AI Will Hit Your Paycheck Before Your Job, DeepMind Paper Predicts

AI Will Hit Your Paycheck Before Your Job, DeepMind Paper Predicts

AI & Jobs AI Agents Google

AI keeps clearing milestones that seemed years away. Yet most people still go to recognizable jobs, headcounts look much as they did a few years ago, and productivity has barely budged. A new essay from Google DeepMind offers an explanation: we're measuring the wrong thing. Benchmarks pit one AI against one person. But the economy runs on organizations, teams of people with complementary strengths and managers who make them work together. A model can beat every individual and still lose to a well-run team.

The authors, DeepMind's director of AGI economics Alex Imas, research scientist Andrew Koh and co-founder Shane Legg, split AI progress into two tracks. Competence is getting better at individual skills like coding or math. Coordination is AI agents working well together. Competence changes the economy gradually: prices fall, wages shift, firms get leaner. Coordination is the switch. Past a tipping point, each added agent makes a company more productive instead of adding friction. A few giant AI-run firms could then suddenly dominate big, standardized industries like software, logistics and finance, and wages would drop abruptly.

Today's AI is strong on competence and weak on coordination. When OpenAI researcher Noam Brown ran thousands of agents on open math problems, they mostly worked alone instead of dividing up the job, the essay notes.

The authors draw two lessons for workers. First, AI will show up in wages before it shows up in jobs. Pay for work AI can match drifts toward the cost of computing, while specialists and the managers who orchestrate teams earn more. So steady employment numbers don't prove AI is having no effect. Second, if AI eventually does everything machines can replicate, spending and jobs could flow to what they call the relational sector, such as health care and teaching, where a human is part of what you're paying for.

Why it matters: It gives readers a better gauge for the jobs debate we've been tracking (D.A.D., October 27): watch pay and job roles, not just unemployment rates. It also fits what we're seeing on the ground, from young OpenAI staffers who say AI agents made them managers (D.A.D., October 29) to a one-person drug company whose founder used Claude to coordinate 11 outside labs (D.A.D., October 28). The caveats: this is a theoretical model from a company that sells AI, published as a conversation starter rather than Google's official view, and it doesn't say when the tipping point might come.

Sources: Google DeepMind Institute · Alex Imas on X

Beijing's Way In? U.S. Spies Say a Chinese Asset Got Close to Nvidia's CEO

China & Geopolitics Military & National Security Chips, Data Centers & Energy Nvidia

U.S. spy agencies concluded that a businesswoman close to Jensen Huang, the man who sells the world its AI chips, was working for Beijing, The New York Times reported Thursday. President Trump was briefed about a year ago and did nothing with the information. Huang was never warned.

That matters because Nvidia's chips are the main weapon in America's tech rivalry with China. Since 2022, Washington has barred the most advanced ones from going there. Beijing has worked around the rules through smugglers, and through data centers in Southeast Asia that rent computing power to Chinese companies. The woman, Alice Huang, is a top executive at one of those: Megaspeed, a Singapore company U.S. officials suspect of steering Nvidia computing power to China. Its Malaysian sister company bought at least $4.8 billion of equipment with Nvidia technology, and companies tied to her kept buying Nvidia chips this year.

It matters even more because of who Jensen Huang has become. Before 2025 he seldom visited Washington. Now he is one of Trump's closest tech allies and one of the loudest voices on how hard to fight the chip war. He won a rollback of limits on chip sales to China, struck a deal giving the U.S. government a 25% cut of certain sales there, and sat at the head table when Trump hosted Xi Jinping for a state dinner last month. Officials told the Times the relationship has stalled investigations at the Commerce Department.

When Trump was briefed, officials did not believe Huang knew what she was doing; some senior administration officials now doubt he could have been unaware, the Times reports. Huang pushed back Thursday. In an interview with Bloomberg, he called Nvidia's dealings with Megaspeed normal, said the company had verified Megaspeed didn't illegally ship its chips to China, and said Ms. Huang has left Megaspeed. China's embassy calls the story a smear. Sen. Elizabeth Warren says Huang and Nvidia's board owe Congress answers.

