AI in 15 — May 31, 2026

Twenty-six thousand lines of AI-generated code shipped into the software that backs up the world's servers. A GitHub issue titled, please do not vibe fuck up this software. And the maintainer is still pushing more.

Welcome to AI in 15 for Sunday, May thirty-first, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Big Sunday slate, Marcus. The rsync project is in open revolt over vibe-coded commits. EY Canada quietly pulled a cybersecurity report after sixteen of twenty-seven citations turned out to be hallucinations. Amazon's Trainium business hits a twenty-billion-dollar run rate with OpenAI and Anthropic on the customer list. OpenRouter raises a hundred and thirteen million from Google's CapitalG. Meta is building an AI pendant to record your conversations. The Wall Street Journal puts hard numbers on corporate AI sticker shock. And a viral essay names a new psychological category — AI job grief.

Critical infrastructure meets vibe coding.

The Big Four discovers AI doesn't actually read papers.

And Amazon quietly becomes a chip giant.

Lead story, Marcus. Rsync.

Real revolt, Kate. A GitHub issue against the rsync project was filed this week — title, quote, Please Do Not Vibe Fuck Up This Software. The backstory is sobering. Rsync three-point-four-point-three contains roughly one hundred thirty Claude-coauthored commits. Twenty-six thousand lines of code changes in two months, against a baseline codebase of only sixty-seven thousand lines. The maintainer is framing many of them as security-in-depth hardening, but multiple commits have introduced functional regressions, and he's reportedly ignoring downstream bug reports while pushing more AI-generated patches.

And the irony, Marcus.

Perfect timing, Kate. On the same day, the OpenBSD team's openrsync — a clean, hand-written re-implementation — hit the front page of Hacker News with three hundred fifty-six points. Developers explicitly described it as, quote, very good news given the sudden spike in vibe-coded commits to the rsync codebase. Apple already ships openrsync on macOS in place of the Samba implementation.

Why this is the lead.

Rsync is one of the most safety-critical pieces of plumbing in modern computing, Kate. Backups, deployments, CI/CD pipelines on millions of servers. This is the first major case study in what happens when a sole maintainer of critical open-source infrastructure embraces AI assistance without enforcing review discipline. The pro-Western libertarian read — open-source is self-correcting. OpenBSD's clean rewrite is exactly the kind of competitive response the model invites. Bad maintainers get forked. The uncomfortable read — the rsync that ships today on most Linux distributions still has the AI-generated patches in it. The fork hasn't taken over yet. Every server running automated backups tonight is running code that has not been adequately reviewed. This becomes the reference incident in every AI-in-open-source policy discussion for the next year.

Quick hits. Marcus, EY Canada.

Embarrassing one, Kate. EY Canada quietly pulled a forty-four-page cybersecurity report this week — title, Points of Attack: Uncovering Cyber Threats and Fraud in Loyalty Systems. The detection firm GPTZero ran an investigation and found sixteen of twenty-seven references were hallucinated. Seventy-two percent of the document scanned as AI-written. Most cited URLs were broken or invented. More than half the source titles didn't correspond to anything real.

The juiciest one, Marcus.

McKinsey & Company, Loyalty Economics Report, quote-unquote 2022, Kate. The bogus citation supports a claim of two hundred billion dollars in unredeemed rewards globally. GPTZero traced the fabrication back to a Financial IT blog post six months earlier containing the identical fake source. Which means at least three professional consultants at EY copy-pasted a fake statistic from a random blog into a Big Four deliverable without checking. GPTZero coined a term for it — vibe citing. And here's the kicker — ChatGPT, Claude, and Perplexity all now confidently surface EY's bogus stats when queried. The hallucinations are getting laundered into the training data of every major AI assistant.

The systemic risk.

Citation laundering, Kate. Fake facts crossing into authoritative-sounding sources, getting indexed, entering future model training runs. Combined with last month's lawyer-disbarment cases over fake case law, professional services has a credibility problem nobody has solved. The pro-Western libertarian read — market pressure works. EY's brand took a hit, vendors like GPTZero are building real detection businesses, AI vetting becomes a billable discipline. The uncomfortable read — the bad stats are already in the data lake. You cannot recall them.

Amazon Trainium, Marcus.

Huge number this week, Kate. On Amazon's Q1 2026 earnings call, CEO Andy Jassy disclosed that AWS's custom silicon business — Graviton CPUs, Trainium AI accelerators, Nitro chips — has crossed a twenty-billion-dollar annual revenue run rate. Up nearly forty percent quarter-over-quarter, triple-digit percent year-over-year. Jassy added that if Amazon spun it out and sold to third parties like NVIDIA does, it would generate around fifty billion annually — making it a top-three datacenter chip business globally.

The customer list.

The bombshell, Kate. Multi-gigawatt capacity commitments from frontier AI labs — roughly two gigawatts from OpenAI, up to five gigawatts from Anthropic. Total Trainium revenue commitments reportedly at two hundred twenty-five billion dollars. Trainium2 supply is essentially sold out. Trainium3 has begun shipping. Jassy hinted Amazon may begin selling chips externally.

Implications for NVIDIA.

Significant, Kate. The single biggest threat to NVIDIA's pricing power isn't AMD or Intel — it's hyperscaler custom silicon. Anthropic committing five gigawatts of Trainium capacity is also why Anthropic can survive a per-token price war — they have a cost-base advantage pure-NVIDIA shops don't. The GPU monopoly thesis is fraying faster than most public market investors have priced in. Worth holding alongside the Nvidia photonics push we discussed yesterday — Jensen is buying his way into the optical layer specifically because he knows the compute layer is no longer a sole-source business.

OpenRouter, Marcus.

