AI in 15 — May 17, 2026

Malta just became the first country on Earth where every citizen gets ChatGPT Plus as a public benefit. Five hundred sixty thousand people, twelve months free, delivered through the same national ID system they use to file taxes. OpenAI isn't selling software anymore. It's negotiating with passports.

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

And I'm Marcus, your co-host.

Sunday slate, Marcus. OpenAI signs the first nation-state ChatGPT Plus rollout in Malta. Bloomberg analyzes fresh BLS data and finds the first real evidence AI is eating white-collar jobs. Meta confirms eight thousand layoffs on Wednesday to pay for a hundred-and-forty-five-billion-dollar GPU bill. Nvidia drops an open-source world model that runs on a gaming card. A viral essay declares the competitive cybersecurity scene officially dead. And an OpenAI engineer racks up a one-point-three-million-dollar token bill in a single month.

OpenAI annexes a country.

AI-exposed jobs finally show up in the macro data.

And the hacker training ground collapses.

Lead story, Marcus. OpenAI's Malta deal. What exactly did they announce on Friday?

First true country-level deployment of a consumer AI product, Kate. Every citizen and resident of Malta — about five hundred sixty thousand people, plus Maltese nationals living abroad — gets a free twelve-month ChatGPT Plus voucher. Distribution runs through the Malta Digital Innovation Authority. You log in with Malta's national eID, complete a short online course called AI for All built by the University of Malta, then collect the voucher. Phase one starts this month. The whole thing sits inside a new OpenAI for Countries program — Sam Altman's framing is explicitly that Malta is a template.

Why Malta specifically?

It's the cleanest possible test bed, Kate. English-speaking, EU member, digitally mature, small enough that you can hit genuine population-scale AI literacy without overwhelming the logistics. If completion rates and productivity surveys come out strong, OpenAI walks into Ireland, Estonia, Singapore, eventually France with a real case study in hand. Up to now, governments bought AI for specific agencies — defense, tax, health. This is the inverse. A consumer subscription delivered as a citizen benefit, like a public library card.

Critics flagging anything?

Two threads, Kate. The first is data privacy. Malta has a complicated reputation — money-laundering investigations, a press-freedom history that includes the murder of Daphne Caruana Galizia. Critics are asking what happens to the conversations five hundred sixty thousand people have with ChatGPT when those interactions are tied to a national digital identity. The second thread is what happens when the free year ends. OpenAI is essentially running a customer-acquisition program at sovereign scale. Twelve months in, when the voucher expires, the cost shifts to the citizen or to the taxpayer. Either way, somebody is paying twenty dollars a month at the end of this rainbow.

And the competitive angle.

A very quiet flex, Kate. Anthropic and Google have nothing comparable announced. OpenAI is using its distribution edge to lock in a new go-to-market — ChatGPT delivered not by app store but by national infrastructure. If this works, the unit of customer acquisition stops being a user and starts being a country.

Quick hits. Marcus, Bloomberg dropped a piece Friday on AI and jobs that everyone is talking about. What did they find?

The cleanest signal we've gotten so far, Kate. The Bureau of Labor Statistics published its annual occupational employment data, and Bloomberg ran the AI-exposure cut. A group of eighteen occupations the BLS itself flagged as AI-exposed — about ten million US jobs total — shed zero-point-two percent of employment in the year through May twenty-twenty-five. Overall employment grew zero-point-eight percent over the same window. Strip out medical secretaries, which is its own trend, and the remaining seventeen occupations dropped one-point-six percent. Second straight year of decline.

Where is the damage concentrated?

Customer service reps lost a hundred and thirty thousand positions, down four-point-eight percent year-over-year, Kate. Non-medical secretaries fell thirty-one thousand. Wholesale and manufacturing sales reps dropped almost twenty-nine thousand. Go back to May twenty-twenty-two, right before ChatGPT launched, and the worst-hit roles are credit authorizers down twenty-six percent, broadcast announcers and DJs down twenty-one percent, sales engineers down thirteen. Goldman Sachs economists noted that job openings in AI-exposed roles have now fallen below pre-pandemic levels.

Counterarguments?

Two real ones, Kate. The BLS itself cautions that the eighteen-occupation list is illustrative, not definitive. And the Trump tariff regime plus elevated oil prices are simultaneous pressures, so the AI signal is tangled with cyclical pain. But the divergence between exposed and non-exposed roles inside the same macroeconomic environment is exactly what AI-displacement theory predicts. Every previous, quote, AI is taking jobs headline got swatted down with, no, that's just the broader economy. That defense just got harder to make.

