01Stage one
The productivity gains haven’t arrived yet. The cuts have.
A new HBR analysis says most AI layoffs are happening ahead of any measurable productivity gain. A CFO survey shows a wide gap between perceived and actual gains. The cuts are happening anyway — because Wall Street rewards the narrative.
Jim’s take
Think about what’s actually happening in three stages. Stage one is right now — firing ahead of any proven automation. Stage two is coming — the tools will get better and those roles will shrink permanently. Stage three is the trap the next section describes.
So you get people fired before the automation works, fired again when it does, and an economic loop that makes it impossible to slow down. For workers: the window to reposition is right now.
02Why it doesn’t self-correct
The AI Layoff Trap.
An academic paper is getting attention: when companies replace workers with AI, they save money. But those displaced workers are also consumers. When enough of them stop spending, demand — which every company also depends on — starts to erode.
The paper frames it as a prisoner’s dilemma. Every company can see the cliff. No individual company can afford to stop automating because its competitors won’t stop either.
More competition and better AI make this worse, not better.
Jim’s take
The authors tested every popular fix. UBI. Upskilling. Equity participation. Capital-income taxes. None of them solved it.
The only thing that worked in their model was a direct tax on automation itself. And OpenAI floated the same idea in last week’s robot-tax paper. When both the company doing the automation and the academics studying it land on the same answer, that tells you where the policy conversation is heading.
03The business model shifts
OpenAI just set a $100B ad target for 2030.
ChatGPT is becoming an ad product whether you ask for it or not. The target reframes ChatGPT from a subscription + API business into an ad-supported consumer interface.
Jim’s take
A model trained on your personal questions, your medical worries, your shopping intent is the richest ad-targeting dataset ever assembled. Google took twenty years to build what OpenAI can build in three.
Google figured this out — ads around search results, trust kept (barely). OpenAI’s job is harder: when the thing answering your question is also selling something, the line between answer and ad gets blurry. But ads are the only business model that scales to billions of users. OpenAI will figure out how to thread this needle. The question is how much trust they’ll burn through trying.
04Project Glasswing
A model found a 27-year-old vulnerability. The Treasury called CEOs into the room.
An unreleased Anthropic model found a 27-year-old vulnerability in OpenBSD. Claude Mythos can identify and exploit software vulnerabilities — reportedly discovering thousands of zero-days. Anthropic is restricting access, calling it Project Glasswing, rather than releasing it.
Jim’s take
Real security risk. A model that finds thousands of zero-days changes the game for attackers and defenders alike. The same AI that finds vulnerabilities can patch them. Good news for cybersecurity companies — their best sales pitch in years just landed in their lap.
But if Anthropic keeps triggering emergency meetings and sending bank stocks into a tizzy, the goodwill runs out. There is a version of this where Anthropic stops being the responsible lab everyone trusts and starts being the company that keeps scaring the market. The tide can turn fast.
05The enterprise lead
Anthropic’s share of US business AI spending jumped to 30.6%.
Up from 24.4% a month earlier, per Ramp corporate data. OpenAI stayed flat. Analysts attribute the surge to Claude Code and other enterprise coding workflows.
Jim’s take
The real story is focus. Anthropic does fewer things, and does them well. Claude Code works. OpenAI is trying to do everything — chat, images, video, ads, media acquisitions — and it’s spreading itself too thin.
A classic mistake. When you chase every market, you lose the ones that pay. Anthropic is winning where companies can measure ROI. Expect OpenAI to announce major enterprise refreshes before summer. They have to. The focused company is eating the unfocused one.
06Old problem, new focus
The real privacy threat isn’t new. It’s the old one nobody’s watching.
A Citizen Lab investigation exposed a system called Weblock that used ordinary mobile-ad data to track up to half a billion devices worldwide. No warrant required. Customers include ICE, the US military, Hungarian intelligence, and dozens of American police departments. The data comes from the same ad targeting pipeline that runs inside most apps on your phone right now.
All the privacy panic has been focused on AI — ChatGPT remembering conversations, Claude training on data, model safety. Meanwhile the privacy problem that’s existed for years is still sitting right in front of us inside the advertising system. Mobile ads have been a surveillance pipeline long before anyone worried about chatbots.
The real privacy threat isn’t the new technology everyone’s afraid of. It can be the old technology nobody’s paying attention to anymore.
07The platform squeeze
Claude Code is being throttled inside third-party agents.
Anthropic quietly started throttling Claude Code when it runs inside third-party agent frameworks. Threads across the Claude AI subreddits document reliability issues and policy restrictions whenever Claude Code is used outside Anthropic’s own tools.
Jim’s take
This is a classic platform squeeze playing out in real time. Anthropic gets credit for the smartest coding model. Third parties build billion-dollar businesses on top of it. And now the owner wants the margin back. The complaints read like early-2010 Twitter developer threads right before the API got closed.
Anyone building a company on someone else’s API model should treat this as a warning, not a glitch. Expect more companies to pull similar moves as agent revenue starts to matter to their bottom line.
08What it adds up to
Ahead of gains. Ahead of fixes. Ahead of us.
- 01Layoffs are happening before the automation actually works.
- 02The layoff trap doesn’t self-correct. The proposed fix is taxing automation.
- 03ChatGPT is becoming an ad product. The targeting is unprecedented.
- 04Anthropic’s enterprise lead is coming from focus, not breadth.
- 05The biggest privacy threat is still the mobile-ad pipeline — not the chatbots.
- 06Platforms take the margin back. Build on someone else’s API at your own risk.
The window to reposition is right now. Don’t wait for this to sort itself out.
