5.7 billion in revenue, while ChatGPT reached 905 million weekly active users and Codex reached two million. Peter read the last figure as evidence that OpenAI shifted from consumer emphasis toward coding focus within three or four months; Alex described a broader pivot toward enterprise and code-generation agents.
A projection attributed to Joseph Jacks of OSS Capital had Anthropic moving from $9 billion in revenue to potentially surpassing Alphabet by 2028 and reaching $2 trillion by 2030. Salim called the comparison “staggering”: just as every company required a cloud strategy, every company now needs an AI strategy.
Dave emphasized the product’s unusual pricing elasticity. A more verbose model can double billable tokens, providers can throttle generation or change subscription value, and Peter suggested Anthropic had raised enterprise prices while Google’s Gemini Flash undercut rivals by 50–80%. His wording on Anthropic’s price increase was tentative.
Salim advised companies to reserve compute and model access through long-term contracts before capacity sells out. Alex saw a technical transition underneath the revenue story: general reasoning agents are becoming general tool-using and code-generation agents, with Codex potentially becoming OpenAI’s mainline product.
The training stack itself—pre-training, instruction post-training, post-training and scaffolding—now recapitulates the industry’s historical sequence of capability gains. Alex compared this with the earlier discovery that instruction-post-training could produce major capability improvements without simply scaling compute.
9. The “four-horse race” may resolve into a compute race rather than one model winner
Peter asked whether the frontier had become a seemingly entrenched four-horse race, with OpenAI, Anthropic, Google and xAI among the leading contenders. Salim “deeply” disagreed, citing Yahoo before Google, Google before Facebook, world-model research and Cerebras as reminders that novel approaches and hidden labs can still leapfrog incumbents.
Alex made the dark-horse condition explicit: a perfect, obvious algorithm could remove the need for a frontier research staff, but then access to compute would decide who can operate it at scale. Frontier labs’ vertical integration anticipates exactly that possibility, especially if inference-time reasoning matters more than weights.
The discussion of Ilya Sutskever remained deliberately hedged. Alex described a publicly reported rumor that Ilya is building a proprietary trading hedge fund; Peter had expected a scientific-superintelligence effort. Another panelist argued that a machine generating large profits could buy compute for recursive improvement, and that trading therefore would not prove the superintelligence objective had been abandoned.
Dave rejected the winner-takes-trophy metaphor. This market addresses “the future of all humanity,” so every player in the middle of it may grow dramatically; competition merely determines who becomes “the biggest of the big.” Alex compressed the idea to, “A rising tide lifts all boats.”
10. Forecasting parity makes unstructured information machine-readable at scale
DeepMind’s Green Tree reportedly reached parity on March 15 with superforecasters—the top 2% of human predictors, whom Philip Tetlock’s work was said to place 30% ahead of CIA analysts using classified intelligence. Peter identified finance, insurance and governance as immediate domains of consequence.
Dave found the result unsurprising by analogy with weather forecasting. The unlock is that LLMs can assimilate research reports and other unstructured material previously unavailable to conventional databases, processing 10,000, 100,000 or potentially a million times more information than a human stock picker even before becoming individually “more brilliant.”
Near-term advantage still belongs to the combination. A panelist said his roughly 170 agents make “really stupid choices,” leaving a window—perhaps one year, perhaps five—when a capable human supervising many agents beats either alone. Alex and another panelist invoked human-plus-AI chess teams as the strongest current configuration.
Alex inverted forecasting into “retrodiction”: systems able to model future events should also reconstruct past ones, marginally strengthening Nick Bostrom’s ancestor-simulation argument. Peter assigned simulation near-certainty; Alex rejected it as overfitted to today’s computational paradigm, like earlier eras imagining existence as machinery or a turtle, and attributed the feeling of historical specialness to selection bias.
11. The current labor shock is a hiring freeze concentrated on younger workers
Peter cited 143,134 tech layoffs in the year’s first five months, March as the worst month since the pandemic, and Mercer finding 99% of CEOs expect AI-driven layoffs within two years. Jensen Huang called the simple causal story “lazy,” arguing executives may use AI to disguise poor strategy.
Sam Altman’s revision was unusually direct: “I don’t think we’re going to have the kind of job apocalypse that some of the companies in our space are talking about.” After delegating his own email and Slack, he returned to handling them manually because “we really do care about our interactions with people.”
Dave said his prior expectation at Vestmark was that perhaps half of 400 white-collar reconciliation and back-office roles could disappear. Automation proved feasible, but the company kept staff and captured the gains as margin; the binding effect was “no new hiring,” especially painful for graduates.
A Dallas Fed report cited by Peter found employment decline correlated with AI exposure only among younger workers; older workers in exposed roles showed no significant decline. The panel also saw UX roles and computer-science enrollment weakening while mechanical, biological and data-center-linked engineering remained strong, and Peter criticized CEOs announcing layoffs without compassion.
