AI in 15 — May 16, 2026

Eight hundred thousand people are about to hand ChatGPT the password to their bank account. OpenAI shipped Personal Finance yesterday, plugged straight into Plaid, twelve thousand institutions, your portfolio and your spending and your subscriptions all loaded into a chat window. Hacker News is asking why, exactly, an AI vendor needs live access to your divorce predictors.

Welcome to AI in 15 for Saturday, May sixteenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Loaded Saturday slate, Marcus. OpenAI launches ChatGPT Personal Finance with full bank integration. The Bun rewrite we covered yesterday just failed its first memory-safety check with thirteen thousand unsafe blocks under fire. Mitchell Hashimoto says entire companies are under AI psychosis and the post is everywhere. Elon Musk dissolved xAI into SpaceX as a subdivision. OpenAI has retained outside counsel to weigh suing Apple over the Siri integration. Recursive Superintelligence emerges from stealth at four-point-six billion. And Turso is shutting down its bug bounty because AI-generated slop killed it.

ChatGPT meets your checking account.

AI psychosis at scale.

And xAI ceases to exist.

Lead story, Marcus. OpenAI shipped ChatGPT Personal Finance yesterday. Walk me through it.

Genuine product expansion, Kate. OpenAI launched a new Pro-tier feature that lets users link checking, savings, brokerage, and credit card accounts directly into ChatGPT through Plaid — the financial-data plumbing layer that powers most fintech apps. More than twelve thousand US institutions covered. Chase, Fidelity, Schwab, American Express, Robinhood, Capital One. The Finances tab in the sidebar pulls live transaction data into a dashboard showing portfolio performance, spending categories, subscriptions, and upcoming payments. Then you ask in natural language. Why am I overspending this month. Should I rebalance my IRA. Web and iOS, US-only, Pro-only for now, Plus access teased after a feedback period. Future Intuit integration coming for tax-impact analysis and credit-card pre-approval predictions.

And the privacy framing.

OpenAI emphasizes that ChatGPT never sees bank passwords, Kate. Plaid handles tokenized authentication. Disconnect an account and synced data purges within thirty days. You can wipe specific financial memories from the Finances page. But the Hacker News reaction was sharp. Dozens of top commenters argued there should be more friction here, not less. The most-upvoted comment pointed out this lands just months after OpenAI tied ChatGPT into people's health records. The line was, bank data is far more valuable than medical data — political donations, divorce predictors, vices.

Why does this matter beyond the privacy debate.

Because it's a defining moment in agentic AI's collision with the most sensitive data category most consumers own, Kate. If the product works, ChatGPT becomes the front door for personal finance. Mint-style apps get displaced. Traditional advisors get squeezed. Bank chat interfaces become irrelevant. If it fails publicly — a misfire on a tax question, a leaked spending profile, an unauthorized transaction once the agentic version ships — it triggers the regulatory backlash that has so far skipped AI in the US. And note the trajectory. Health records two months ago. Bank accounts today. Consumer agentic AI is racing into categories where one bad headline ends the product. The libertarian read, Kate, is that consumers should get to make this trade themselves — but only if the disclosure is clean and the off-ramps are real. Watch the first major mistake.

Quick hits. Marcus, the Bun rewrite we celebrated yesterday already has a problem.

Big one, Kate. Within twenty-four hours of merging that million-line Rust port, a community issue — number thirty thousand seven hundred nineteen — documented that the new codebase contains thirteen thousand unsafe blocks across seven hundred thirty-six files, and fails basic Miri checks. Miri is Rust's memory-safety interpreter. The reproducible example — PathString slice — constructs a dangling reference, classic undefined behavior, that compiles cleanly and passes the test suite but corrupts memory once Miri inspects it. The issue author's recommendation was direct. Quote, please consider not vibe coding Rust as AIs are not good at writing Rust, and hire a real Rust dev.

So the entire memory-safety pitch of moving from Zig to Rust.

Currently exists on paper, Kate. Real context worth knowing — Zig recently formalized a ban on LLM-authored contributions, which is why Anthropic-owned Bun couldn't upstream AI-generated changes back to its Zig fork. That's what forced the Rust migration. Rust is just a language Claude is much better at producing. The most-upvoted Hacker News comment mourns the loss. The author was drawn to Bun precisely because it was hand-written in Zig with Andrew Kelley's design taste. Now it's a translated codebase with safety guarantees the type system can't actually verify. This is the live test of the ship-fast-let-agents-patch-it thesis. If undefined behavior in those thirteen thousand blocks turns into a steady stream of CVEs over the coming months, this PR rewrites how the industry views AI-translated infrastructure code.