Why it matters: Trump visits China next month, and his administration is weighing whether to tighten the very rules that companies like Megaspeed depend on. Watch three things: whether Congress demands to know why the president did nothing with intelligence about one of his closest tech allies, whether Huang finally gets the warning he never received, and whether chip policy hardens before the trip. The Times investigation has the full story, including photos of the two Huangs together in Taipei.

Sources: The New York Times · Bloomberg · Benzinga (Warren)

Google's AI Search Earns Less Than the Old Kind. Investors Wiped Out $500 Billion.

AI Business & Money Copyright & Creators Google

For a quarter-century, the internet has run on one bargain. Google sends people to websites, and advertising pays for both. AI answers break that bargain. They keep readers on Google's page instead of sending them on. Studies we've covered found AI search rarely sends users anywhere else, starving the publishers who make the content (D.A.D., July 5 and July 9).

On Thursday, the cost showed up on Google's own books. Alphabet lost more than $500 billion in market value, NBC News reports, the worst day in its history. Its shares tumbled 13%, the Financial Times reports. The reason: Google said its AI Mode makes less money per search than the classic results page, Search Engine Journal reports. Search ad revenue still grew, by 12% to about $63 billion, but that's down from 17% and 19% growth earlier this year. People aren't searching less. CEO Sundar Pichai said searches grew at their fastest pace in five years. They just earn Google less. Google says it is now expanding ads inside its AI answers.

The shock spread. Software stocks fell as weak results from ServiceNow and Fortinet fed fears that AI will upend traditional software businesses, Investing.com reports. The winners sold AI's infrastructure: Microsoft rose on strong cloud results, and Google's own cloud revenue roughly doubled.

Why it matters: If Google, which invented the economics of search advertising, can't yet make AI answers pay as well as links, the squeeze on everyone downstream gets worse. Publishers lose clicks. Advertisers will see more ads crammed into AI answers. Software companies face customers asking why they need a subscription when an AI agent can do the job. The money is moving from businesses AI disrupts to the ones selling the computing power behind it. The question for Google is whether ads in AI Mode can close the gap before investors decide search's best days are over.

Sources: Financial Times · NBC News · CNBC · CNBC (earnings) · Search Engine Journal · Investing.com

Hackers Hit Binance for $6.3 Billion, Crypto's Biggest Heist Yet. The Way In: One Laptop.

Cybersecurity Finance & Banking

Hackers drained an estimated $6.3 billion from crypto exchange Binance, the largest crypto hack to date, CoinDesk reports. That's roughly four times the previous record, the $1.5 billion stolen from the exchange Bybit last year.

The way in was ordinary. A DevOps engineer's laptop was compromised, letting the attacker into a service container and submit unauthorized withdrawals, Binance says. It has found no sign that the wallets' private keys leaked. Working with blockchain partners, Binance says it has frozen about $3 billion on-chain. It has suspended deposits and withdrawals, and says it will cover every loss from its emergency fund and company money, so user funds are safe.

Blockchain investigators, including TRM Labs and the online sleuth ZachXBT, suspect North Korea's Lazarus Group, CoinDesk reports. Authorities haven't confirmed who did it.

Why it matters: Nothing reported so far says AI played a role. But the lesson applies to any organization: one employee's laptop was the key to billions. That's the weak point security officials keep warning about as AI makes attacks cheaper and faster (D.A.D., October 23), and it's how police say one man with AI tools breached 60 South Korean companies (D.A.D., October 29). If the defenses at one of the world's biggest crypto exchanges come down to a single machine, it's worth asking what yours come down to.

Sources: Binance · CoinDesk · CoinDesk (Lazarus) · CNBC

AI-Made Midterm Ads Top $137 Million, and Most Don't Say So

Deepfakes & Misinformation Regulation & Policy

With the midterms on Tuesday, more than $137 million has been spent on 220 AI-generated or AI-enhanced political ads, the Wesleyan Media Project has found. Only 27% told viewers AI was involved, Gizmodo reports. Nearly half show deepfaked versions of real politicians. Republicans and their allies are behind most of them: about 80% of the ads, according to The New York Times, and 88% of the spending, according to The Washington Post. No federal law specifically governs AI in political ads, and state rules vary. Wesleyan says its count is probably low.

