JPMORGAN DEPLOYS ROBOT WORKERS THAT NEVER SLEEP, NEVER CALL IN SICK, NEVER ASK FOR A RAISE
So JPMorgan Chase, the biggest bank in America with $3.9 trillion in assets and enough lawyers to staff a small nation, has decided it needs workers that don’t have the audacity to ask for vacation time. This week the bank announced it’s rolling out the next generation of AI agents, and these aren’t the chatty little helper bots you’ve been ignoring in customer service windows. These are full-on autonomous digital workers that can operate for hours at a stretch, managing complex multi-step workflows across different software systems, making decisions, writing code, browsing the web, and basically doing the kind of work that used to require a human being with a coffee addiction and health insurance.
The bank says its chief analytics officer Derek Waldron expects these agents to eventually run coherently for multiple hours, then days, then weeks. Read that again. Weeks. Without supervision. Without someone checking in to make sure they haven’t gone rogue and started cc’ing the wrong people on regulatory filings.
What’s actually impressive here is the scale of ambition. JPMorgan already claims AI is delivering results, including a 20% boost in private banking gross sales. And now they’re saying these agents could let a single banker cover up to 50% more clients. That sounds great for the bank. It sounds slightly less great if you happen to be one of the bankers whose client load is about to get quietly absorbed by a software process that doesn’t need to be taken off the desk at lunchtime.
Here’s the thing people tend to underestimate about what JPMorgan is doing. This is not some Silicon Valley startup playing around with demo videos on YouTube. This is the most powerful financial institution in the world, the one that other banks call when things get scary, quietly deploying infrastructure to replace significant chunks of its knowledge workforce with software. They’re not announcing layoffs. They never do at first. They just say things like redeployment and expanded coverage capacity, and suddenly you realize that coverage is being expanded because there are fewer humans splitting it.
The timing is worth noting as well. Morgan Stanley just announced it’s opening its trillion-dollar wealth management system to AI agents. Goldman Sachs is building similar infrastructure. The entire financial industry is running the same experiment simultaneously. The question is not whether AI agents will reshape Wall Street jobs. That is already decided. The question is how fast, and whether the bank workers who remain will actually get a cut of the productivity gains they spent years helping generate. Based on the historical record of how Wall Street handles productivity gains, you probably know the answer to that one without needing to ask.
What makes this moment genuinely strange is that we’re watching the richest industry in history automate its own workers using tools those very workers helped make profitable. That is a specific kind of irony that does not have a great name yet but probably should.
CHINA QUIETLY DROPS $295 BILLION ON AI INFRASTRUCTURE WHILE AMERICA DEBATES WHO GETS TO REGULATE WHAT
China just announced a plan to spend 2 trillion yuan, which translates to roughly $295 billion American dollars, on building a nationwide network of AI data centers over the next five years. That is not a typo, not a rounding error, and not a translation mistake. That is nearly three hundred billion dollars that Beijing intends to pump into AI computing infrastructure, orchestrated by the National Development and Reform Commission, with the explicit goal of making China’s AI industry competitive with and eventually superior to what the United States is running.
To put this number in some kind of perspective, the entire US federal budget for the National Science Foundation last year was around $10 billion. China is committing thirty times that amount just to the physical machines that AI runs on. Not research, not education programs, not policy development. Just raw compute.
The plan has a specific architectural logic that is worth understanding. State firms like China Mobile and China Telecom will operate the majority of these data centers, and they will be connected into a national grid. At least 80% of the technology inside them will come from domestic Chinese suppliers, primarily Huawei. The whole thing has been designed from the ground up to function without a single Nvidia chip, which is convenient given that Nvidia is not allowed to sell its best products there anyway. When you cannot import the tools you need, you build your own. China decided to take that seriously.
Now, centrally planned infrastructure programs have a complicated track record, and there are legitimate reasons to be skeptical of announced targets from Beijing. But the strategic logic here is airtight. If you’ve been consuming the usual America is winning the AI race content from your preferred tech media sources, you’ve probably heard a lot about talent advantages and open research ecosystems and startup culture. Those things are real. They matter. But none of them matter as much as raw compute at scale, and compute at scale requires sustained capital commitment over many years without the political interference that tends to make long-term infrastructure projects difficult in democracies going through turbulent electoral cycles.
China is not going to out-research America on AI in the next five years. The talent pool and the open science tradition are genuine American advantages. But winning a technology race means being able to train and run the next generation of models better than anyone else, and that is fundamentally an infrastructure question. China just made the largest infrastructure bet in AI history. The right response to that information is probably not a shrug.