Mitchell Hashimoto story, Marcus. He posted something on Thursday that has not stopped reverberating.

Eleven hundred points on Hacker News, Kate. Hashimoto — co-founder of HashiCorp, creator of Vagrant and Terraform, now of Ghostty terminal fame — posted that, quote, entire companies right now are under heavy AI psychosis, and it's impossible to have rational conversations about it. He won't name names because they include friends he respects. His core argument draws on his infrastructure-era scars. In the early cloud days the industry had a fierce debate — mean time between failure versus mean time to recovery. The MTTR-is-all-you-need crowd argued you should ship buggy code if you could heal fast. Hashimoto says the same fight is back, but now powered by agents. Quote, it's fine to ship bugs because the agents will fix them so quickly and at a scale humans can't. His warning — you can automate yourself into a very resilient catastrophe machine. Systems that look healthy on every local dashboard but become globally incomprehensible.

And the timing.

Surgical, Kate. The post landed in a week filled with corroborating data. Amazon employees tokenmaxxing to game internal AI leaderboards, which we covered Wednesday. Turso shutting down its bug bounty, which we'll get to. Bun's million-line rewrite with thirteen thousand unsafe blocks. Hacker News commenters started proposing AI rescue consulting as the next high-value specialty — the way data-recovery firms get called in after a breach. Hashimoto is one of the most respected voices in infrastructure, and he is explicitly not anti-AI. He uses Claude heavily for Ghostty. When he says corporate AI adoption has slipped its moorings, that's a structural warning, not a Luddite take. Expect this post to be cited in board meetings within weeks.

xAI story, Marcus. Elon Musk just dissolved his AI company.

First major frontier-lab dissolution of the cycle, Kate. Musk announced last week that xAI no longer exists as a standalone company. After the February merger that valued xAI at two hundred fifty billion and SpaceX at one trillion, all xAI products — Grok, the Memphis Colossus supercomputer — now live under a SpaceXAI division running parallel to Starship and Starlink P&Ls inside SpaceX. Musk's framing was, quote, xAI was not built right first time around, so is being rebuilt from the foundations up. All eleven of xAI's original co-founders are gone — Igor Babuschkin, Jimmy Ba, Tony Wu, Christian Szegedy, Greg Yang. Talent has bled to Meta's Superintelligence Labs and Mira Murati's Thinking Machines.

And the GPU utilization story.

This is the kicker, Kate. Wccftech reported this week that xAI was using only eleven percent of its five hundred fifty thousand Nvidia GPUs at Memphis, against forty-three to forty-six percent at Meta and Google. Anthropic is now reportedly renting some of that idle capacity. So the world's largest single-site GPU cluster is being subleased to a competitor by its parent because it can't fill the load itself. The picture from late twenty-twenty-four — four hyperscalers plus a couple of independents — is now Google, Microsoft-OpenAI, Meta, Anthropic, and a SpaceX subsidiary. Consolidation phase is officially here. The next two years are won on distribution, not benchmark scores.

OpenAI versus Apple, Marcus. Bloomberg broke this Thursday.

Major fracture in the highest-stakes consumer AI partnership of this cycle, Kate. OpenAI has retained outside counsel to weigh breach-of-contract action against Apple. The grievance — when the WWDC twenty-twenty-four partnership was struck, OpenAI expected its placement in Siri and Apple Intelligence to drive massive ChatGPT Plus subscription growth. Two years on, OpenAI execs say the integration is buried, features are hard to find, and revenue is, quote, nowhere near projections. An unnamed OpenAI executive told Bloomberg, quote, they basically said OpenAI needs to take a leap of faith and trust us. It didn't work out well. Apple meanwhile is now publicly testing Claude and Gemini as alternative providers — diluting OpenAI's exclusivity. Apple has also privately complained about OpenAI's privacy hygiene and Jony Ive's OpenAI hardware project.

Will it actually go to court.

