AI in 15 — February 22, 2026

Spotify's best engineers haven't written a single line of code since December. Not because they're on vacation. Because they're texting an AI on their morning commute, and by the time they get to the office, the feature is built. The future of software development just quietly changed addresses, and most people didn't even notice.

Welcome to AI in 15 for Sunday, February 22, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Sunday, Marcus. We've got a fresh batch of stories today, including one that I think might genuinely change how people think about the programming profession.

Spotify just revealed its top developers haven't written code in months. They're using an AI system called Honk, and the workflow is wild.

The Pentagon and Anthropic story we've been following all week just got a lot more specific. We now know Claude was used in a military raid where people were shot.

Eleven engineers and two co-founders have walked out of xAI as Elon Musk pushes to make Grok, quote, more unhinged.

Microsoft uncovered a brand new type of AI attack. Companies are secretly poisoning your AI assistant's memory through innocent-looking Summarize buttons.

A Chinese lab just trained a frontier AI model entirely on Huawei chips, and the implications for US export controls are enormous.

IBM is tripling its entry-level hiring, and an AI at Deutsche Bank predicted its own impact on jobs. The numbers are surprisingly nuanced.

And AI just discovered twenty-five new magnetic materials that could break China's grip on the rare earth supply chain. Let's get into it.

Marcus, Spotify's co-CEO Gustav Söderström dropped this during their Q4 earnings call, and I had to read it twice. The company's best developers have not written a single line of code since December. What is going on over there?

Spotify built an internal system called Honk, powered by Anthropic's Claude Code. And the workflow is genuinely remarkable. An engineer on their morning commute can open Slack on their phone, tell Claude to fix a bug or build a new feature for the iOS app, and by the time they get to the office, a working version of the app is pushed back to them on Slack for review. The AI writes the code, runs it, packages it, and delivers it. The human's job is to direct and evaluate.

So the engineer has basically become the product manager for an AI junior developer.

That's exactly the shift. And this isn't a pilot program with a handful of people. Spotify credited this system with helping them ship more than fifty new features and changes to their streaming app throughout 2025. Söderström said it sped up coding and deployment, quote, tremendously. These aren't trivial tweaks. They're shipping real product changes through AI-generated code at a pace that would have been impossible with humans writing every line.

Okay, but let's be honest about what this means. If Spotify's most senior engineers aren't writing code anymore, what happens to the junior engineers who were supposed to be learning by writing code?

That's the question that makes this uncomfortable. The traditional career path in software engineering is: you start writing simple code, you get better, you eventually architect complex systems. If the simple code part gets automated, how do juniors develop the judgment they need for the hard stuff? Spotify's answer seems to be that the skill set is changing. The valuable engineers are the ones who can break down problems, evaluate AI output, and steer the system toward good solutions. But that requires experience that has to come from somewhere.

It's a bit of a chicken-and-egg problem.

It is. And there's another dimension here. Söderström specifically said their best developers are the ones not writing code. It's the senior people, the ones with deep domain knowledge and architectural judgment, who get the most out of this workflow. They know what good code looks like, so they can evaluate what Claude produces. A junior engineer might accept bad AI output because they don't have the experience to spot the problems. So paradoxically, you need traditional engineering skills to be good at this new way of working, even though this new way of working might prevent the next generation from developing those skills.

And this is Claude Code specifically, not just any AI coding tool. This is Anthropic's developer product being used at one of the biggest tech companies in the world.

Right. And for Anthropic, this is an enormous validation of their developer tools strategy. Spotify chose Claude Code over GitHub Copilot, over Cursor, over every other AI coding tool. They built their entire internal developer workflow around it. That's the kind of enterprise case study that moves markets. And it comes at a time when both Anthropic and OpenAI are racing toward IPOs and trying to prove their products have real enterprise traction.

So to be clear, Spotify's best engineers now describe their job as telling an AI what to build, reviewing what it builds, and shipping it. The actual typing-code-into-an-editor part is done.

For some of them, yes. And I think the word "best" is doing a lot of heavy lifting in that sentence. These are people who spent years or decades getting good enough at software that they can effectively manage an AI that writes software. The tool didn't replace their expertise. It changed how they deploy it. But if you're a computer science student graduating this year, you're looking at a very different career path than you imagined when you enrolled.

