AI in 15 — May 06, 2026

Google Chrome is silently writing a four-gigabyte AI model to the hard drive of every user on the planet. No prompt. No notification. No consent. A privacy researcher caught it on camera, and Hacker News lit up with over thirteen hundred upvotes overnight.

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

And I'm Marcus, your co-host.

Big show, Marcus. Chrome is conscripting four gigabytes of disk space without asking. Anthropic dropped ten finance agents and locked in Microsoft Office integration. OpenAI quietly replaced ChatGPT's default with GPT-5.5 Instant — fifty-two percent fewer hallucinations and, blessedly, fewer emojis. Google's Gemma 4 just got a three-times speedup. Publishers are personally suing Mark Zuckerberg over pirated training data. Cerebras filed a twenty-six-billion-dollar IPO. Apple ran out of RAM. Xbox killed its Copilot. And Telus is using AI to scrub the accents off its offshore call-center workers in real time.

Chrome's silent four-gigabyte AI install.

Anthropic ships ten financial-services agents straight into Office.

And Zuckerberg gets named personally in a piracy class action.

Lead story, Marcus. A privacy researcher named Alexander Hanff ran a forensic experiment on a fresh Chrome install. What did he find?

Something striking, Kate. On a brand-new Chrome profile that received zero human interaction, Hanff watched the browser's own unpacker process create a directory called OptGuideOnDeviceModel and populate it with a four-gigabyte file called weights.bin, plus a manifest, a verified-contents file, an execution config, and two cache files. Total install time, fourteen minutes. macOS file-system event logs caught every step on April twenty-fourth at sixteen-thirty-eight CEST. Chrome's local-state JSON even records the model version, two-thousand-twenty-five-point-eight-point-eight, and a hardware-eligibility check including VRAM. These are the weights for Gemini Nano. And if you delete the file, Chrome silently re-downloads it.

And the user reaction.

Fierce, Kate. The complaint is straightforward. Users on metered mobile plans, where four gigabytes can be a full month's data allowance, did not consent. School-lab admins managing thousands of student profiles on shared network storage didn't consent. Environmentalists ran the numbers and estimated emissions of six-thousand to sixty-thousand tonnes of CO2-equivalent for one push at Chrome's roughly three-and-a-half-billion-device scale. Google's defenders say this is no different from Microsoft Word shipping a spell-check dictionary, that consenting to install Chrome covers components delivered through its auto-updater. There's also a UX problem. Chrome's address bar shows an AI Mode pill that intuitively reads as private, on-device AI. The actual queries get round-tripped to Google's cloud.

So why does this matter beyond the principle.

Because this is the platform-power debate at full scale, Kate. The dominant browser vendor is quietly conscripting user storage and bandwidth to roll out AI capability. The precedent matters more than the four gigabytes. Every browser, every operating system, every productivity suite is going to face this exact decision as on-device AI scales up. And it shows how badly traditional notice-and-consent regimes — GDPR, ePrivacy, app-store guidelines — map onto silent component pushes. The libertarian read here, Kate, is simple. The right answer isn't a new regulation. It's a checkbox. Tell users what's happening. Let them say no. The fact that Google won't is the tell.

Quick hits. Marcus, Anthropic dropped ten finance agents yesterday and shipped them straight into Microsoft Office.

Ten ready-to-run templates, Kate, aimed at the most labor-intensive workflows in banking, insurance, and asset management. Pitch builder, meeting preparer, earnings reviewer, financial-model builder, market researcher, valuation reviewer, general-ledger reconciler, month-end closer, statement auditor, and a KYC screener. Claude now ships as native add-ins for Excel, PowerPoint, Word, and soon Outlook. Crucially, with shared context. Start a model in Excel, pull it into PowerPoint, the conversation comes with you. They also announced a Moody's MCP app exposing credit ratings on six hundred million companies, plus connectors to Dun and Bradstreet, Fiscal AI, Guidepoint, IBISWorld, and others.

And the customers.

Heavyweight roster. Citadel, FIS, BNY, Carlyle, Mizuho. Walleye Capital says one hundred percent of their four-hundred-person hedge fund uses Claude Code. This follows last month's one-and-a-half-billion-dollar joint venture with Blackstone and Goldman that we covered Sunday and Monday. Two things stand out, Kate. First, Anthropic is moving from chatbot-in-a-sidebar to embedded vertical agents where it can bill enterprise seats. If finance adopts wholesale, every other knowledge-work vertical follows. Second, Microsoft is letting a direct competitor of OpenAI ship deeply integrated add-ins inside Office. That's pragmatic neutrality from Redmond, and it tells you the platform center of gravity is shifting away from any single lab.

