Pope vs AI, Anthropic's Digital God, AI Job Loss Narrative Flips, Open Source Wars

Key points

• **Pope Leo XIV's encyclical 'Magnifica Humanitas' (235 pages, 42,000 words) is the AI-regulation Schelling point.** Core argument: *'technology is never neutral, and technology takes on the characteristics of those who build, finance, and control it.'* Calls for worker-retraining, child safety guardrails, a ban on autonomous weapons (Skynet rule). Google/Amazon/Meta lobbied the Vatican April 29 to soften it; he refused. Mirrors Leo XIII's 1891 anti-industrial-revolution encyclical — Gurley's rebuttal: *'From 1891 to today the work week went 60→34 hours, real wages up 8-10x, global poverty 75%→under 10% — he got the whole thing precisely wrong.'*
• **Sacks' 'who guards the guardians' frame: the centralisation risk is real but government is the likely culprit.** *'I get very worried about... if some government agency can give notes to the model developers and they start telling them your definition of safety is not expansive enough.'* Cites the social-media trust-and-safety expansion (microaggressions, disinformation, transphobia) as the template. Prefers antitrust used aggressively as a check vs an FDA-for-AI. **Direct continuation of the centralisation-vs-decentralisation lens and [Pax Silica supply-chain framing from Issue 05 Helberg](/issues/2026-05-17).**
• **Gurley's 'Dr Frankenstein theory' of Anthropic is the sharpest single take of the week.** *'I've literally in the past 30 days read everything I can about Anthropic... I don't think they think they're writing software. I think they're midwifing a deity here. And I don't know which one I'm more afraid of, the regulatory capture or this second theory.'* Built on Dario's 'Machines of Loving Grace' (*'a capitalist economy of AI systems which then give out resources to humans based on some secondary economy of what the AI systems think makes sense to reward'*) + Chris Olah's 80-page constitution + Amanda Askell as chief philosopher. **Counterpoint to the bullish Anthropic-EBIT-positive framing in [Issue 06 Gavin Baker / All-In](/issues/2026-05-24)** — same company, opposite lens.
• **Chamath's game-theory overlay on the doom messaging: it's GTO, not (only) dogma.** *'If you want to be unexploitable... get three or four entities in a room, close the door, and dominate those other three or four entities. You set the rules. Your counterparty is unable to track at the level of technical capability you have.'* And: *'if the refs don't understand the game, you'll run over the game.'* Plus the halo effect — *'if you polled the intellectual elite on who's most caring, they'd probably put Anthropic first because they've been out with the doomerism talk.'*
• **'There is no single best model anymore' — Rogo's financial-analyst eval is the commoditisation receipt.** Chamath quoting the paper: *'At the top of the leaderboard Opus 4.7, GPT-5.5, Sonnet 4.6 appear almost indistinguishable, separated by less than 3/10 of a percentage point. Read superficially, the results suggest convergence.'* Raises the central investment question: *'what's the ROI on all this incremental [training] spend?'* if evals asymptote. **Direct continuation of [Gavin's Pareto-frontier-convergence framing in Issue 06](/issues/2026-05-24)** — but now framed as a margin/ROI risk to the frontier labs, not just a competitive note.
• **Enterprise AI reality check — the unwind is the story, not the adoption.** Chamath/Gurley: a Fortune-20 CEO asked for $1B in AI opex savings; six months in, *'the team has spent $200 million on tokens with minimal results'* (Vivek Garipalli tweet). **Microsoft announced it's killing its Claude licenses.** Kirkland & Ellis is spending ~$500M to roll its own frontier model. Counter-signal on Claude quality: *'Claude for Excel is better than Copilot, not by a little, by a lot'* (Gurley); Jason hit his Claude Pro token limit and paid another couple thousand dollars. **Control-plane / hot-swap architecture is the enterprise answer** — 8090 (Chamath) + Abacus (incubated by All-In) build headless products so the Fortune 1000 isn't locked to one frontier lab.
• **Sacks' open-source-ban thesis: Washington is laying 'predicate facts' to justify banning open-weight models.** *'You look at a lot of the rhetoric around how models need guardrails and that with open-source models the guardrails can be removed... they're trying to put predicate facts in the public record to justify an action later on.'* The paradox: *'our adversary, the Chinese, of all people... is leading the open source [open-weight] movement and the United States is centralizing'* (Jason). If the US bans open weights, cloud providers stop serving them and the rest of the world leapfrogs.
• **Apple as the intelligence-sovereignty dark horse.** Jason: *'intelligent sovereignty is different than privacy... you can't tell me what to think, you can't use your AI to analyze my photos, my emails, my messages and tell me how to interpret the world.'* The hardware enabler: M5, 48-128GB, the rumoured Mac Studio with a terabyte — enough to run open-weight SMLs locally. **First real Apple-AI-thesis surface in the digest — worth tracking against the consumer-device AI cohort.**
• **Elon's training-complex rewrite in C — the cost-of-training-is-collapsing receipt.** Chamath: *'Elon was like, we've rewritten the entire training complex in C and it's an order of magnitude increase and we can run it on 220,000 GPUs.'* The math: *'every 1% [efficiency] equals hundreds of millions in compute.'* Pushes toward *'$10 million training runs'* vs *'the big $10 billion training runs.'* **Continues the domain-specific-architecture + cheaper-training thesis from [Issue 06 Dwarkesh / Reiner Pope + Gavin TerraFab](/issues/2026-05-24).**
• **The job-loss vibe shift + 'AI washing' as securities-fraud risk.** Sacks: *'everyone's coming around to the idea that the job apocalypse is massively overblown'* — Sam + Dario walking it back, Goldman CEO, explosion in job postings. A securities-litigation partner (Donnie King, Akerman) is warning clients that attributing non-performance to AI could be **securities fraud via puffery** — Wix's May-28 layoff memo cited as a candidate. Gurley's structural counter to the margin-windfall thesis: *'there's no scenario where you just do more for less and everyone has 70% operating margins. Someone else comes along and does more for less and lowers the price'* — a productivity boom, not a profit windfall.

Notable quotes

Themes

• AI regulation & centralisation of power
• Anthropic safety-narrative as strategy
• Model commoditisation & training ROI
• Enterprise AI adoption reality vs hype
• Open-source / sovereignty / Apple