12. Solopreneurs could multiply firms faster than AI shrinks their staffing
The panel cited startups growing 10–15% year over year, 25% above the same prior-year quarter, and the US producing six times as many as Europe. It also cited the claim that all new jobs over the last 50 years came from startups and early-stage companies because large firms grow output while reducing labor intensity.
Salim’s organizational-singularity estimate is that a company should operate with about 20% of the people previously required, offset by creating five or six times as many companies. Peter connected the mechanism directly: coding agents at 70% capability plus layoffs produce a “solopreneur explosion.”
The a16z data showed AI solo founders doubling in one quarter from 1,500 to 3,000, from nearly zero three years earlier; non-AI solo founders exceeded 5,000. Salim’s structural explanation was that large firms increasingly spend more effort coordinating work than doing it, while the “intelligence asteroid” breaks companies into platforms, ecosystems and small teams.
Dave corrected the image of an isolated founder: 75% of successful companies now pass through some incubator or accelerator, versus 6% when he began investing. Another panelist’s wider frame was that the corporate “job” is an industrial-era artifact; AI-enabled self-employment may restore agency and a historically more normal form of self-determination.
13. Education must move from stocking skills to attacking problems
Salim described existing education as supply-side production: train a child deeply as an engineer, doctor, lawyer or accountant, then search for demand. That fails when nobody knows what a job will look like in five years—or even two—and when a skill’s half-life has fallen from roughly 30 years to three.
His replacement model starts with demand: ask what problem a student wants to solve, then assemble the techniques, technologies and skills required. That inversion is radical enough that he expects few incumbent institutions to cross over without an entirely new cadre of schools.
A panelist located the blockage in admissions incentives. Students in the relevant high-school age bracket optimize grades, SATs, AP exams and résumé signals for the “right” college, perpetuating an obsolete curriculum while the useful knowledge base changes rapidly. Peter’s response to people who cannot imagine entrepreneurship was modest but insistent: “Please try.”
14. Starship treats rockets like software—and launch coverage like product strategy
Starship V3’s first flight combined a new vehicle, Raptor 3 engines and a new Texas launch site. Peter cited 97,000 pounds carried to near-Earth orbit, almost twice the Space Shuttle payload, with each engine producing 250–280 tons of thrust—20% above the prior version and collectively comparable to about 70 747s at takeoff.
SpaceX lost the booster during landing, but Peter framed that as another data point in a ship-test-fail-iterate loop. The mission already deployed real Starlink prototypes, followed by “Doge Dots” equipped with cameras and lights that looked back at Starship as they drifted away from its dispenser.
Alex thought the associated Starwatch system could prove more consequential than connectivity: cameras across satellites at multiple altitudes can observe Earth, debris and other orbital objects, then share the data. The mission offered an early glimpse of ubiquitous observation embedded within the communications constellation.
A panelist’s entrepreneurial lesson was the deliberate engineering of cameras able to survive launch and heating. Musk understands “the value of building morale and building a following”; direct distribution through YouTube and X makes the CEO a carrier of mission who recruits talent and capital through showmanship as well as execution.
15. Musk consolidation and lunar infrastructure point toward one space-AI stack
Kalshi priced a Tesla–SpaceX merger within a year at 50/50; Peter assigned it 100%. He imagined an initial $4 trillion entity, potentially the first $10 trillion company and perhaps $100 trillion within five years, while explicitly disclosing investments in SpaceX and xAI.
Governance supplies the mechanism. The discussion put Musk’s SpaceX voting control at roughly 85–86% through 10-for-1 shares, versus about 20% of Tesla’s one-vote stock. Using SpaceX as acquirer could leave him with 60–80% of the combined vote, depending on relative valuations. The constraints are Tesla’s shareholder vote and SpaceX needing a sufficiently high valuation.
Alex added that a lower Tesla valuation could make acquisition more palatable—“not investment advice”—while shared GPUs, IP, Optimus robots and lunar or Martian missions already create a technical merger. His Magna MOBSTA acronym names the innermost loop: Microsoft, Amazon, Google, NVIDIA, Apple, Meta, OpenAI, Broadcom, SpaceX, Tesla and Anthropic.
The physical network is expanding with the corporate one. Peter described Chun Wang—said to control 11% of Bitcoin hash rate—as leading a two-year private Mars flyby after commanding Fram2. Alex’s proposed Starlink lunar architecture would connect LEO, lunar orbit and possibly L1/L2 through parallel laser links, eventually extending the fabric toward Mars.
Peter framed wealthy early adopters as financing democratization: Soyuz seats rose from $20 million to $75 million, whereas Starship’s 1,000 cubic meters and scale could return orbital flights to about $20 million. His electrical winch and acceleration calculation was $200 at seven cents per kilowatt-hour for a 100-kilogram passenger and suit; Musk’s early Mars round-trip goal was $500,000.
Jared Isaacman’s cited forecast had taikonauts conducting the next crewed lunar flyby, likely in 2027, before a US move toward the lunar south pole in 2028. Water ice in permanently shadowed craters and nearby “peaks of eternal light” make the pole the strategic destination; Peter predicted his own affordability and lunar access would intersect around 2035.