photo: Efrem Efre · Pexels
Signal & Noise · Episode 7

Anthropic at the Vatican.

2026-05-27 6-day window 2026-05-22 → 2026-05-27

Pope Leo XIV cites Anthropic in his AI encyclical.

Uber COO: no productivity gains.

Narayanan deflates Google's $916-OS claim. curl maintainers drowning in AI-bug reports.

01 / Summary

The window, distilled

Pope Leo XIV cites Anthropic in his AI encyclical. Uber COO: no productivity gains. Narayanan deflates Google's $916-OS claim. curl maintainers drowning in AI-bug reports.

02 / Headlines

What landed

  1. Pope Leo XIV's AI encyclical cites Anthropic
    Magnifica Humanitas — first papal encyclical on AI; Chris Olah present at the Vatican rollout; Anthropic Institute formalized as parallel research arm
  2. Uber COO: no proportional AI productivity gains
    Andrew Macdonald publicly admits Uber not seeing gains proportional to spend — Marcus calls it the bubble-pop indicator
  3. Cherny ships /usage + endorses 'use auto'
    Claude Code's observability primitive (breakdown by Skill/Agent/MCP/Plugin) ships; auto-mode declared the #1 tip
  4. curl maintainer drowning in AI security reports
    Daniel Stenberg: 4-5× more reports than 2024, average >1/day; AI-assisted vulnerability research breaks OSS maintenance economics
  5. Mollick: Choosing to Stay Human
    AI-detection arms race + 'feels like being lied to' — Mollick names which work to KEEP human; convergence with Chollet and Graham
03 / Threads

What moved across voices

The Vatican Moment — Anthropic's institutional legitimacy spike

Pope Leo XIVChris OlahBoris ChernyJack Clark

Pope Leo XIV's Magnifica Humanitas explicitly draws on Anthropic-style interpretability framing — the encyclical's name was chosen in dialogue with Leo XIII's 1891 Rerum novarum on capital and labor, casting AI as the labor question of this century. Chris Olah present at the Vatican; Cherny amplifies inside Anthropic. Anthropic Institute formally announced as parallel research arm led by Clark. The Builder reads it as mainstream legitimacy. The Skeptic reads it as institutional capture working both ways.

Productivity bubble fissures become audible

Gary MarcusArvind NarayananDaniel Stenberg via Simon Willison

Four independent voices on the same beat in one week. Marcus on Uber COO Andrew Macdonald admitting no proportional gains. Marcus on retirement-fund AI-bubble exposure. Narayanan deflates Google's $916-OS claim with independent-eval critique. Stenberg names curl-maintenance breakdown under AI-assisted security reports. The gap between productivity claims and productivity reality is now publicly auditable — and not from skeptics only.

AI-written prose as trust-destroyer

Paul Graham via WillisonArmin RonacherEthan MollickFrançois Chollet

Paul Graham: 'It feels like being lied to.' Ronacher: issue reports laundered through LLMs lose the human voice and become guesswork. Mollick ships Choosing to Stay Human — explicit framing of what to keep human. Chollet: 'Whenever AI tells me I'm absolutely right, my trust drops.' Recognizable AI writing now actively damages relationships. The Builder side has no clean answer; the only adopted position is Mollick's: choose.

Software 3.0 operationalized — Claude Code, Datasette Agent, FST

Andrej KarpathyBoris ChernySimon WillisonLakshya AgrawalErik Schluntz

Cherny ships /usage as harness observability; 'use auto' becomes the official #1 tip. Willison ships Datasette Agent + 3 plugins, ending a 3-year LLM-Datasette convergence. Schluntz's 'ask without derailing' feature lands. Agrawal et al. publish FST (fast-context + slow-weight) at arXiv — 3× sample efficient over RL, 70% less catastrophic forgetting. Karpathy's Sequoia transcript canonical: December 2025 as agentic inflection point; verifiability × training attention as capability rubric; LLM Wiki + LLM Council as named patterns.

