Got this newsletter from a friend? Subscribe for free →

Happy Wednesday morning! One minute, the neocloud panic had everyone expecting the AI bubble would burst any second. The next minute, Anthropic’s public IPO filing, with its surprisingly strong revenue and growth numbers, had everyone making deals again. If the Jim Cramer rule is to be believed (that is: he’s always wrong), then right now you should believe the AI business is back on track.

But the truth is probably somewhere in the middle, and the march of AI news shows no sign of stopping either way.

→ Google DeepMind’s new world model
→ What Anthropic’s life sciences team has been up to
→ Donald Trump wants to ban the phrase “artificial intelligence”

...and all the AI news we couldn’t fit here.

The news: Google DeepMind just released Genie 4, the latest in its line of world models. You sort of have to try it—or watch one of Google’s demos—to understand it. It’s a little like a model that generates FPS video games, and a little like a video generator like Sora, and a little like neither of those. But it’s… certainly something. Genie 3, released last year, generated silent interactive environments and characters that looked like animation. The new version appears to nail photorealism, and the characters talk on command. 

The takeaway: Yann LeCun, formerly of Meta, but now of Advanced Machine Intelligence Labs, believes the fabled AGI won’t be possible without world models. Google thinks so too, writing in its blog post about Genie 4, “Understanding and simulating the world is a crucial aspect of building AGI.” 

The news: If you’re like me, when you picture an LLM being used for drug discovery, you picture something a bit like a hacking model—a crude, brute-force attempt to find promising molecules, which can then be handed over to human scientists for testing. A new blog post about the efforts of Anthropic’s life sciences arm helpfully zooms in a bit by providing case studies. For instance, in one, a team of researchers aiming to make a blood cancer treatment work on solid tumors used Claude Code to propose shorter genetic code in the cells used for the therapy. One promising code is 70% shorter, and 10 times as effective by one measure.

The takeaway: Some case studies about early, preclinical work are a reminder that LLM-generated wonder drugs still very much do not exist. But Claude isn’t just spitting out random molecules, and the breadth of what’s being tried is, at the least, very intriguing.  

The news: Yep, he’s still at it. President Trump’s latest entry in the “Super Intelligence” rebrand saga was a Truth Social post in which he said he was on the verge of “BANNING” the phrase “artificial intelligence” for federal employees. He has already issued an executive order making Super Intelligence standard phrasing in official federal documents. But apparently he’ll go even further to keep this story going. 

The takeaway: Tech CEOs had started to embrace Trump’s phrase, but even his biggest cheerleader, Elon Musk, has slipped and said “AI” recently. This is a reminder that, yes, somehow, the president still cares about this. So if you’re an AI business leader who made an unwise promise to Trump about saying “Super Intelligence,” you can’t relax yet, sorry.

She Trusted ChatGPT With Her Medication Decisions. It Cost Her Everything.

Silicon Valley parents want the tech they help build out of their children’s schools

Research Sheds New Light On DeepSeek’s Safety Risks

California cities push back on data centers with bans, moratoriums

See How Anthropic Arrived at Its Mind-Boggling $40 Trillion Market Opportunity

In China, the AI bubble has already burst

Inside the Billionaires’ War Against California’s Proposed Wealth Tax

Amazon’s anti-tokenmaxxing memo says engineers should be pickier about using AI

Amazon tells software engineers what really matters as AI ushers in the era of the 'Renaissance Developer'

A bust-up between mathematicians is fueling fears that AI could undermine science