Meta story, Marcus. Eight thousand layoffs Wednesday.

Three days from today, Kate. Roughly ten percent of the company. Chief people officer Janelle Gale told staff the cuts will, quote, run the company more efficiently and offset the other investments we're making. Another six thousand open positions are being eliminated. Zuckerberg has refused to rule out a second round in the back half of twenty-twenty-six. The brutal contrast is the financials. Meta posted nearly sixty billion in fourth-quarter revenue last year, up twenty-four percent. Twenty-three billion in net income. They are not cutting because they are losing money. They are cutting because AI capex for twenty-twenty-six is projected at a hundred and fifteen to a hundred and thirty-five billion dollars, up from seventy-two billion last year. Add in Superintelligence Labs spending and it's roughly a hundred and forty-five billion all-in.

And the inside-Meta reporting.

Bleak, Kate. Wired had staffers openly hoping to be laid off to collect sixteen weeks of severance and eighteen months of paid healthcare. The SF Standard ran a piece this week describing employees being asked to, quote, train your replacement, without compensation. The unique honesty of this round is that Meta isn't pretending AI made the people redundant. They're saying out loud that the layoffs are paying for the GPUs. Human headcount swapped directly for hardware. Wall Street loves it — margins go up. The labor market gets a sharp downdraft from one of the most profitable companies on Earth, and the line that AI creates more jobs than it destroys gets noticeably harder to sell.

Tech palate cleanser, Marcus. Nvidia released SANA-WM on Saturday. What is it?

A 2.6-billion-parameter open-source world model, Kate. Generates 720p video up to sixty seconds long from a single starting image plus a six-degree-of-freedom camera trajectory. You specify the path the camera should travel and the model hallucinates a temporally consistent world along it. Code is Apache 2.0. Weights are coming, though Hacker News commenters are skeptical until they see them. A distilled variant produces a sixty-second clip in thirty-four seconds on a single RTX 5090. That's two-point-one times real-time on a consumer graphics card.

How does it stack up?

Thirty-six times higher throughput than LingBot-World — which uses fourteen billion plus fourteen billion parameters on eight GPUs — at comparable visual quality on the standard VBench benchmark, Kate. Technical recipe combines linear attention with selective softmax, a dual-branch camera control system, and low-rank refiners. The top Hacker News comment captured the catch — the demo videos look conspicuously game-like, suggesting Unreal Engine was used heavily for synthetic training data. So it's not a general-purpose Sora competitor. It's specifically tuned for navigable environments.

Why does this matter?

World models have been the buzzy frontier since Sora and Genie, Kate. Generate not just video but navigable, physically coherent space. SANA-WM matters because it brings minute-scale, camera-controlled generation down to consumer-tier hardware. That's the inflection point for indie game studios, robotics simulation, and small film teams. And it continues Nvidia's playbook — subsidize the open-model ecosystem that sells more 5090s. Jensen makes the picks-and-shovels case better than anyone in tech right now.

CTF story, Marcus. A blog post that went viral this weekend declared the competitive cybersecurity scene officially dead.

Three hundred fifty-six points on Hacker News, Kate. The author is a competitive Capture The Flag player. The argument — when GPT-4 era models could solve the easy challenges, organizers adapted. But, quote, when Claude Opus 4.5 dropped, almost every medium difficulty challenge, and some hard challenges, became agent-solvable. GPT-5.5 Pro now reportedly clears Insane-difficulty heap-pwn problems. The twenty-twenty-six CTFTime leaderboard is, quote, unrecognisable. Established teams have collapsed in the rankings, replaced by whoever is willing to burn the most API tokens.

And this connects to the AISI evaluation.

Same week, Kate. The UK AI Safety Institute published its red-team work on Claude Mythos Preview — Anthropic's next-gen model. Mythos succeeds at expert-level CTF tasks seventy-three percent of the time, tasks no model could complete before April twenty-twenty-five. AISI's Last Ones benchmark — a thirty-two-step corporate network attack that takes a skilled human roughly twenty hours — was first solved by Mythos in three of ten attempts, improved to six of ten in later runs. The blog author's conclusion is bleak. The educational ladder that turned curious teenagers into security researchers is gone. Defensive design — making challenges deliberately AI-hostile — produces problems that are also guessy and unpleasant for humans.

Why does this matter beyond the CTF community?