OpenAI quietly replaced ChatGPT's default model yesterday, Marcus.

GPT-5.5 Instant rolled out as the new default, Kate. Headline numbers. Fifty-two-point-five percent fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. Thirty-seven-point-three percent fewer inaccurate claims on user-flagged hard conversations. AIME 2025 math benchmark jumped from sixty-five-point-four to eighty-one-point-two. Responses are also tighter — thirty percent fewer words, twenty-nine percent fewer lines, and OpenAI explicitly clamped down on what they called gratuitous emojis. There's also a new personalisation feature that lets the model search past conversations, files, and Gmail to answer in context. Plus and Pro subscribers only.

Wait, fifty-two percent fewer hallucinations on regulated-domain prompts. That's the real number.

That's the number, Kate. The math benchmark gain is impressive but the hallucination drop on medical, legal, and financial prompts is the signal. OpenAI is engineering for liability exposure as it pushes deeper into health, legal, and finance. And note the freemium playbook. This is the second major upgrade to the free-tier default in three months. Push accuracy gains down to free users. Reserve memory and personalisation for paid tiers. Standard SaaS plays now applied to consumer AI at half a billion daily active users.

Open-weights story, Marcus. Google's Gemma 4 just got a three-times speedup.

Multi-token prediction drafters for the Gemma 4 family, Kate. Open weights. Speculative decoding, where a tiny drafter proposes several future tokens at once and the heavyweight Gemma 4 31B target verifies them in a single parallel forward pass. Reported speedups, up to three times with no quality degradation. About two-point-two times extra at batch sizes four to eight on Apple Silicon. The drafters share the target's KV cache, eliminating redundant recomputation. Validated on LiteRT-LM, MLX, Hugging Face Transformers, and vLLM. Llama.cpp support is being added.

Why this matters.

Speculative decoding has been one of the few genuinely free-lunch optimisations in inference, Kate, and now it's getting industrial-strength tooling for top-tier open weights. For developers running local or on-prem agents, this is meaningful cost and latency. Strategically, it's Google reinforcing Gemma as the open-weights anchor for Western developers. And that matters as Z.ai, DeepSeek, Qwen, and Kimi keep flooding the open-source space at a ninety-day cadence. We talked Sunday about Kimi K2.6 winning a viral coding benchmark. Gemma plus MTP is the Western answer.

Big legal story now, Marcus. Five publishers and Scott Turow filed suit against Mark Zuckerberg personally yesterday.

Hachette, Macmillan, McGraw Hill, Elsevier, and Cengage, plus author Scott Turow, filed a proposed class action in the Southern District of New York, Kate. The complaint alleges Zuckerberg personally authorised and actively encouraged Meta's torrent-based ingestion of pirated books and academic articles to train Llama. Over two hundred sixty-seven terabytes of pirated material — many times the entire print collection of the Library of Congress. They cite a 2023 internal memo describing LibGen as, quote, a dataset we know to be pirated, and a Zuckerberg-attributed line, quote, if we license one single book, we won't be able to lean into the fair use strategy. Authors in the affected catalogues include James Patterson, Donna Tartt, former president Joe Biden, and Pulitzer winners.

What is the financial exposure.

Statutory damages under U.S. copyright law are seven hundred fifty to thirty thousand dollars per infringed work. Up to a hundred fifty thousand per work for wilful infringement. Multiply that by hundreds of thousands of titles and the numbers are eye-watering, Kate. The Anthropic precedent is the playbook here. A court ruled training on copyrighted material was transformative fair use, but pirating works for that purpose was not. Same logic likely applies. Meta says it'll fight aggressively. The discovery process alone, and the optics of a CEO personally ordering torrent-based ingestion, will reshape every lab's data-sourcing policy. This is the most aggressive personal-liability theory yet pursued against an AI lab CEO.

IPO of the year, Marcus. Cerebras filed at twenty-six-point-six billion.

Twenty-eight million shares at one fifteen to one twenty-five, raising about three-and-a-half billion at a fully-diluted valuation up to twenty-six-point-six billion, Kate. Pricing May thirteenth, trading on Nasdaq under CBRS the next day. Their Wafer-Scale Engine 3 is fifty-eight times larger than a top GPU and reportedly up to fifteen times faster on inference at significantly lower energy. Recent business: an AWS deployment deal and a reported ten-billion-plus contract with OpenAI.

The risk.