04 / How to Do It This Week

The practitioner synthesis

Prompting & inference 04
Tools, repos, libraries 06
Architectural & model-selection 04
Methodological frames 07
  • Models / Apps / Harnesses three-layer split
    When evaluating any AI capability claim, separately ask which model, which app, which harness — improvements in one layer get conflated with the others
  • Verifiability × training-attention rubric
    Before delegating a task, score (a) is the success signal automatic? (b) was this task emphasized in lab training? Both high = trust; either low = supervise
  • Independent-eval-or-it-didn't-happen
    When labs publish productivity-cost numbers, weight the claim down until independent replication exists; if none in 2 weeks, the number is marketing
  • Audit-the-fine-print discipline
    For any lab productivity claim, ask: which fraction of which work, on what tasks, with what acceptance rate? Most numbers don't survive the question
  • Goodhart-check before measuring AI usage
    Tokens-per-employee is the Claudeonomics anti-pattern; measure decisions-improved-per-dollar instead. When a measure becomes a target, it ceases to be a good measure
  • What's now possible that wasn't
    Replace 'what can AI speed up?' with 'what information transformation was impossible before?' — Karpathy and Chollet converged on this framing the same week
    via Andrej Karpathy + François Chollet · x.com/fchollet/status/2058982905368773040
  • Issue-report hygiene against LLM-laundering
    When receiving an issue: ask the human to state in their own voice (a) what they did, (b) expected, (c) happened. Don't accept LLM rewrites that hide the actual observation
    via Armin Ronacher via Simon Willison · lucumr.pocoo.org/2026/5/24/pi-oss/
Papers worth a closer read 03
  • Learning, Fast and Slow: Towards LLMs That Adapt Continually
    Fast-context + slow-weight composition; 3× more sample-efficient than RL alone, 70% less KL divergence from base, continues to learn where RL stalls — production candidate for continual-learning workloads
  • MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
    Skill-centric lifecycle with unit tests + runtime feedback; treats skills as long-lived testable assets — adjacent to JC-OS's learning-graduation discipline
  • BRANE: Natural Language Query to Configuration for Retrieval Agents
    Per-query config selection beats static config tuning — matches best-fixed accuracy at up to 89% lower cost on MuSiQue, BrowseComp-Plus, FinanceBench
05 / Quotes

Lines worth carrying

We keep finding things that are mysterious, even unsettling. We find structures that mirror results from human...
Chris Olah at Vatican (Magnifica Humanitas presentation)
If enough other companies report the same, the bubble pops.
Gary Marcus, on Uber COO Andrew Macdonald
The rate of incoming security reports is 4-5 times higher than it was in 2024 and double the speed of 2025 — meaning that on average we now get more than one report per day.
Daniel Stenberg, curl maintainer
A lot of the emails I get from founders are now written in a hard-hitting journalistic style. I know they're written by AI... It feels like being lied to.
Paul Graham via Simon Willison
Thinking of AI as a productivity booster for prior workflows is the wrong framing.
François Chollet
Use auto.
Boris Cherny (Claude Code lead, #1 tip)
06 / Position shifts

What changed in the stance map

PersonThemeShiftNote
Anthropic (institutional) Cultural legitimacy ESCALATED From 'we publish' to 'we are cited by the Vatican'. Anthropic Institute formalized; Chris Olah at papal encyclical rollout
Boris Cherny / Claude Code Distribution + observability NEW THEME /usage shipping; 'use auto' official endorsement; teaching-basics push signals demographic expansion beyond power-users
Ethan Mollick AI / human boundary SHIFTED From 'use AI to do work' to 'choose what to keep human' — additive axis, not retraction; convergence with Graham, Chollet, Ronacher this week
François Chollet AI-as-productivity framing SHIFTED 'Productivity-booster for prior workflows is wrong framing' — same-week convergence with Karpathy on 'what's now possible?'
Gary Marcus AI productivity bubble ESCALATED Three-prong case: Uber COO admits no gains; retirement-fund exposure piece; OpenAI/Anthropic headline math audit
Arvind Narayanan Lab-claim audit NEW PUBLIC ANGLE First explicit 'audit the lab claim' piece this year — Google's $916-OS deflation; companion-piece to Marcus same week
Simon Willison Agentic security ESCALATED Copilot Cowork exfiltration + curl-maintainer pressure together name new failure mode: agentic features create exfiltration vectors AND break OSS maintenance economics simultaneously
07 / Cross-references

Who built on whom

08 / Source registry

The voices

Andrej Karpathy
@karpathy
Bear Blog · YouTube · X · Anthropic
Simon Willison
@simonw
simonwillison.net · X · GitHub
Garry Tan
@garrytan
YouTube · X · YC
Ethan Mollick
@emollick
One Useful Thing substack · X
Gary Marcus
@garymarcus
Marcus on AI substack · X · CNBC
Cassie Kozyrkov
Decision Intelligence substack
Nathan Lambert
Interconnects substack · Allen AI
Arvind Narayanan
Normal Tech substack · Princeton
Yejin Choi
papers · talks · NVIDIA/Stanford
Melanie Mitchell
@MelMitchell1
Santa Fe Institute · X
Yannic Kilcher
@yaborobot
YouTube
Sebastian Raschka
@rasbt
Ahead of AI substack · X · GitHub
Lakshya Agrawal
@LakshyAAAgrawal
papers · X · UC Berkeley
MIT Technology Review
AI section editorial
Jack Clark
@jackclarkSF
Import AI substack · X · Anthropic Institute
Boris Cherny
@bcherny
X · Anthropic / Claude Code
Chris Olah
@ch402
X · Anthropic interpretability
François Chollet
@fchollet
X · ARC · Keras
Erik Schluntz
@eschluntz
X · Anthropic / Claude Code