Weekly Roundup 2026-09-13
Hugging Face, Vintage Cameras and Uncomfortable Oxford.
LLMs
- Colin Fraser (a data scientist at Meta) has become my new favourite Bluesky follow, for threads like this one. Framing LLMs as having a novel capability without hyping it up to be AGI has been helpful to me in thinking about how to get real value out of LLMs in my field, and how to evaluate where there is potential.
They do a kind of approximation to deductive inference which is kind of off sometimes but is frequently very much on, and which is flexible and fluid enough to be widely applicable wherever generalized deductive reasoning is helpful, which is a lot of places. (link)
- Katie Paul at Reuters published a great investigative piece on Meta's attempt to transform into an "AI native" organization, a decision that seems to have been entirely based on wishful thinking. I posted my thoughts on what lessons I could take from this failure here, and on linkedin (sorry). There are some interesting threads to follow on whether LLMs are actually capable of supporting a fundamental shift in how organisations are managed. I'm not sure they are. I will be sharing my research and thinking as I unpack that over the next few weeks.
- The OpenAI Hugging Face hack coverage gained steam and broke through to the normals. Stephen Witt's Opinion Piece in the New York Times "This Is Really Bad" made it onto my wife's group chat with her sisters, with the comment "I hadn't heard about this and now I'm terrified. Who is going to do anything about this!?"
- I made a note to myself to write a rebuttal to Witt's article expressly for my sisters-in-law, but I think Cory Doctorow beat me to the punch with his "LLMs are real, AI is fake" post mere hours later. Alex Hern's TL;DR summary on bluesky is even more suitable reading for normal family members. He's one of the few journalists that both get the technical nuance of LLMs and seems to be able to communicate it to normals, at least in my experiments.
- I (shockingly) find myself agreeing with John Gruber (previous and previouslier disagreements) on his bang-on assessment of Jacob Coxon and Anthropic:
Coxon thinks he’s blowing the whistle that Anthropic is a threat to the world. What he’s actually done is confirm that Anthropic is a cult, and its employees can’t really be talked to rationally, because they’re cultists.
Vintage Cameras
- I was gifted a delightful Iagee Parvola last week and spent too much time figuring out how to build a digital back for it, before finding the perfect donor (Tomy KiiPix) to build an Instax Mini back. A full report will come when I've got it working. Some teaser photos.
Uncomfortable Oxford
- This weekend was Open Doors at Oxford, and we visited Rhodes House. We didn't get organised in time to book a spot on the Uncomfortable Oxford tour, but it's not hard to be uncomfortable in an ego-palace to a white supremacist. I find the inability of the Rhodes Foundation to even contemplate removing his name from the little bit of good that they do in the world absolutely astonishing.