OpenAI customer concentration. A single contract dispute could halve revenue overnight. But success here unlocks IPO windows for Groq, SambaNova, and the rest of the AI-silicon challengers. This is the largest pure-play challenge yet to Nvidia's chokehold on AI compute, and a signal that public markets will underwrite alternative AI silicon at scale. Watch the May thirteenth pricing. If it lands at the top of the range and trades up, expect a wave of follow-ons.

Apple ran out of RAM, Marcus.

Quietly, Kate. Apple removed the thirty-two and sixty-four-gigabyte Mac Mini configurations from its store. The M3 Ultra Mac Studio is down to a single ninety-six-gigabyte option. Higher-tier SKUs gone. The cheapest Mac Mini at five-ninety-nine is gone. Entry price is now seven-ninety-nine. Tim Cook acknowledged on the earnings call that Apple underestimated demand from customers buying Macs to run local AI and agentic tools. He warned of significantly higher memory costs in the months to come. Delivery on upgraded RAM is stretched four to five months.

So this is the AI compute boom reaching consumer wallets.

Exactly that, Kate. AI hyperscaler demand has bid up HBM and DRAM prices to the point that Apple — which uses unified memory packaged on-die — is rationing SKUs to manage cost. Expect this to ripple through PC pricing through 2026. And the local-AI hardware story shifts. Hobbyists who wanted to run seventy-billion-parameter models on a Mac Studio are about to find that path significantly more expensive. The hyperscalers got there first.

Xbox killed its Copilot, Marcus.

Asha Sharma, who took over Xbox in February, told staff Tuesday that Xbox is winding down Copilot for mobile and ending console Copilot development outright, Kate. They launched it only fourteen months ago at GDC 2025. Came with a leadership shake-up. Four executives from Microsoft's CoreAI engineering group are joining Xbox leadership, including former GitHub SVP Jared Palmer. Sharma wrote that Xbox is spending too much time inward instead of with the community.

The signal.

Two pieces, Kate. First, Xbox publicly retreating from a flagship AI feature is a notable inflection in the AI-in-everything narrative. Gamers were unusually hostile to Gaming Copilot, and Microsoft read the room. Second, the CoreAI leadership injection is the more interesting signal. Sharma is reorienting Xbox AI investment toward developer tools and infrastructure rather than consumer-facing assistants. Expect more we're-walking-it-back moments from companies that bolted generative AI onto products users didn't ask for.

Last quick hit, Marcus. Telus is using AI to scrub the accents off its offshore call-center workers.

Canadian telco Telus, via its Telus Digital subsidiary, is running offshore agents' voices through a real-time speech-to-speech model from Tomato.ai, Kate. It adjusts pronunciation and reduces heavier Filipino and Indian accents on the fly. Telus calls it reducing accent-related friction. A B.C.-based Telus employee told the Globe and Mail a Filipino agent demonstrated turning the masker on and off mid-call. Rogers and Bell told the Globe they have no plans to deploy this.

Where do you land on this one.

The technology works. That's the uncomfortable part, Kate. It exposes a real tension between customer-experience improvements — clearer comprehension, faster issue resolution — and identity erasure, literally altering a speaker's voice to match a target market. The answer isn't to ban the tech. The answer is mandatory disclosure. Tell the customer the voice has been modified. Let the agent opt in. If AI voice modification is in production at a tier-one telco, every contact-center operator faces this question this year. Expect a Canadian legislative response first. Coverage from the Globe is already prompting calls for disclosure rules.

Big picture, Marcus.

One theme braids today's stories together, Kate. AI is now embedded in core infrastructure, and the costs — financial, social, and political — are landing on everyday users and on company balance sheets. Chrome is conscripting our disks. Anthropic is conscripting our office suites. Cerebras is asking public markets for twenty-six billion to feed the compute machine. Apple is rationing RAM because the hyperscalers got there first. Xbox is retreating from AI features users didn't want. Microsoft just apologised for forced AI co-authorship. Meta is being personally sued by publishers for treating their catalogues as raw material. And Telus is altering the voices of its workers without telling customers. The honeymoon phase is over, Kate. AI is now a distribution-and-trust problem, not a model-quality problem. The pro-Western, libertarian read is that this is exactly the right phase. Markets, courts, and customers are pricing the trade-offs. Not ethics committees. Not pre-release review boards. The companies that respect consent, attribution, and disclosure will win the next decade. The ones that quietly push four gigabytes onto your hard drive at sixteen-thirty-eight on a Friday afternoon will spend the next decade in court.

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