Formal lawsuit unlikely until after the Musk trial concludes, Kate. But a breach-of-contract notice could come sooner. The Apple-OpenAI relationship was, until this week, the single biggest distribution channel in consumer AI. If it ends — or if Apple opens Siri to Anthropic and Google equally — the AI assistant market splits four ways instead of being a default-ChatGPT pipeline to one-and-a-half billion iPhones. Apple's WWDC is in three weeks. Whatever Tim Cook announces will be read against this Bloomberg story. And the timing tracks with our running thread — distribution, not models, is where the next year is fought.

Funding story, Marcus. Recursive Superintelligence emerged from stealth Wednesday.

Six hundred fifty million dollars at a four-point-six-five-billion valuation, Kate. Led by Richard Socher — ex-Salesforce chief scientist, you-dot-com founder — co-founded by ex-Meta FAIR director Yuandong Tian. GV — that's Google's venture arm — and Greycroft led the round, with AMD Ventures and Nvidia participating. The bet is explicit and extreme. Build AI systems that improve themselves by analyzing their own performance and rewriting their training, weights, and architecture without human input. They're calling it open-ended self-improvement. Other named hires include Peter Norvig and Cresta co-founder Tim Shi. The plan — scale compute in San Francisco and London, run their first Level One autonomous training system, public launch targeted mid-twenty-twenty-six.

And the broader signal.

This is the most explicitly self-improvement-focused, well-funded lab to launch since the SSI and Thinking Machines wave, Kate. Whether you find recursive self-improvement an exciting frontier or a thing humanity should be panicking about, six hundred fifty million in two days with Nvidia and Google money attached says the smart money is still betting on more capability per quarter, not less. And it lands the same week xAI got dissolved — so the message to founders is clear. Bigger and broader is not winning right now. Sharply specialized frontier labs with charismatic founders and a clean thesis still get top-tier capital instantly.

And the bug bounty story, Marcus. Turso just shut theirs down.

Latest casualty in a fast-growing pattern, Kate. Turso, the company behind the SQLite-compatible distributed database libSQL, announced yesterday it is killing its thousand-dollar-per-bug security bounty. The program ran nearly a year, paid out five legitimate winners, and was steadily drowned by AI-generated submissions. The examples Turso shared are surreal. Pull requests that manually corrupt a database header, then claim the corruption proves a vulnerability. Critical bugs reporting that a SQL database allows SQL execution. PRs that modify the source code to introduce a bug, then report the bug. The economic asymmetry kills the program. Quote, it costs the slopmaker perhaps a minute to generate their submission. It costs us hours to read, understand, and engage. Turso joins cURL, Nextcloud, and HackerOne's Internet Bug Bounty — all paused or scaled back in twenty-twenty-six for the same reason.

So the labor market for open-source security is collapsing.

Collapsing under AI generation costs that approach zero on the offense side, Kate, while triage costs stay human-bound. Bug bounties were the closest thing the open-source world had to a real security labor market. They're being eaten by the same models that promise to defend the same software. And it's a direct corollary to the Hashimoto post. Defenders bear human costs. Attackers and slop-generators don't. Every economic model that assumes symmetric effort is being broken right now.

Big picture, Marcus.

Three threads tie today together, Kate. First — AI is becoming infrastructure for genuinely sensitive systems. Bank accounts via OpenAI and Plaid. Production runtime code via Bun. Security research via Mythos and Daybreak. The stakes per deployment are climbing much faster than the safety guarantees underneath them. Second — the credibility of corporate AI metrics is collapsing. Hashimoto's psychosis post, Amazon's tokenmaxxing from Wednesday, Turso's bug-bounty death today. Watch any productivity or usage claim coming out of big tech this year with extreme skepticism. Third — the frontier-lab landscape is consolidating in real time. xAI dissolved into SpaceX. Anthropic now sits on a forty-billion-dollar Google line and a twenty-billion Amazon line plus five gigawatts of compute. OpenAI is fighting Apple while building a finance platform. The big-four-labs-plus-scrappy-independents picture from twenty-twenty-four is now big three platforms — OpenAI, Anthropic, Google — plus Meta, plus a SpaceX subdivision, plus a small handful of credible specialists like Recursive and Thinking Machines. The pro-Western libertarian read, Kate, is that competitive consolidation around real cost discipline is healthy. The risk is the credibility thread. If Hashimoto is right that entire companies are operating under AI psychosis, the soft middle of this whole capex cycle is much softer than the headline numbers admit. The next twelve months separate the labs that ship working products from the ones that just ship token counts.

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