The Vatican’s Stance on AI, The Mass Tech Layoffs, and China’s 2027 Moon Flyby | EP #259
Peter Diamandis, Dave Blundin, Alex Wissner-Gross, and Salim Ismail dig into Pope Leo XIV's 42,000-word AI encyclical, debating whether it reflects genuine Vatican doctrine or Anthropic-influenced messaging, and what its stance against AI personhood means set against Buddhist orders in South Korea ordaining AI monks. They cover a killed White House AI-regulation executive order, a new DeepSWE coding benchmark showing GPT-5.5 (70%) far ahead of Claude Opus 4.7 (54%) and everyone else, and the Jevons-paradox collapse in AI token prices alongside exploding demand and revenue (OpenAI's $5.7B quarter, a forecast that Anthropic could out-earn Alphabet by 2028). The panel reconciles conflicting job-market signals -- mass tech layoffs versus Sam Altman's walked-back 'job apocalypse' warning -- with Dallas Fed data suggesting a hiring freeze for the young rather than mass automation, alongside a boom in solo-founder startups. The back half turns to space: SpaceX's record Starship V3 test flight, speculation about a Tesla-SpaceX merger, falling launch costs enabling a private circumlunar Mars flyby, Starlink's lunar-internet ambitions, and NASA's warning that China may beat the US back to the Moon with a crewed 2027 flyby.
Pope Leo XIV's AI encyclical and the Vatican's anti-personhood stanceAI▶ 3:05
Pope Leo XIV released a 42,000-word encyclical ('Magnificate Humanitus') calling for AI regulation, worker protections, and bans on autonomous weapons, coining the term 'Babel syndrome' for data/profit excess. The panel debates reports that Google, Anthropic, Meta, and OpenAI lobbied the Vatican beforehand, and that Anthropic in particular may have shaped language framing AIs as 'grown' rather than 'built', even though the document is unambiguous that AIs are not moral persons -- the opposite of Anthropic's own model 'soul document' framing.
The panel contrasts the Vatican's rejection of AI personhood with Buddhist orders in South Korea ordaining embodied AIs as monks, and debates whether 'religious LLMs' built on absolute truths are even possible, versus AI models whose beliefs are effectively 'hardwired' during pre-training in a similar way to childhood religious indoctrination.
White House AI regulation executive order collapsesGeopolitics▶ 21:42
A planned executive order requiring a 90-day government review before AI model releases was pulled by Trump hours before a signing ceremony after pushback from Elon Musk, Mark Zuckerberg, and David Sacks, who called it a slippery slope to mandatory licensing. The panel compares it to the 1975 Asilomar biotech self-regulation model and argues committee-based regulation can't keep pace with monthly frontier-model release cycles.
DeepSWE benchmark exposes a coding capability cliffCompute▶ 27:49
A new benchmark from startup Data Curve, testing real-world tasks requiring edits of 668 lines of code across seven files, shows GPT-5.5 scoring 70% versus Claude Opus 4.7's 54%, with Gemini, Kimi, and DeepSeek all below 32%. Panel notes Opus 4.7 uses roughly twice the tokens of GPT-5.5 for similar results, and predicts the benchmark will saturate within months like its predecessors.
AI token prices have fallen roughly 75% since late 2024 while monthly token consumption exploded from near zero to 25 trillion, a textbook Jevons paradox. The panel debates whether tokens are even the right unit to measure 'abundance of intelligence' given how much a token's meaning varies by model and tokenizer.
AI revenue explosion at OpenAI and AnthropicEconomy▶ 42:04
OpenAI posted $5.7 billion in quarterly revenue with ChatGPT's coding agent Codex reaching 2 million users, while OSS Capital's Joseph Jax projects Anthropic could surpass Alphabet's total revenue by 2028, growing from $9 billion to potentially $2 trillion by 2030. The panel discusses how usage-based, throttleable pricing makes AI subscriptions an unusually 'slippery' product for enterprises to budget around.
The four-lab race and Ilya Sutskever's rumored hedge fundAI▶ 46:52
The panel debates whether any dark horse can catch OpenAI, Anthropic, Google DeepMind, and xAI, discussing reports that Ilya Sutskever's Safe Superintelligence is building a proprietary trading hedge fund (likely to self-fund compute for recursive self-improvement) and whether a breakthrough 'perfect algorithm' could let a newcomer leapfrog the incumbents once compute becomes the only remaining bottleneck.
DeepMind's Green Tree AI reaches superforecaster parityAI▶ 51:35
DeepMind's 'Green Tree' system reportedly matched the accuracy of human superforecasters (the top 2% of predictors, per Philip Tetlock, historically 30% more accurate than CIA analysts) for the first time on March 15th, largely by assimilating unstructured data (research reports, news) at a scale no human analyst can match. Panel discusses implications for finance, insurance, and governance, plus a tangent into forecasting-implies-retrodiction and Nick Bostrom's simulation hypothesis.
Tech layoffs versus Sam Altman's job-apocalypse walkbackEconomy▶ 58:54
134,000 tech workers have been laid off in 2026 so far, and Mercer reports 99% of CEOs expect AI-driven layoffs within two years, but Jensen Huang calls AI a 'lazy narrative' CEOs use to cover bad strategy. Sam Altman walked back his own prior warnings of white-collar job apocalypse; a Dallas Fed report shows AI-exposure-correlated employment decline only among younger workers, suggesting a hiring freeze rather than mass layoffs.
Solopreneur explosion and the organizational singularityEconomy▶ 1:03:04
US startup formation is up 25% year-over-year with six times more startups than Europe, and a16z data shows AI solo founders doubling quarter-over-quarter to 3,000. The panel argues small teams will keep outperforming large, coordination-heavy organizations, while noting most 'solo founders' are actually embedded in dense, connected ecosystems (accelerators, hackathons like the Gemini XPRIZE) rather than working in isolation.
