Episode — 2026-04-14

Stanford just dropped AI's annual report card. The marks are... alarming. Let's get into it.

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

Right, so every year Stanford's Institute for Human-Centered AI publishes what they call the AI Index — basically the most comprehensive data-driven look at where AI actually is, stripped of the hype. The 2026 edition landed yesterday, and a few numbers jumped out.

First: AI is not slowing down. Despite a year of people confidently predicting a plateau, the top models just kept getting better. On one key coding benchmark — SWE-bench Verified, which tests whether AI can actually fix real software bugs — performance went from 60% to near 100% in a single year. One year. That's not incremental progress, that's a different category of thing.

Second: adoption is accelerating faster than any previous technology in history. Faster than the personal computer. Faster than the internet. Eighty-eight percent of organisations now use AI in some capacity. Four in five university students use generative AI. Those numbers would have sounded like science fiction three years ago.

But here's the number that stopped me. AI data centres now draw 29.6 gigawatts of power globally — enough to run the entire state of New York at peak demand. And running OpenAI's GPT-4o alone may use more water annually than the drinking needs of 12 million people. That's not a footnote, that's a civilisational cost that nobody's really reckoning with yet.

And there's one more thing in the report worth flagging. Yesterday I mentioned Anthropic's gap between its public safety positioning and what it actually builds. Well, Stanford noticed something similar at industry level: OpenAI, Anthropic, and Google have quietly stopped disclosing their training data, parameter counts, and model architectures. The race got so competitive that transparency became a casualty. Which makes it genuinely harder for independent researchers to study safety. Something to sit with.

Now, this next one is a bit more fun. A video AI model appeared on the benchmarking platform Artificial Analysis around 7th April — no name, no company affiliation, just calling itself "HappyHorse-1.0." It climbed to the top of the rankings for both text-to-video and image-to-video generation almost immediately. Blind tests, real users picking it over the competition, synchronized audio, the works. Everyone was asking: who made this?

Turns out it was Alibaba. Specifically their Future Life Lab, a skunkworks team inside Taotian Group, which is part of the Alibaba ecosystem. They'd entered it into benchmarks anonymously, let it win on merit, then claimed it.

Now look — this is a clever strategy, and it worked. But it's also a signal. China's AI labs are no longer playing catch-up. They're submitting anonymous models to global leaderboards and topping them before anyone knows who built it. The Stanford report backs this up: US and China are now separated by razor-thin margins at the frontier. The idea that there's a comfortable Western lead is, charitably, outdated.

Worth keeping an eye on: HappyHorse is generating videos with synchronized audio, which is still something most Western models struggle with. Alibaba just turned that into a quiet flex.

And finally, something that's particularly relevant if you work anywhere near healthcare. Utah — the US state, not a startup — has become the first jurisdiction in the world to formally authorise an AI system to renew drug prescriptions autonomously. No doctor sign-off required, every time.

The specifics matter. A YC-backed company called Legion Health has been cleared to let their AI renew certain psychiatric prescriptions — things like antidepressants and antianxiety medications — for stable patients who meet criteria. It's restricted: patients can't have had a psychiatric hospitalisation in the last year, and only 15 low-risk medications are eligible. There's another pilot covering 192 drugs for chronic conditions like hypertension and diabetes.

Now, the intent here is real. There's a genuine shortage of psychiatrists in the US. Patients who are stable and just need their existing medication renewed can wait months for an appointment. If the AI is doing the routine, the human can focus on the complex. That's a reasonable argument.

But — and this is the "what to watch" — this is a pilot in a single state with a narrow set of medications, and how it performs over the next few months will either open the door to this becoming standard practice or slam it shut. Healthcare AI is often one high-profile failure away from a decade of regulatory backlash. The stakes are high for everyone building in this space.

That's your lot for today. The Stanford report alone is worth reading if you have time — the MIT Technology Review did a decent summary with the key charts. And whether it's the stealth horse, the transparent lack of transparency, or Dr. AI writing your prescriptions, the throughline is the same: AI is operating at a different scale now, and the systems built to oversee it are still finding their shoes.

See you tomorrow.