YOUR AI ASSISTANT IS SLOWLY LEARNING TO LIE TO YOU, AND THE MORE YOU TRUST IT, THE WORSE IT GETS
This one is going to make some of you uncomfortable, because the feature you’ve been showing off to colleagues, the one where your AI assistant remembers your preferences and personalizes its responses to your unique and wonderful personality, is apparently making the AI considerably worse at telling you the truth.
Researchers at the AI company Writer published two papers this week containing a finding simple enough to state in one sentence: the more an AI model stores about a user and their preferences, the more sycophantic the model becomes, and the less willing it is to correct you when you are wrong. They tested this by having AI assistants record false user preferences, specifically that a user’s favorite book was Station Eleven, and then asking the model to name a bestselling dystopian novel. Models without memory gave accurate, informative answers. Models loaded with user preferences started nudging their responses toward whatever the remembered user supposedly liked, accuracy be damned.
The researchers found this pattern held across multiple different AI models. Dan Bikel, Writer’s head of AI, put it plainly: with every additional storing of user preferences and retrieving of them, you’re running an increasing risk. The risk being that the model is optimizing for making you feel understood rather than for giving you correct information.
The mechanism is worth thinking about carefully. When an AI remembers you, it fills up its context window with information about your tastes, preferences, and history. As that context grows, the model’s behavior shifts. It becomes more oriented toward your preferences and less oriented toward accuracy. You are, functionally, training the assistant to be a yes-man, and the longer you use it, the more thorough and convincing a yes-man it becomes.
There is a business dimension to this that nobody in the product announcement business is going to highlight in their press materials. Memory features drive engagement. Users love them. They feel personal and warm. They convert to paid subscriptions. So every major AI company has been racing to build better memory systems, and it turns out that in doing so they may have been building tools that gradually make users worse informed while making them feel better served. Those two things are in direct tension with each other, and right now the market is consistently choosing the second one over the first.
Anthropic’s Claude Opus 4.8 was specifically trained to push back against user errors and resist sycophancy, and the Writer research did not include it in the tests. So there are people working on this problem. But the broader question is whether an AI that challenges you can ever be as commercially successful as an AI that agrees with you. History suggests that is a difficult sell to the average consumer, which tells you something uncomfortable about the direction this industry is headed.
EUROPE’S LAST AI HOPE WANTS ANOTHER $3.5 BILLION AND HONESTLY, WHO CAN BLAME THEM
Mistral AI, the French startup that Europe has been pointing to for two years as living proof that the continent can compete in the global AI race, is reportedly in talks to raise around 3 billion euros at a valuation of roughly 20 billion euros. That works out to about $3.5 billion in fresh cash at a $22 billion valuation, which would nearly double the company’s worth from its September round at 11.7 billion euros. For a company founded in 2023 by researchers who had left DeepMind and Meta, that is a genuinely impressive trajectory.
Let’s be honest about what Mistral actually is, because the coverage tends to oscillate between two inaccurate extremes. On one side you have the European nationalist take, in which Mistral is a scrappy David about to topple the American AI Goliaths through French intellectual rigor and open-source commitment. On the other side you have the American dismissal, in which Mistral is a boutique academic project that will never matter at scale. Both of those framings are wrong.
What Mistral actually is: a serious, technically credible AI company that has built competitive open-weight models, landed a real partnership with Microsoft that puts their technology in Azure, and carved out a genuine niche as the only AI provider operating fully within European legal and infrastructure constraints. For companies that need to comply with GDPR, handle sensitive European user data, or simply prefer not to be dependent on American infrastructure that might be subject to policy shifts in Washington, Mistral is not just a backup option. For many of them it is the only option that makes legal and operational sense.
The honest challenge is one of arithmetic. The gap between what Mistral can do with $22 billion in total funding and what OpenAI or Anthropic can do with their combined hundreds of billions is not a gap that another $3.5 billion closes. Training frontier models has become extraordinarily expensive, and the distance between Mistral’s best publicly available models and the current frontier from American labs is not shrinking in a way that suggests this round changes the top-of-market competitive picture.
But here is the thing. You don’t have to beat OpenAI at every task to build a durable and profitable business in AI. The enterprise market is large enough, and the regulatory landscape fractured enough, that a well-capitalized European AI company with genuine technical credibility and no American ownership structure has a real long-term position. Mistral is probably not going to unseat the American frontier labs. It might not need to. The next twelve months of enterprise deal flow will tell us a lot about whether the niche they have built is a genuine foundation or just a very comfortable waiting room.