I might just have to make a weekly news roundup part of the regular rotation here, as once again the noteworthy stories crowding my inbox are so numerous this morning I feel compelled to comment on each of them in turn. If you hate roundups, skip this post – but here are the top stories I find worthy of additional comment today:
- OpenAI’s Math Problem = AI’s Data, Explainability, and Sourcing Problems. Yesterday OpenAI claimed to have solved a long standing maths problem having to do with Navier-Stokes equations for fluid motion, which for those of us who stopped doing maths in high school, is a Really Big Deal. But as Wired reports, it turns out OpenAI’s work may well have been juiced by prompt data input by two researchers working on the same problem. Adding to the intrigue: One of those researchers also works at Anthropic, and to advance his work, he and his colleague had used OpenAI’s Codex as well as Anthropic’s Claude platform. I’m sure we’ll get the HBO movie in due time, but for now, I think the larger questions are more interesting, starting with this one: Is OpenAI using “private” prompt data to train its internal systems? The answer, given by OpenAI, is “we can’t be sure.” “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models,” OpenAI wrote, which sounds an awful lot like a yes to me. Bonus additional questions raised by this dust up: Will there every be a way to know for sure? Probably not, because there’s not an industry defined approach to auditing AI systems at scale – and likely, there never will be. Like all of Big Tech, AI platforms really, really don’t want to be regulated or held to account for how their products are used, including internally.
- Serious Tremors In Bond Land. If I ever tell you I understand how bond markets work, put me in stocks and toss tomatoes at my head. But I know this: When bonds get the sniffles, the world gets the flu. And right now all signals from the global bond market indicate a terrible head cold. The tipping point: The US has begun to intervene in bond markets in a way that puts the Trump administration at odds with the most powerful force in our economy not named NVIDIA: Wall Street traders. From the FT: “Slowly, and sometimes subtly, but surely, the voices calling out the administration’s inconsistencies and irrationalities are getting louder.” That might sound like pulling punches, but in the pages of the FT, it’s an all out fist fight. I’m watching this one closely.
- Meta Puts a Genie on a Leash. If you read my rants regularly, you know I am deeply skeptical of using AI agents from companies whose core business model is selling you something other than AI agents. Of course, the first company to introduce personal AI agents at scale is….Meta. Dubbed “Muse,” Meta’s new AI agent can work on your behalf across the web, and initial reviews are generally positive. Meta’s video (screen shot above) introducing Muse shows the agent buying school supplies based on email lists, signing a user up for a 10K run, and rebooking cancelled flights. Meta has given Muse a cuddly mascot, which Google Gemini, when prompted, calls a “whimsical, cream-colored plush sprite.” OK, but I’ll call it what it is: terrifying*. Nearly 15 years ago I warned you all to not build your digital lives inside platforms like Facebook. I’ll repeat the warning, updated for the AI era: Digital agents like Muse may seem like magic genies, ready to do your bidding. But they’re on a leash held by Meta, and they work for Meta’s business model, not for you.
*It also looks like a thumb with handsless arms, but whatever.
—