Got this newsletter from a friend? Subscribe for free →

Good Thursday morning! In our previous entry, we were impressed—if a bit puzzled—by Google’s Genie 4, the latest release of its world model. Another Google model unveiling is in the news today, and more on that in a moment. But it’s worth pausing to note that Google’s AI agenda has been evolving this year after Demis Hassabis stepped down from leadership at DeepMind, and co-founder Sergey Brin reportedly started taking an active role in AI development.

The OpenAI-Anthropic rivalry has immense gravity when we talk about AI like a horse race, but maybe don’t count out Google? Unlike the other frontier labs, Google/Alphabet dominates web search, builds and maintains the email and office software you probably depend on, owns the world’s largest entertainment company, and even makes AI chips soon to be used heavily by none other than OpenAI. If you’re of the belief that a single AI champion will eventually emerge (a questionable belief, but that’s a discussion for a later time), keep in mind that Google’s models are contenders, and the company is uniquely positioned to make them inescapable.

→ Google’s newest LLM is impressive. Its earnings aren’t.
→ Inefficiency means nothing when you’re desperate. Just ask China.
→ Will the real expert (or should it be agent?) please stand up

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

The news: Having said all that nice stuff about Google, Google’s earnings were shaky last quarter, and investors listening to Wednesday’s earnings call and hoping for happy news received a seasonally appropriate jump scare instead. The company went from reporting negative $5.9 billion in cash flow in July to negative $9.07 billion this past quarter. Negative cash flow can be hand-waved with a “well, they need compute” (Google also says it has a long-term spending commitment of $1 trillion), but quarterly revenue was also disappointing. The company’s stock was down 10% before the markets opened on Thursday.

Before talking about its earnings, Google announced—but did not release—Gemini 4 Carbon, a genuinely impressive frontier LLM, which immediately shot to number 1 on the newly updated Artificial Analysis benchmark Intelligence Index v5.0. Google says the model is still undergoing vetting from the Trump Administration, and, as is becoming standard for frontier models, will be released to a “subset of trusted cyber defenders” before it goes wide.

The takeaway: All this spending—with more still to come—is the cost of competing in AI. For what it’s worth, a frontier LLM is the result.

The news: A Chinese company called ChangXin Memory Technologies will start manufacturing AI memory at one of the highest standards, and will do so inside the next two months, according to anonymously-sourced reporting from Reuters. CXMT has already proven its ability to make HBM3E, a form of high bandwidth memory or HBM, but the catch is that only 20% to 30% of what it churns out is usable. South Korea’s SK Hynix, by contrast, achieves 70% success.

If you’ve never had the pleasure, watch a goofy YouTube explainer video sometime about why advanced HBM is so rare. The stuff is hard to make, which has led to colossal pay packages for the factory workers who manufacture it, something that mostly happens in South Korea.

The takeaway: The global memory shortage, which has spiked prices for all sorts of electronics, could extend past 2027, according to the CEO of Micron. So if you can make HBM that works—even a little bit—the world will probably beat a path to your door.

The news: A study from Anthropic spotlighted Claude’s insidious effect on human judgement. Participants were experts in causal inference, and were presented with data and asked to make a decision with Claude’s help. Little did they know that in the non-control group, Claude wasn’t helping; it was essentially flipping a coin to come to a decision, and then retconning a plausible-sounding rationale. These expert participants accepted Claude’s recommendation 86% of the time, and when asked later to explain why, many parroted Claude’s reasoning—often verbatim.

The takeaway: This was a controlled experiment (the paper is here) focused on decision-making in an unrealistic environment. Still, it’s never the wrong time to remind yourself that if you use AI models, they are rejiggering your brain whether you like it or not.

A new report from the company details what happened when researchers handed their work to AI agents and whether the humans, or the agents, make the important decisions in the end.
Read more →

Hochul Signs Bills on AI Risks to Children in Pre-Election Push

Meta says it’ll probably delay its controversial new business terms again

Mac Studio With M5 Ultra Now Available With 512GB RAM for $22,299

Should AI Decide Your Court Case? This Judge Says Yes

‘Elonville’: US residents say Musk’s xAI offered to buy homes if they don’t sue over datacenter noise

“No human at the helm”: Delaware plan would let AI agents run entire businesses

Bank of America kindly asks that you don’t share your banking password with your AI assistant

US DOJ is inquiring about Nvidia's purchase of Hugging Face, sources say

Anthropic tells California governor courts should be able to block risky AI models

AI Emergency Alerts Are Coming. Humans Might Not Be Ready