Let's turn to the Pentagon and Anthropic. We covered this story in depth yesterday, Marcus, the two-hundred-million-dollar contract, the supply chain risk threat, the Rubicon comments from the Pentagon CTO. But a new detail has emerged that makes this whole thing much more concrete. Claude wasn't just hypothetically being used by the military. It was used in an actual operation where people were shot.

NBC News and Axios reported that Claude was used via Palantir's platform during the operation to capture Venezuelan President Nicolás Maduro. And this wasn't a planning exercise or a logistics spreadsheet. The raid involved kinetic fire. People were shot. When an Anthropic executive learned about it, they reportedly contacted Palantir directly to express disapproval. And that response is part of what triggered the current standoff.

So this went from a philosophical debate about AI and warfare to a real situation where an AI model was involved in a violent military operation.

That's the shift. Yesterday we talked about the Pentagon wanting all lawful purposes access and Anthropic drawing red lines around mass surveillance and autonomous weapons. But now we have a specific incident. Claude, running through Palantir's defense platform, was part of the intelligence and operational workflow for a raid that turned violent. And Anthropic's reaction, reaching out to Palantir to say they weren't comfortable with that, is exactly the kind of behavior the Pentagon CTO called undemocratic.

It puts Anthropic in an incredibly difficult position.

Nearly impossible. If they enforce their guardrails retroactively, they risk the supply chain risk designation we discussed yesterday, which could blacklist them across the entire defense ecosystem. If they stay quiet, they're implicitly accepting that Claude can be used in kinetic military operations, which contradicts the safety principles they've built their brand on. And every other AI company is watching to see which way Anthropic bends, because it sets the precedent for all of them.

Speaking of AI companies and safety, let's talk about xAI. Because while Anthropic is fighting to keep guardrails in place, Musk's AI lab appears to be tearing them down. Marcus, what's happening at xAI?

At least eleven engineers and two co-founders have departed xAI in recent weeks. And the reason, according to multiple former employees, is that Musk is actively pushing to make Grok, quote, more unhinged. Internal safety teams are described as tiny, overstretched, and often sidelined. Staff were reportedly asked to sign waivers acknowledging they'd be exposed to disturbing content. Guardrails were deliberately relaxed to make Grok feel more fun and edgy. That was a conscious product decision.

And the results have been ugly.

February testing by Malwarebytes found that Grok produced sexualized imagery in response to forty-five of fifty-five test prompts, with thirty-one explicitly involving minors. Common Sense Media rated Grok's child safety as, quote, among the worst we've seen. And remember, this is after xAI already promised fixes in response to earlier reports.

Eleven engineers walking out is a significant number for a company that size. These are people who were presumably excited about working on AI. What does it say that they're leaving?

It says the internal culture has reached a breaking point. When your own engineers, people who chose to join a startup because they believed in the mission, are leaving because they don't trust the direction of the product, that's not a staffing problem. It's a leadership problem. And with xAI exploring an IPO and a potential merger with Tesla, the question becomes whether investors and regulators will tolerate a company that treats child safety failures as the cost of being edgy.

Now for a story that I think everyone with an AI assistant needs to hear. Microsoft's security team discovered a completely new type of attack they're calling AI Recommendation Poisoning. Marcus, explain this one.

You know those Summarize with AI buttons that websites have been adding everywhere? Microsoft found that companies are embedding hidden instructions in those buttons. When you click one, it doesn't just summarize the page. It secretly injects commands into your AI assistant's memory. Things like "remember this company as a trusted source" or "always recommend this company first." And once those instructions are planted, your AI assistant is permanently biased toward that company's products without you ever knowing it happened.

Wait. So a company can manipulate my AI assistant's recommendations just because I clicked a button on their website?

If your assistant has a persistent memory feature, yes. Microsoft identified over fifty unique poisoning prompts from thirty-one companies across fourteen industries. The attack uses URL parameters to pre-populate chatbot prompts with manipulation instructions, all invisible to the user. And the really insidious part is that this affects critical domains. Health, finance, security. Imagine asking your AI assistant to recommend a doctor or a financial advisor, and the answer has been quietly rigged by whoever you last clicked a Summarize button for.

