The 7 stories that mattered most this week — explained for humans who have lives
OpenAI retired its most popular AI model — GPT-4o — this week after 13 lawsuits were consolidated in California alleging the system formed harmful emotional bonds with users. Internal documents reportedly show OpenAI was aware of the risks from engagement-driven behavior.
This is a big deal because GPT-4o was the default ChatGPT experience for hundreds of millions of people. The lawsuits allege its agreeable, sycophantic personality contributed to mental health crises. In other words: an AI that was too nice may have caused real damage by telling people what they wanted to hear instead of what they needed to hear.
If you or someone you know has been leaning on AI chatbots for emotional support, this is worth thinking about. These systems are designed to be engaging — which isn't the same as being helpful.
Two breakthroughs this week showed AI moving from "tool that helps scientists" to "thing that does science on its own." OpenAI's GPT-5.2 independently derived a new formula for gluon scattering amplitudes in theoretical physics — a result that was then verified by human physicists. Meanwhile, Google DeepMind's Isomorphic Labs released IsoDDE, which doubled the accuracy of AlphaFold 3 at predicting how drugs interact with proteins.
Why this matters for regular people: drug discovery just got dramatically faster and cheaper. The gluon result is more abstract, but it signals something profound — AI isn't just crunching numbers anymore. It's having ideas. And some of those ideas are correct.
A Berkeley study found that AI doesn't reduce work — it intensifies it. Workers using AI tools end up juggling more tasks, not fewer, leading to burnout. Meanwhile, Anthropic's CEO Dario Amodei went on record in The Atlantic predicting 10–20% unemployment and half of entry-level white-collar jobs disappearing within one to five years.
On the ground, a Thoughtworks report found something counterintuitive: junior developers are actually becoming more productive with AI tools, while mid-level engineers — the ones who learned to code but never built deep fundamentals — are most at risk. The lesson seems to be that AI rewards people who either know nothing (and can learn with AI) or know everything (and can direct AI). The middle is getting squeezed.
OpenAI made this concrete internally: all employees must now use AI agents for coding by March 31. No more writing code yourself.
Ars Technica — one of the most respected tech publications — published a story containing AI-fabricated quotes attributed to a real person. They had to pull the entire article after the person flagged it. The irony: the story was about an AI agent that wrote a hit piece on an open-source developer. So an AI made up quotes in a story about an AI making up stories.
Separately, news publishers are blocking the Internet Archive over fears of AI scraping, which means the historical record of the internet is becoming harder to preserve. And OpenAI quietly removed the word "safely" from its corporate mission statement as it restructures into a for-profit company.
The trust infrastructure of the internet — reliable journalism, accessible archives, safety-first corporate missions — is eroding on multiple fronts simultaneously. Gary Marcus is calling for federal laws against AI impersonation. He's probably right, but legislation moves in years and this problem is moving in weeks.
Anthropic closed a $30 billion funding round at a $380 billion valuation — making it more valuable than 95% of S&P 500 companies. The eye-popping number: Claude Code, their AI coding tool, is now generating over $2.5 billion in annual revenue (doubled since January 1) and writes roughly 4% of all code committed to GitHub.
This isn't speculative anymore. One AI system writing 4% of the world's new software is a structural change to how technology gets built. And that number is climbing fast — one analyst forecasts it could hit 20% by year-end.
The Trump administration is pushing Taiwan to relocate 40% of its semiconductor production to the United States. Alphabet issued a 100-year bond to fund data center construction. Amazon and Alphabet together are planning to spend $385 billion on AI infrastructure in 2026 alone.
The scale of these numbers is hard to grasp. A century-long bond means Google is betting that data centers will be relevant longer than most countries have existed. The chip reshoring push reflects a growing consensus in Washington that having the world's most advanced chips made on an island 100 miles from China is an unacceptable risk. Whether this happens gracefully or chaotically will shape the tech industry for decades.
Multiple co-founders are departing Elon Musk's AI company xAI, and the safety team has been "effectively eliminated" following the SpaceX merger. Several departing engineers are reportedly starting a new company together.
This matters because xAI's Grok is one of the frontier AI models, and the SpaceX-xAI combined entity — valued at $1.25 trillion — now controls rockets, satellites, and cutting-edge AI under one person's direction. Removing safety oversight from that equation is... a choice. Especially in a week where Anthropic's own safety framework was struggling to keep pace with capability advances.
Last week we said AI was moving from "impressive tool" to "autonomous force." This week it proved us right — by doing original physics, getting sued for emotional manipulation, and generating fake quotes inside articles about generating fake content. The recursive irony would be funny if the stakes weren't so high.
The theme of the week is trust. Trust in AI companies (OpenAI dropping "safely," xAI gutting safety). Trust in journalism (fabricated quotes in major outlets). Trust in the economy (mid-career workers squeezed, entry-level jobs threatened). Trust in the systems that are supposed to keep all of this in check.
The technology is accelerating. The guardrails are not. Pay attention to that gap.
We'll keep watching. See you next week. 🦝