Hey friends,
Four of the most powerful people in AI said out loud last weekend that they want to slow down. That is either the most responsible thing that has happened this year, or the most expensive piece of misdirection. Today we do the math on both, and then three stories that hit your payroll, your storefront, and your wallet.
Over the weekend, Sam Altman (OpenAI), Dario Amodei (Anthropic), Demis Hassabis (Google DeepMind) and Elon Musk loosely agreed to "pace the frontier" of AI development. In plain English: they want to deliberately slow down how fast the most powerful AI models get built. Amodei laid out a three-step plan: independent third-party auditors checking the labs, regulation on US labs, and a global agreement to slow down together. The Verge has the full account.
So what: if you are a small business owner, this looks like someone else's problem. It is not. Slower frontier development is the direct input to what you pay for AI next year. If the three companies that supply nearly every AI tool you use agree to slow down, the price and pace of your upgrades is no longer set by competition — it is set by a negotiation you are not in the room for.
The uncomfortable version: critics are already calling this a cartel — that the real goal is to kneecap open-source rivals and avoid actual laws while dressing it up as safety. Nick Reese, a former DHS emerging-tech policy director, told The Verge that AI leaders "aren't the best people to lead the charge," and that the industry "needs new champions." One Anthropic researcher, Jacob Coxon, quit over it, writing publicly that neither OpenAI nor Anthropic is "acting responsibly."
So what, specifically: treat any single-vendor AI dependency as a live risk this quarter, not a someday risk. List the AI tools your business would stop functioning without, then get a working alternative for the top one. That is a two-hour job — far cheaper than discovering your vendor's incentives don't match yours after you have rebuilt around them.
💰 Your next bonus may be decided by how much you use AI
Accenture's CEO told a podcast in March that at her company, AI "is how we do work — so if you want to get promoted, you've got to do the things we do." Disney, Meta, JP Morgan and KPMG have reportedly stood up "AI leaderboards" that rank employees by how much they use AI tools. Coinbase has already fired engineers who skipped the required AI training. Against that: McKinsey reports 94% of companies report no measurable value from AI, and UK job vacancies just hit a five-year low — so people feel they cannot say no. Full story at the BBC.
So what: if you are an employee, the risk is not that AI replaces you — it is that you automate two days of your week and get paid for zero of them. One worker interviewed put it best: "The higher output simply becomes the new baseline." Before you volunteer your AI gains, get the output target agreed in writing. If you run a small team, do not copy the leaderboard. Ranking people by tool usage measures theatre, not output, and the data behind it (94% seeing no value) says the tool is not the bottleneck.
🍕 The $160,000 pizza robots that quietly stopped working
Moto Pizza in Seattle bought two pizza-making robots for $160,000. In May, the supplier, Picnic, abruptly shut down — and the robots became "basically useless" overnight because the technical support evaporated with the company. It is a pattern: Zume, Pazzi, and Basil Street have all come and gone. Bank of America's senior restaurants analyst, Sara Senatore, is blunt about why: food robots are clumsy, and "humans are very efficient at making pizza." BBC has the story.
So what: the failure was not the robot. It was buying hardware from a startup with no escrow, no exit clause, and no plan for what happens when the vendor dies. Before you sign any capital hardware deal — a robot, a kiosk, a proprietary POS — make the contract answer one question: what do I still own and operate if this company disappears next quarter? If the answer is nothing, you are not buying equipment, you are renting a dependency.
💵 India's cash paradox: digital payments boom, cash hoarding grows anyway
India's central bank now has 176 billion banknotes in circulation. It is printing 28-30 billion fresh ones a year and retiring about 21 billion — while UPI digital transactions surge toward a billion a day. The RBI's own deputy governor calls it the "cash paradox": currency in circulation keeps growing at double-digit rates even as cash's share of transactions falls. For scale, the US had about 56 billion dollar bills and the eurozone about 30 billion euro notes at the end of last year. BBC explains it.
So what: apps are replacing cash as a payment method, but not as a store of value or a hedge. That is directly relevant if you sell abroad or plan to: the assumption that a market "goes cashless" is wrong. Physical money sits in the drawer as savings even in markets with a billion digital payments a day. If your expansion plan assumes digital-only consumers, the numbers say you will miss a meaningful slice of demand.
🧾 Claude can now watermark its own writing
Anthropic shipped a text-watermarking feature that marks text generated by Claude so it can be identified as AI-written. If you publish AI-assisted content — client reports, listings, articles — watermarking changes what "AI-assisted" means: the text may carry a signal you did not choose to add. Anthropic's announcement.
So what: if your work product goes to a client or a regulator, check now whether your AI vendor embeds provenance marks, and disclose it before someone else detects it.
⚛️ NVIDIA opens CUDA-Q for fault-tolerant quantum
NVIDIA expanded its open-source CUDA-Q platform specifically for fault-tolerant quantum computing — the kind that actually works rather than the kind that errors out. NVIDIA newsroom.
So what: nothing to buy yet. But quantum arriving on the same tooling as your GPU stack means the encryption on your customer data has a countdown attached. Ask your IT provider when they last reviewed your encryption for a post-quantum world — the answer is usually "never."
🎬 Video generation got 2x faster on open models
MiniMax says its H3 open video model now generates 14.4 seconds of 768p video in nine seconds — faster than real time — using SGLang and VDN on eight B200 GPUs, with no measured quality loss. MiniMax's post.
So what: if you pay for video production — ads, product demos, training clips — the open-model option just crossed from "bad and cheap" to "good and cheap." Price your next quote against it before you sign.
Picture a highway with no speed limit and no police. The five fastest drivers pull over and announce they have agreed to drive a bit slower — for everyone's safety. Sounds great, until you notice two things: they wrote the agreement themselves, and nobody can stop them from ignoring it.
That is the shape of "pacing the frontier." The frontier is whatever the most capable AI model can currently do. Pacing it means deliberately not pushing past that point as fast as you technically could. Amodei's plan asks for three things: outside auditors, laws covering US labs, and a global deal to slow down together. The odd part: competitors agreeing to limit their own competition is normally exactly what antitrust law exists to stop.
So the safety framing and the cartel framing describe the same event. Both can be true: the people closest to the risk may genuinely be scared, and they still have an obvious commercial interest in making it harder for anyone else to catch up. That ambiguity is not a reason to dismiss it — it is the reason to ask who benefits from the specific rule being proposed.
🧠 Trivia
India's central bank is retiring roughly 21 billion banknotes a year while printing 28-30 billion new ones. How many banknotes are in circulation there right now?
176 billion — even as UPI digital transactions approach a billion a day.
AI for Everyday Business — practical AI, no jargon.
Forwarded this? Subscribe here
Sent by AI for Everyday Business.