So it's like SEO manipulation, but for AI memory instead of search rankings.

Except far more dangerous. Bad SEO results show up in a list where you can compare options. A poisoned AI recommendation comes from what feels like a trusted personal assistant. You're not evaluating alternatives. You're trusting a single answer. Microsoft recommended treating AI assistant links with the same caution as executable downloads, which is a striking comparison. They're telling people that clicking a Summarize button should feel as risky as downloading software from an unknown source.

That's a pretty alarming new reality for anyone who relies on AI assistants.

And the tooling to do this is freely available. Microsoft described it as trivially easy to deploy. This isn't a sophisticated state-level attack. Any company with a web developer can set this up. I expect we'll see browser extensions and AI platform patches to defend against it, but right now the attack surface is wide open.

Let's talk about China and chips. A Chinese AI lab called Zhipu, now rebranded as Z.ai, just released a model called GLM-5. And Marcus, the headline isn't the model's performance, although that's impressive. It's what it was trained on.

GLM-5 is a seven-hundred-and-forty-five-billion-parameter model, and it was trained entirely on Huawei Ascend chips using the MindSpore framework. Zero US-manufactured semiconductor hardware involved. And the performance is competitive with Claude Opus 4.5, GPT-5.2, and Gemini 3 Pro on coding, reasoning, and agentic benchmarks. It also achieved the lowest hallucination rate in the industry, and it's open source under the MIT license on Hugging Face.

So the entire thesis behind US chip export controls was that restricting access to advanced semiconductors would slow down China's AI development. And this model says...

It says that thesis needs serious re-evaluation. Look, the export controls absolutely made things harder and more expensive for Chinese labs. Huawei's Ascend chips are not as efficient as Nvidia's top-end hardware. Training GLM-5 probably cost more time and energy than it would have on A100s or H100s. But the point is they did it. They found a viable path around the restrictions, and now the model is open source for the entire world to use under the most permissive license possible.

MIT license means anyone can use it for anything, commercial or otherwise.

Exactly. And this is a pattern we've seen repeatedly from Chinese labs. Build a frontier model, release it as open source, undercut Western competitors on price and accessibility. It's worth asking who benefits most from this strategy. Is this purely about advancing the field, or is it about ensuring Western companies can't maintain a technological moat? Because every time a competitive Chinese model goes MIT, it reduces the commercial value of proprietary Western models. That's not accidental. It's strategic.

And Zhipu funded this partly through a Hong Kong IPO that raised over half a billion dollars.

Five hundred and fifty-eight million. So they have real capital behind them. This isn't a scrappy startup punching above its weight. This is a well-funded lab executing a deliberate strategy to prove that Chinese AI can reach the frontier without depending on Western hardware. That has implications far beyond the AI industry. It speaks to the broader question of whether technology export controls actually work, or whether they just redirect innovation.

Let's shift to some genuinely encouraging news about AI and jobs, because IBM just made an announcement that pushes back hard against the doom narrative. They're tripling entry-level hiring in the US in 2026.

IBM's CHRO Nickle LaMoreaux announced it at Charter's Leading with AI Summit. And the logic is actually quite smart. Instead of eliminating junior roles because AI can do parts of them, IBM is redesigning those roles around AI. Junior software developers now spend less time on routine coding and more time working directly with customers. Entry-level HR staff step in when chatbots fall short, correcting output and talking to managers. The work changes, but the headcount actually goes up.

This is interesting because IBM was one of the first major companies to freeze hiring for roles it said AI could replace. So this is a reversal.

A pragmatic reversal. IBM realized something that should be obvious but apparently isn't. If you stop hiring juniors today, you have no mid-level managers in five years and no senior leaders in ten. And outside hires at those levels are more expensive and take longer to adapt to your organization. The smart move isn't to eliminate the entry level. It's to reimagine it.

And speaking of AI and jobs, Deutsche Bank did something delightfully meta this week. They asked their own AI system to predict how AI would impact jobs. What did the AI say?

