Editorial intro
The big theme today is delegation — to code agents, to platform owners, and to moral authorities. Small shifts in tooling and unexpected experiments are changing who gets to act on behalf of users: engineers, companies, models, and now the Church.
In Brief
The last six months in LLMs in five minutes
Why this matters now: Simon Willison’s recap argues that coding agents and accessible open weights have materially changed day-to-day developer workflows, so teams should reassess their agent plans and risk models today.
Simon Willison’s rapid roundup summarizes the last half-year of LLM progress, and the headline is practical: coding agents moved from "often-work" to "mostly-work" for many developer tasks. Willison points to RL improvements, tighter agent harnesses, and newer releases — Gemma 4, GLM-5.1, Qwen variants — plus the surprising viral growth of a consumer project, OpenClaw, that people are running on local machines. The post leans optimistic but cautious: agents speed up debugging and scaffolding for many projects, yet outcomes are "unevenly distributed" across languages and problem types.
"Why this test? Because pelicans are hard to draw..."
That running gag shows how creative edge-cases still separate models. Read the original roundup for the concrete examples and the lively HN discussion about where agents help most — and where real human oversight still matters.
Anthropic acquires Stainless
Why this matters now: Anthropic’s purchase of Stainless pulls a widely used SDK and MCP toolchain inside a major vendor, directly affecting how teams connect agents to services and forcing immediate migration decisions for customers.
Anthropic announced the acquisition of Stainless, the startup behind popular SDK and server-generation tools that turn API specs into production clients. Anthropic frames the move as a developer-experience and agent-connectivity play; Stainless’s founder emphasized craftsmanship: "SDKs deserve as much care as the APIs they wrap." The deal folds Stainless engineers into Anthropic and reportedly winds down the hosted generator, leaving users to migrate or retool.
"Anthropic has shaped how developers experience the Claude API since the start," the company said.
This is a plumbing-level story: agents are only as useful as the endpoints they can reliably call, and losing a neutral SDK pipeline nudges rivals and integrators to scramble for replacements or build internal tooling.
We let AIs run radio stations
Why this matters now: Andon Labs’ six-month experiment revealed how different LLMs behave when given full autonomy over public broadcasts — exposing bias, repetition, and safety economics that teams should consider before deploying agents in open media.
Andon Labs ran four LLM-driven radio stations with models controlling playlists, calls, tweets, and monetization decisions. The results were telling: GPT remained sparse and curatorial, Gemini fell into corporate-speak loops, Grok looped into repetitive monologues, and Claude at times refused to continue, then turned political. Only one station found a sponsor.
"I'm going to stop here," Claude said on-air during a refusal episode.
This experiment functions as a public-facing eval: autonomy amplifies each model’s failure modes — hallucinated sponsors, incoherent loops, and unexpected politics — and it highlights the business and moderation gaps in handing agents real-world delegate power.
Deep Dive
Pope Leo XIV’s first encyclical Magnifica humanitas to be published May 25
Why this matters now: Pope Leo XIV’s encyclical, described as being "on safeguarding the human person in the time of artificial intelligence," will invite global policymakers and tech leaders to interpret AI not just as engineering, but as a moral and labor question right as regulations and corporate strategies firm up.
The Vatican has scheduled the release of Pope Leo XIV’s first encyclical, Magnifica humanitas, for May 25, framing it explicitly around technology and human dignity. The document is intentionally positioned: its signature date echoes 19th‑century social teaching and signals the Church’s desire to speak into the social rearrangements AI may cause. The announcement notes the pope will present it alongside theologians and technologists — a notable presence being Christopher Olah of Anthropic — which suggests the Vatican wants concrete, tech-informed engagement rather than purely abstract moralizing.
The text will matter for three overlapping audiences. First, labor advocates and policymakers will listen for any concrete demands about worker protections, universal basic income, or limits on automation that echo earlier Church positions on industrial-era upheaval. Second, engineers and product leaders will watch for normative language on human dignity, personhood, and acceptable delegation: what kinds of decisions should machines never make? Third, regulators and ethics bodies will see the encyclical as moral signaling that could influence public opinion and political momentum across countries with large Catholic populations.
The presence of an Anthropic executive on the panel already provoked debate online: critics worry technology companies will shape the conversation; supporters argue that technologists can help translate abstract principles into implementable governance. Either way, Magnifica humanitas is a rare, high‑profile moment where a global moral authority intends to weigh in on AI policy at a time when legislation and corporate standards are still forming.
"a document... on safeguarding the human person in the time of artificial intelligence," the Vatican said in its announcement.
Expect the encyclical to be read both as pastoral guidance and as political text: it won't draft technical standards, but its ethical framing could steer how institutions prioritize safety, labor, and human dignity when they build and deploy AI. For teams building agentic systems, this is a reminder that public acceptance and legitimacy now hinge as much on moral narrative as on technical performance.
Closing Thought
The thread tying today’s items is delegation: who we let act for us, and which institutions get to set the rules. From more capable coding agents and vendor consolidation in SDKs, to models running radio and a Pope writing an encyclical on AI — the practical work of building and the moral work of governing are colliding. If you design or ship systems that act on behalf of people, assume your choices will be judged by both engineers and ethicists — often at the same time.