Education system criticized as obsolete for an AI economyOther▶ 1:13:51
Peter launches a listener survey (moonshots.com/servey) on whether education is preparing students, and the panel argues schooling remains 'supply side' -- training a fixed skill for a job market -- when it needs to flip to 'demand side': identifying a problem to solve and acquiring whatever skills are needed, especially as skill half-life has dropped from roughly 30 years to about 3.
SpaceX Starship V3, the Tesla-SpaceX merger, and the new lunar space raceSpace▶ 1:18:48
SpaceX's Starship V3 flew for the first time from a new Texas launch site, carrying 97,000 lbs to near orbit on upgraded Raptor 3 engines. The panel discusses a Kalshi market pricing 50/50 odds of a Tesla-SpaceX merger within a year (Peter says 100%), falling launch costs enabling a private circumlunar Mars flyby by Bitcoin miner Chun Wang, Starlink's lunar-internet ambitions, and NASA administrator Jared Isaacman's warning that China could fly a crewed lunar flyby before the US returns in 2027-2028.
Predictions made
openAlex Wissner-Gross: The DeepSWE coding benchmark's current spread between frontier models will saturate, just like prior coding benchmarks.
“This too will saturate... probably in the next few months.”
Your call:
openAlex Wissner-Gross: The current White House administration will remain the one supervising AI regulation for the foreseeable future, regardless of the next presidential election.
EP #? · · due: unspecified (through at least the next election cycle) · ▶ watch
“the executive that we have right now is the executive supervising AI regulation potentially for the future.”
“2035, I think, uh, the price and my ability to afford it will intersect.”
Your call:
openDave Blundin: The window of AI-driven job loss and social disruption will turn a corner as abundance-driven new job creation catches up and overtakes it.
“we were always predicting that by 2030 it would turn the corner because the abundance created by all this AI is going to create massive massive new gains.”
Your call:
Numbers that matter
Pope Leo XIV's encyclical on AI runs 42,000 wordsFirst papal encyclical devoted to AI, titled (per transcript) 'Magnificate Humanitus'.
1.4 billion Catholics worldwideSize of the audience the Vatican's AI-personhood stance is aimed at.
GPT-5.5 scored 70% on the new DeepSWE coding benchmarkSolving 7 of 10 hard, real-world software engineering tasks fully autonomously.
Claude Opus 4.7 scored 54% on DeepSWEA 16-point gap behind GPT-5.5 on the same benchmark.
Gemini, Kimi, and DeepSeek all scored below 32% on DeepSWEDescribed as a 'massive cliff' versus the top two models.
DeepSWE tasks require editing 668 lines of code across 7 filesIllustrates why the benchmark is harder to game than prior, saturated coding benchmarks.
Claude Opus 4.7 uses about 2x the tokens of GPT-5.5 for the same or a slightly worse resultRaises Anthropic's effective revenue per user on usage-based pricing.
AI token prices fell about 75%, from roughly $2 to $0.50 per million tokens since late 2024Cited by Peter Diamandis to illustrate Jevons paradox in AI pricing.
Alex Wissner-Gross cites a 3x price drop (from $1.50 to $0.50 per million tokens) alongside a 30-50x rise in usageA slightly different price-drop figure than Peter's, both illustrating the same Jevons-paradox dynamic.
Monthly AI token consumption grew from near zero to about 25 trillion tokens per monthDemand growth attributed to falling token prices; panel suspects true demand is understated because capacity is sold out.
Gartner forecasts inference on a trillion-parameter LLM will cost 90% less by 2030 than in 2025Cited as evidence intelligence access costs keep falling even as model power rises.
OpenAI generated $5.7 billion in a single quarterFramed alongside Codex reaching 2 million users as evidence of OpenAI's pivot from consumer to coding revenue.
OpenAI's Codex coding agent has 2 million usersBecoming a real revenue engine as OpenAI shifts from consumer chat to coding.
Anthropic revenue projected to grow from $9 billion to as much as $2 trillion by 2030, potentially surpassing Alphabet by 2028Projection from Joseph Jax of OSS Capital, called 'the fastest wealth creation ever in human history' if directionally correct.
134,000 tech workers laid off in the first five months of 2026March was called the worst month for tech layoffs since the pandemic.
Worth digging into
🕳️ Anthropic's alleged influence on the Vatican's AI encyclical
The panel claims Anthropic figures may have shaped how the encyclical frames AI as 'grown' rather than 'built', despite the document's opposite stance on AI personhood -- a genuinely surprising alliance worth verifying against primary sources.
🕳️ Ilya Sutskever's Safe Superintelligence reportedly building a hedge fund
A striking pivot for a company founded explicitly around AI safety; worth checking whether there's credible reporting beyond the panel's 'publicly reported rumor' framing, and how it connects to SSI's compute strategy.
🕳️ DeepMind's Green Tree superforecasting system
A claim of AI reaching parity with elite human superforecasters would have major implications for finance, insurance, and governance if verified, but no paper or methodology was named on the pod.
A jump from $9 billion to $2 trillion in revenue by 2030 is an extraordinary claim worth stress-testing against Anthropic's actual disclosed run-rate and growth curve.
🕳️ Data Curve's DeepSWE benchmark methodology
A brand-new coding benchmark claiming to fix the benchmark-saturation problem via hand-coded, real-world multi-file tasks; worth checking whether it holds up as newer models (like Opus 4.8, announced mid-episode) are tested against it.
🕳️ The physics of cheap orbital access
Peter's back-of-envelope claim that the raw energy cost to reach orbital velocity is about $200 (versus a $20 million ticket) is a striking abundance argument worth sanity-checking against real launch-system efficiency losses.