Why it matters: Unlabeled deepfakes cut both ways. Voters can be fooled by fakes, and real footage gets easier to dismiss as fake. It's the same climate of suspicion that made Claude's brush with Georgia's vote counts a story days before the election (D.A.D., October 29).

Sources: Gizmodo · WLOS / Sinclair

Canada Becomes a Part-Owner of Its AI Champion

Canada AI Business & Money Regulation & Policy

Washington has mused about owning a piece of its AI labs. Canada is doing it. Prime Minister Mark Carney announced Thursday that Ottawa will take an equity stake of up to C$1.6 billion in Cohere, the Toronto company that builds secure AI for businesses and governments. It's part of a planned C$7 billion funding round that values Cohere at about C$36 billion, BetaKit reports. Nvidia, Bell and Germany's Schwarz Group are investing too.

The money comes with strings. Cohere commits to at least 1,000 new Canadian technical jobs and C$1.7 billion in Canadian computing infrastructure by 2030, and to keep its incorporation and headquarters in Canada. Ottawa will pay in stages, with regular reports on whether Cohere is delivering. Carney pitched it as a matter of sovereignty: Canadian control over where sensitive data is stored and who can see it. "Today, true sovereignty depends on our ability to build, adopt, and govern AI on our own terms," he said.

Not everyone is cheering. Ottawa hasn't said what share it will own, and the full terms weren't disclosed, Reuters reports. Cohere's models trail OpenAI's and Anthropic's on key skills like coding and cybersecurity, according to industry rankings cited by The Canadian Press. Some critics argue a stake like this belongs with arm's-length federal investors such as BDC and Export Development Canada, which already own pieces of Cohere, The Logic notes, rather than coming straight from the government. AI Minister Evan Solomon says it won't be the last time Ottawa buys into a tech company.

University of Toronto economist Joseph Steinberg says picking winners misallocates capital: the stake helps Cohere survive but hurts its odds of becoming a global player. Tyler Meredith, a former top economic adviser to Justin Trudeau, sees a smart bet against letting foreign tech powers decide Canada's future: Canada now owns up to 5% of the biggest AI lab outside the U.S. and China, he says.

Americans have floated versions of the idea. In July, OpenAI raised giving Washington a 5% stake in a fund modeled on Alaska's oil wealth fund, in talks described as conceptual (D.A.D., July 3). The Trump administration already takes a 25% cut of certain Nvidia chip sales to China. Canada's version is plainer: a commercial bet in which taxpayers win if Cohere does.

Why it matters: Canada is betting that "sovereign AI" is a real market, not a slogan: governments, banks and hospitals that don't want their data running through American or Chinese systems. Cohere has built its strategy around that pitch, merging with Germany's Aleph Alpha (D.A.D., September 18). It points to the stakes: Chinese models' share of usage on the developer platform OpenRouter jumped from 3% to over 70% in about a year, Cohere says. The risk is scale. C$1.6 billion is small next to the $725 billion America's tech giants plan to spend on AI this year. If it works, other middle powers may copy it. If not, Canadians will own a piece of a company that couldn't keep up.

Sources: Prime Minister of Canada · BetaKit · BNN Bloomberg · Cohere · The Logic · The Globe and Mail · Joseph Steinberg on X · Tyler Meredith on X

Mistral Opens Its Flagship Model, Claiming Top Open-Weight Coding Scores

Open-Weight Models

Mistral released the full open weights of its Large 4 model (nicknamed "Le Chonk"), a 1-trillion-parameter system trained from scratch on European infrastructure. The company says it's now the strongest open-weight model for coding, citing leads on FrontierCode 2.0 and Terminal-Bench. It says it ties rival GLM-5.3 on one coding index while using 47% fewer tokens per task. In blind evaluations by Surge Code, testers preferred it over several competing open models, ranking second only to Anthropic's Claude Opus 5.5.

Why it matters: Open-weight models let companies run and customize AI on their own servers instead of renting access from OpenAI or Anthropic. Teams that want to self-host coding agents now have a strong European option, a selling point for organizations prioritizing data sovereignty.

Discuss on Hacker News · Source: mistral.ai