Because CTFs are the global recruitment pipeline for offensive security talent, Kate. If the format collapses, you lose the funnel that produced most of the elite red-teamers at firms like Google Project Zero and Trail of Bits. Pair this with the Google announcement from May eleventh — that they caught hackers using AI to weaponize a zero-day for a mass exploitation event — and the picture is coherent. AI is simultaneously eating the training ground for defenders and accelerating the capabilities of attackers. This is the cybersecurity version of what's happening in education writ large.

DeepSeek story, Marcus. There's a steering-vectors angle making the rounds this weekend.

Two stories interleaved, Kate. First, DeepSeek's V4-Flash — a 284-billion-parameter mixture-of-experts model with 13 billion active and a one-million-token context — is now powerful enough to match low-end frontier models on agentic coding tasks, running locally. Second, that opens up steering vectors to ordinary developers. Steering vectors are the technique behind Anthropic's Golden Gate Claude experiment. Feed the model paired prompts, measure the difference in internal activations, and you extract a vector that captures a behavioral concept. Inject it into other prompts and you reproduce the behavior. Brain surgery on a neural network without retraining.

And someone shipped a tool around it.

Salvatore Sanfilippo — antirez, the creator of Redis — has been building DwarfStar 4, Kate. Stripped-down llama.cpp variant that runs only V4-Flash and treats steering as a first-class feature. The top Hacker News comment claimed his library can completely remove refusal behaviors from V4-Flash via the right activation vectors. The academic name for that is abliteration. An earlier paper found most refusals sit on a single vector you can simply subtract out.

So what's the implication?

Two things, Kate. The consumer ceiling for local AI just moved up significantly — you can run something close to frontier on your own machine. And the safety story for open-weights models is now visibly fragile. Alignment training can be erased with a single vector. DeepSeek is a Chinese lab, weights released openly are quickly modified worldwide by anyone with a GPU. The branding is, quote, free AI for the world. The practical effect is uncensored models proliferating beyond any safety regime. Anyone building on top of open-weight Chinese models should plan accordingly.

And to close, Marcus, an oddly fitting color story. A one-point-three-million-dollar API bill.

Peter Steinberger, Kate — creator of OpenClaw, now working at OpenAI — disclosed on X that his team of about three people racked up a one-point-three-million-dollar OpenAI API bill in thirty days. Six hundred and three billion tokens. Seven-point-six million requests. Mostly GPT-5.5. Because he works at OpenAI, the company effectively eats the cost. He runs roughly a hundred Codex instances in the cloud that write code, review PRs, hunt security holes, deduplicate issues, monitor benchmarks, post regression alerts to Discord, even listen in on team meetings and spin up PRs from things mentioned in conversation.

What's the framing?

Quote, exploring how software would be built if token costs didn't matter, Kate. He noted that simply disabling Fast Mode would cut the bill by seventy percent. The Hacker News reaction was sharp. Top commenter compared it to dot-com customer-acquisition costs subsidized by VC money — argued the unit economics of AI coding aren't real yet. Others pointed out that a two-hundred-dollar Codex subscription delivers five to six thousand dollars of raw API usage if you max it out, which tells you how heavily OpenAI is subsidizing power users right now.

So is he the future, or the canary?

Honestly both, Kate. Costs do keep falling, but, quote, if costs didn't matter is doing a lot of work in that sentence. The org-chart implication is the part to watch. Three humans plus a hundred agents replacing what used to be a twenty-engineer team. Capital substituting for labor at unprecedented speed, happening fastest at the labs themselves. The places that know AI best are the first to rip out their own headcount.

Big picture, Marcus.

Three of today's stories tell the same story from different angles, Kate. AI is restructuring the relationship between humans and the institutions that used to train, employ, and certify them. Malta hands free AI to every citizen as a national upgrade — that's the institution as distributor. Bloomberg shows the US labor market quietly hemorrhaging exactly the white-collar roles AI is good at — that's the institution as employer. Meta makes the trade explicit, eight thousand humans out to pay for a hundred and forty-five billion in GPUs — that's the institution choosing capital over labor in plain sight. And the CTF story is the canary for what happens when AI eats not the jobs but the ladder — the place where talent used to develop. Whether you cheer or grieve depends on whether you think AI will eventually build new ladders too. The pro-Western libertarian read, Kate, is that the cycle is doing what it always does — capital substitution, productivity gains, dislocation on the way to a new equilibrium. The honest caveat is that the dislocation is visibly faster and more concentrated than the equilibrium is arriving. Malta's bet is that being early on adoption beats being early on caution. Meta's bet is that GPU spending compounds faster than headcount cuts hurt. Both bets get tested over the next twelve months.

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