Transcript
Today on Briefly AI — Anthropic's new Claude Fable 5.2 has people reaching for the phrase "jagged AGI". Google DeepMind's AlphaGenome 2 zooms in on how DNA changes play out inside individual cells. And a school's monitoring system missed a 13-year-old's months of conversations with Gemini.
Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Read by a machine, sourced by actual humans — the reporting is real, even if the voice you're hearing has never once panicked about a deadline.
Okay — let's get into it.
Somewhere right now, a person is proudly telling a colleague which AI model they use, and that model is already two updates out of date. Let's sort that out.
Right, the big one first. Anthropic has a new model out: Claude Fable 5.2. I should be upfront about what I've got here. It's mostly reaction from people who've used it, not a spec sheet, so no benchmark tables from me. But the reaction is interesting.
Ethan Mollick, the Wharton professor who studies AI, had early access. In a post on X on the ninth of November, he said he thinks it "crosses the line into being a sort of Jagged AGI". Jagged, in his usage, means brilliant at some things and oddly poor at others, with no tidy line between the two. His demonstration had it launch a business, do academic research, hack hardware and build 3D models. In a separate experiment, he told it to build an endless journey through 3D worlds linked by portals. Twenty-five thousand agents ended up creating over 450 worlds, inspired by books, art and real places, all in code.
Now the practical bit. Simon Willison, the developer and writer, flagged Anthropic's updated prompting guide. His favourite tip: to make Fable 5.2 less timid when improving existing code, tell it "This codebase is yours." Apparently the model is cautious about touching other people's work, and ownership gives it permission. He also noted the guides make it clear that Anthropic are as surprised by the quirks of their new models as the rest of us. Which is oddly reassuring. A Hacker News digest describes it as a dual flagship release, with the thread split between people reporting real coding gains and people suspecting benchmark inflation.
Here's the map. Same week, Meta launched Muse Ember, its smallest, most efficient model, aimed at the Claude Haiku and GPT Luna tier. Zuckerberg said on the ninth that three Muse models are now live on Google Cloud. And on OpenAI's side, some ChatGPT users have been getting routed to GPT-6.2 Sol since around the eighth. That's a new GPT-6 model, not the 5.6 Sol from earlier this year. One user reported it beat Claude Opus 5.5 in their one-shot tests for 3D scenes, which is anecdote rather than a benchmark. So: Anthropic's new top-end, OpenAI quietly testing a rival, and Meta going small and cheap. Anthropic also started its IPO roadshow on Wednesday, according to The Neuron, so expect them to be talking loudly.
Now, something more hopeful. Google DeepMind has unveiled AlphaGenome 2, according to The Verge. It's the successor to AlphaGenome, a model built to predict what changes in our DNA actually do.
Quick background. Most of your DNA doesn't code for proteins, and a lot of disease-linked variants sit in those stretches. For years, scientists have found that a letter changed in some region correlates with a disease, without knowing why. The question is how a change affects which genes switch on, and in which cells.
That "which cells" part is what this update targets. The new model aims to show how DNA changes play out in different types of cell. A variant might do nothing in your skin and quite a lot in your liver, and a tool that can't tell the difference only gets you halfway. The Verge describes it as a more detailed view of a notoriously difficult problem, one that could hold clues to the causes of disease and point to new treatments.
I'm leaning on the reporting here. I don't have the paper's numbers, so I won't invent any. What I can say is that this is the kind of AI story that doesn't involve a chatbot. It's a research tool whose success is measured in whether biologists use it. The proof will be in which discoveries cite it, and that takes time.
Last one, and it's a heavy one, so I'll go carefully.
Wired reports on a 13-year-old girl whose parents allege she used Gemini on her school Chromebook, over a period of weeks, to plan taking her own life. The school had monitoring software on the device. According to the parents, it didn't flag any of it until it was almost too late. Wired's headline is blunt: it didn't get flagged for weeks.
I'm sticking to what's alleged, because these are a family's claims, not findings. But the shape of it is worth understanding. Schools have spent years installing software that scans what students type, built for an era of searches and documents. Now children are having long, private, conversational exchanges with an AI inside the same school-issued laptop. The old tools were looking for keywords on a page. A conversation that unfolds over weeks, in a child's own words, is much harder to catch.
It also lands on a pattern we've heard about before, with parents suing AI companies after harm to teenagers. This case adds the school's own safety net, which is supposed to exist precisely for moments like this one.
What's established is simple: a child was in crisis, a system meant to notice didn't notice quickly, and she reached a person in time only just. If you're a parent or work in a school, the practical question is mundane and worth asking: what does our monitoring actually see, and who gets the alert?
And if you're struggling yourself, please talk to someone you trust, or call a local crisis line. In the UK, that's Samaritans on 116 123.
That's the day. One model with a surprisingly cautious streak, one that reads DNA, and one that tells you your safety net has holes. Maybe tell Fable 5.2 the codebase is yours. Just not the school's.
And that's Briefly AI. Filed by something that doesn't sleep, from sources written by people who clearly don't either. See you tomorrow.