I just happened to stumble across this blog from Cory Doctorow from last September that resonated with your final point there: pluralistic.net/2025/09/27/e... I'm sure it is true that there is a hell of a lot of room for optimisation of models once the bubble bursts and more is open sourced ...

Tech firms have over-leveraged themselves to the point where AI is "too big to fail" - the only way they can repay their investments is if we all become dependent on LLMs for everything. But it simply isn't suited to most tasks. Don't buy the hype, but DO use AI where it is the right tool! (9/9)

This is a great example of the nature of a Language Model - wonderful at dealing with patterns and the logic of a document - linguistic structure! But it simply doesn't deal with *knowledge* or *facts*, and asking it to do that is intrinsically going against the grain of how it is structured (8/9)

That really highlights what's going on. Internal reasoning within a text, and basic cross referencing (like King's reigns) play to the strengths of an LLM - a Language Model. But asking it show sources highlights the weaknesses - it has, at best, done a traditional search retrospectively! (7/9)

Gemini has even cited our very own @layersoflondon.bsky.social - but that is just as much of a tell that this is all fake as the random Ealing church website! Layers is a great for medieval parishes, but I'm pretty sure that Gemini's visual reasoning can't interpret those complex map overlays (6/9)

Asking Gemini to annotate the transcription also got great results, and certainly showed reasoning. The identification of 'church of St Benedict' as St Benet Sherehog is particularly impressive - so I asked how? The logic is correct. Where it all goes wrong is in the citations ... (5/9)

Inspired by Humphries' blog, we used Gemini Pro 3.5 Preview in Google AI Studio, and began uploading and prompting for transcription into JSON and Markdown formats, with great results! (better than volunteers...) Then we started to try the annotation and reasoning that the blog highlighted (4/9)

We hold the 'Great Wall of Cheapside', a vast set of handwritten index cards created by Derek Keene and his team during the 1980s and 90s, summarising medieval and early modern City of London property deeds. Narratives of each location were published but the cards contain a wealth of names etc (3/9)

There was a flurry of interest a few months ago in response to Mark Humphries' article on using Google Gemini 3.5 Pro Preview for transcribing a C18th commercial ledger generativehistory.substack.com/p/the-sugar-... So at @chppc.bsky.social we decided to have an experiment (2/9)

Also @neilstewart.bsky.social and Rebecca from the IHR's UoL cup winning quiz team!

Oh I'm sure London Haberdashers supplied Colchester chapmen! But Colchester definitely had its own regular shipping with Cologne. Stuart Jenks (whose data we're using) has been finding interesting stuff in Colchester court books re: German merchants suing rural weavers for debts - so direct trade