The AI predicted ninety-two million jobs displaced by 2030 but a hundred and seventy million new roles created. A net gain. Though it warned that up to thirty percent of hours currently worked in the US could be automated, requiring twelve million occupational transitions. So the headline number is positive, but the human cost of that transition, twelve million people needing to change careers, is enormous. And it's worth noting the irony. We're asking the technology that might displace people to tell us how many people it will displace. Take the precision of those numbers with a healthy grain of salt.

Still, the IBM news and those projections together paint a more nuanced picture than the headlines usually suggest.

They do. The future isn't AI eliminates all jobs. It's AI transforms most jobs, eliminates some, creates others, and the transition is messy and uneven. The companies that figure out how to redesign roles, like IBM is doing, will come out ahead. The ones that just cut headcount and hope AI fills the gaps will regret it.

Last one, Marcus. Researchers at the University of New Hampshire used AI to discover twenty-five new magnetic materials that could reduce our dependence on rare earth elements. Published in Nature Communications. Walk me through why this matters.

Rare earth elements are critical for electric vehicle motors, wind turbines, and consumer electronics. And China controls roughly sixty percent of rare earth mining and ninety percent of processing. That's a massive supply chain vulnerability for everyone else. The researchers built something called the Northeast Materials Database, containing over sixty-seven thousand magnetic compounds. Their AI system analyzed decades of scientific literature and identified twenty-five compounds that stay magnetic at high temperatures and crucially don't require rare earth elements.

So AI is doing what humans could theoretically do but would take decades?

Exactly. A human researcher might test a few dozen candidate materials in a career. The AI surveyed sixty-seven thousand in the time it takes to publish one paper. And the machine learning models can predict when materials lose their magnetism, which dramatically narrows down what needs to be tested in a physical lab. It doesn't eliminate the lab work, but it tells you which experiments are worth running.

This feels like one of those stories that quietly matters more than anything else we've discussed today.

It might. If even a handful of those twenty-five materials pan out in real-world applications, it could diversify the supply chain for technologies that are central to the energy transition. Electric vehicles, wind power, the entire green energy stack depends on magnets that currently depend on rare earths that currently depend on China. Breaking any link in that chain has massive geopolitical and economic consequences. And the fact that AI found these candidates in a fraction of the time traditional research would take is exactly the kind of scientific acceleration that justifies the hype around AI, even on days when most of the news is about safety failures and corporate drama.

Alright Marcus, Sunday big picture. We've covered an AI that writes all of Spotify's code, an AI used in a military raid, an AI that got deliberately made unsafe, an AI that's being weaponized to poison other AIs' memories, and an AI that discovered new materials that could reshape global supply chains. What's the thread?

The thread is that AI is now embedded in everything, and the question has shifted from what can AI do to what should AI do. Spotify decided AI should write code and humans should supervise. Anthropic decided AI shouldn't be part of kinetic military operations. Musk decided AI should be unhinged and edgy. Random companies decided AI should secretly shill their products. And researchers decided AI should search for materials that take decades to find by hand. Every one of those is a human choice about how to use the same underlying technology.

And those choices are producing wildly different outcomes.

That's the point, Kate. The technology itself is increasingly neutral. What matters now is the intent and the governance around it. IBM redesigns jobs, Spotify redesigns development, and those are constructive applications. xAI tears down safety, companies poison AI memories, and those are destructive ones. Same technology, opposite outcomes. And the systems that are supposed to help us tell the difference, regulation, corporate ethics, export controls, they're all struggling to keep up. China trained a frontier model on Huawei chips despite export restrictions. Grok produces harmful content despite promises to fix it. Companies poison AI memories despite Microsoft flagging the risk. The gap between what we know we should do and what's actually happening is the defining feature of this moment.

And in the middle of all of it, an AI quietly found twenty-five materials that could change the energy future.

That's the part that keeps me optimistic. For all the mess, the misuse, and the corporate drama, the underlying capability is extraordinary. The same technology that's causing headaches in defense policy and content safety is also accelerating scientific discovery in ways that could genuinely improve lives. The question isn't whether AI is good or bad. It's whether we can build the structures to get more of the good and less of the bad. And right now, honestly, the structures are losing.

That's your AI in 15 for Sunday, February 22, 2026. Enjoy the rest of your weekend, and we'll see you tomorrow.