Today's Hallucination HQOpenAI Hires Its Own Sceptic, Which Is Either Brilliant or a Hostage Situation
Paul Christiano, one of the most respected voices in AI safety and a man who has publicly worried that advanced AI might kill us all, is joining the OpenAI Foundation's board. Christiano leads the Alignment Research Center and has, to put it mildly, complicated feelings about the organisation he's now helping govern. Poachers hiring the gamekeeper. The fox chairing the henhouse committee. Truly, we live in interesting times.
Source: TechCrunch
OpenAI Quietly Solves a 100-Year-Old Maths Problem, Academia Nervously Clears Its Throat
OpenAI claims to have cracked the Navier-Stokes problem — one of seven Millennium Prize challenges so notoriously hard that solving any one carries a $1 million reward and the permanent, slightly smug satisfaction of being correct. The result is being treated as mathematically genuine, which is the good news. The bad news, for anyone with a PhD and a mortgage, is that an AI just did in months what humans couldn't manage in a century. Peer review, one assumes, is ongoing and uncomfortable.
Source: The Verge
Companies Are Spending Less on AI Per Head. The Hyperscalers Are Pretending Not to Notice
AI spending per employee at major firms dipped in August, and before anyone reaches for an explanation involving beach holidays, the underlying trend is real: token costs are falling, cheaper models are multiplying, and enterprises are getting more for less. This is, of course, excellent news for everyone except the cloud providers who built their quarterly projections on the assumption that the spending frenzy would continue indefinitely. It did not. Gravity, as ever, remains employed.
Source: TechCrunch
Suno and the Record Industry Make Music Together, Which Sounded Unlikely Six Months Ago
Suno's new v6 music model is the first trained with actual record label cooperation — licensed data, not contested data — marking a notable shift from the industry's previous strategy of simply suing everyone. The model was built from scratch with a clean dataset, which Suno is keen to emphasise, presumably because the alternative generated rather a lot of legal correspondence. Whether the music is any good remains, as always, a matter of taste. Whether the lawyers are satisfied is a matter of contract.
Source: The Verge
Healthcare AI Can Read Your Records. Now Comes the Harder Part: Fitting Into the System
The technical capabilities of healthcare AI are, by most accounts, genuinely impressive — models that parse long clinical histories, interpret complex terminology, and cross-reference documentation with speed no human can match. The problem, as MIT Technology Review outlines, is integration: hospitals run on legacy software, fragmented databases, and workflows built before anyone had heard of a large language model. Brilliant at diagnosis, less brilliant at navigating a forty-year-old patient records system running on something your IT department would rather not discuss.
Source: MIT Technology Review
As always, the models are getting smarter. The paperwork is not.
|