There was a great paper written a couple of years ago about the worsening quality of the advice received by dictators as their rule gets more precarious, and the dynamics that lead to it: roblog.co.uk/2023/04/dict...

I finished it a couple of weeks ago. I had to set aside some time on a quiet Saturday morning to read the last couple of chapters, both because I didn't want it to end and because I had a sense of how I might feel when it did. "Whimsical and devastating" is exactly right – thank you for it.

An absolute pleasure, and Merry Christmas right back at ya

(Here's a gift link just in case it's paywalled: giftarticle.ft.com/giftarticle/... – the link is in the footer links at the end of the article)

Looks like it's linked from here: www.ft.com/content/c26f...

(And shark pyjamas! It's an amazing AI/Photoshop job though, really can't see any dead-giveaway image artefacts about it)

Yeah, this is it IMO – that and the general situation. Why would they be together on Christmas morning at all, let alone in a house with a Jaws-themed Christmas tree?!

Yeah I think both of these points of view are very valid. I think even a light bit of hypertext/generative interactivity sprinkled on the otherwise traditional index would be interesting though (e.g. even just the ability to group index entries semantically as well as ordering them alphabetically)

Like, we have indexes in paper books precisely because we can't search them, computers were good at searching text before LLMs and are waaaaay better at searching text after them. We should be imagining how to more effectively search books, not how to more efficiently create indexes

The need to keep the output as "TOPIC: page #, page #, page #" is an annoying limiting/complicating factor, which shows their initial premise to be a bit silly. Why get an LLM to replicate a dead-tree index format, when you can get it to give you a hyperlinked semantic search over the entire book?!

Look forward to seeing how you get on with this. My mind went to a sort of semantic index approach: treat each page as a separate document, generate embeddings, do some sort of kNN to find thematically similar pages, feed those into an LLM to summarise the (topics | people | other works) referenced

Thanks! Appreciate that. AI is by a million miles the hardest-to-navigate technical-ethical dilemma I've experienced in my life, especially since the position I think is probably morally and practically correct is so different from that of my "tribe" (soft-left creatives and craftspeople)

I tried to write a bit about this at the start of the year; IMO the correct framing is to oppose "machinery hurtful to commonality", as the Luddites put it, and the answer is to try to create machinery that is helpful rather than hurtful roblog.co.uk/2025/01/the-... & roblog.co.uk/2025/08/ai-s...

From a retailer perspective this is what's so frustrating. You have people say "DPD are awful, I'll never buy from you if you use them" and others say "I'd only ever choose DPD if I could", repeated for each and every courier. But it's all down to how good their local driver is

Yes, you're right – the Americas were geographically isolated for thousands of years, and Australia too, which means the identical ancestors point for humans in ~1490 will have been ~50,000 years earlier. Migration has collapsed the IAP for currently living humans to a much more recent date

Here's more to dig into: en.wikipedia.org/wiki/Identic...

(Mitochondrial Eve is not technically the most recent common ancestor, just the most recent common ancestor through strict matrilineal descent, but I believe we can't tell who those other candidates might be)