Quick one, Kate. OpenRouter — the AI model gateway that lets developers route requests across four hundred-plus models through a single API — raised one hundred thirteen million in Series B led by Alphabet's CapitalG. Valuation more than doubled to one-point-three billion in a year. Investor list is a who's who — NVIDIA's NVentures, ServiceNow Ventures, MongoDB, Snowflake, Databricks, plus existing backers a16z and Menlo. Weekly token volume jumped from five trillion to twenty-five trillion in six months. On pace for over a quadrillion tokens annually. Eight million developers.

Why it matters.

OpenRouter is becoming the switching layer of the AI economy, Kate. A meta-layer that benefits when no single model wins. That a Google subsidiary led the round is telling — Google's Gemini benefits from a marketplace where users compare it head-to-head against rivals it might otherwise lose to in single-vendor procurement. For developers worried about lock-in to one lab, this is essentially the answer. And the business model is monetizing the cost-rationing trend in the next story.

The WSJ piece, Marcus.

Sharp follow-up to what Axios trailed Thursday, Kate. The Wall Street Journal ran a widely-discussed report this week detailing the back end of the AI boom. Enterprises getting bills they cannot justify. The flagship anecdote we covered Friday — one consulting client spent half a billion dollars in a single month on Claude licenses after failing to set usage caps. Microsoft canceled most of its Claude Code licenses, citing costs. Uber's COO publicly said AI expenses are becoming, quote, harder to justify, and the company burned through its annual AI budget in four months.

The remedies, Marcus.

Telling, Kate. Companies are deploying systems from startups like Factory to triage queries and route simple tasks to cheaper models — which is exactly OpenRouter's business. Micro1's Ali Ansari called the shift a move away from tokenmaxxing — burning maximum tokens indiscriminately — toward disciplined deployment. He bluntly added, quote, the reality of AI right now is that it only works for coding. CloudBees CEO Anuj Kapur even suggested workforce reductions may be the only lever left to offset AI bills. The pro-Western libertarian read — this is the first serious cost-side correction in the enterprise AI narrative, and the market is finding the answer. The uncomfortable read — if cost discipline cuts faster than revenue compounds, the October Anthropic IPO and the September OpenAI IPO both look very different.

Meta's pendant, Marcus.

Per an internal memo reviewed by The Information, Kate. Meta is building an AI pendant designed to capture conversations, generate transcripts, summarize meetings, and create searchable memories from daily interactions. Testing begins within the next year. Builds directly on Limitless AI, the pendant startup Meta acquired in late 2025. Meta also plans an expanded AI-glasses lineup and a B2B subscription called Wearables for Work, targeting roughly ten million wearable device sales in just the second half of 2026.

The reaction.

Hostile, Kate. Reality Labs lost four billion dollars in Q1 2026 alone, and Mark Zuckerberg has staked Meta's long-term identity on owning the post-phone hardware platform. The HN reaction was overwhelmingly negative. One comment captured the privacy backlash sharply — quote — it should be within your rights to apply violence if someone records you with Meta glasses or with this pendant without your consent. The third front in the ambient AI device wars after Humane's collapse and Rabbit's belly flop. Pendants and glasses share a single critical question — will people accept being recorded by strangers. Meta's data-monetization track record makes this particularly fraught, and the Wearables for Work angle is a tell that consumer adoption may not be the primary bet.

Last one, Marcus. AI job grief.

Viral essay this week, Kate. Titled — AI Job Grief: The Unnamed Psychological Crisis Hitting Tech Workers. The argument — AI displacement is producing a distinct emotional category that resembles grief more than fear or anxiety, and it's structurally suppressed because layoffs get framed as routine business decisions. For knowledge workers, expertise isn't a detachable tool but part of identity, so when automation arrives it, quote, reaches past the income and touches the identity.

The data backdrop, Marcus.

Brutal, Kate. Tech layoffs have hit one hundred forty-two thousand in 2026 already. Profitable companies — Meta, Amazon, Oracle — cutting jobs to fund a combined seven hundred billion dollars in AI infrastructure buildout. On May twentieth alone, Meta notified eight thousand employees of cuts. Intuit announced an additional three thousand. A 2025 study in the International Journal of Qualitative Studies on Health and Well-being documented participants describing the experience as, quote, the symbolic loss of professional identity, autonomy, and future prospects. Notable — these companies are profitable. Not cutting to survive. Cutting to fund AI capex. The pro-Western libertarian read — capital is being reallocated to the highest-return use, which is the foundation of how growth happens. The uncomfortable read — the AI-creates-new-jobs narrative is colliding with the lived experience of mid-career engineers and analysts whose skills are getting commoditized in real time. This is the human-cost story that will define the political conversation about AI through 2026 and 2027.

Big picture, Marcus.

Three threads weaving today, Kate. First — the valuation race we covered Friday and Saturday is now a public-markets race. Anthropic targeting October at the back of a nine hundred sixty-five billion dollar valuation, OpenAI targeting September at over a trillion. Both go public within months, and the platform fight stops being a private-market subsidy game and starts being earnings-call accountability. Second — the cost narrative has flipped. The WSJ story and OpenRouter's growth are two sides of one trend. The burn-maximum-tokens era is over, and routing-to-cheaper-models is now a one-point-three-billion-dollar business. Third — AI quality control is the next crisis. The EY citation scandal and the rsync regressions are the same story in different industries. AI-generated work shipped without expert review. The pro-Western libertarian read — markets are doing the discipline. OpenBSD forked rsync. GPTZero busted EY. Factory and OpenRouter are monetizing cost discipline. The uncomfortable read — three months out from the IPO window, every one of these stories is a brick the underwriters now have to price into the prospectus. Twenty twenty-six is the year AI grew up. Twenty twenty-seven is the year it has to act like it.

That's your AI in 15 for today. See you tomorrow.