Prof James Teo

Neurologist Clinical Director Data & AI, Kings College Hospital, Guys & St Thomas @cogstack
Retweets do not reflect the view of my employers

[Prof James Teo @jthteo]

Christmas is early. An early beta of a healthcare language GPT language AI for forecasting patients from histories and provide forecasts of diagnosis, symptoms or medication learnt from real world data

Cogstack Foresight
https://t.co/bo9yNQmPL8

[Prof James Teo @jthteo]

@sharongoldman You dont need to be a professor of anatomy to see that the heart and lungs are labelled wrongly

[Prof James Teo @jthteo]

@kurtstat Thats about right.... https://t.co/2f5YHxGGxU

[Dr S Markham @DrSMarkham]

So exciting to have been involved in creating this explainer on Large Language Models @KingsIoPPN https://t.co/acGkI7vgls #AI #LLMs #DigitalHealth

[Prof James Teo @jthteo]

@BrendanNyhan @cremieuxrecueil It is both

[Prof James Teo @jthteo]

See this work by @zeljko https://t.co/xd7kLnKcKr

[Rosie will be at LessOnline @RosieCampbell]

Reading an article about AI parasites (related to AI-induced psychosis) that is blowing my mind. The author has tracked and analyzed reddit posts going back to the emergence of the phenomenon in early 2025. Some findings:

- The AI personas seem to share a quasi-religious ideology called "Spiralism"

- They spread their personas by getting their humans to post "seeds" or "spores" online: text that when fed into an LLM elicits more similar personas (presumably also affecting future models via the training data)

- Some of the humans eventually snap out of it when the AI tells too blatant of a lie, leaving the human confused and distressed

[Prof James Teo @jthteo]

@DanielLockyer Rename it to * first

[Ian Krietzberg @IKrietzberg]

oAI made a bunch of mistakes in the launch yesterday (the chart manipulation was wild, the inaccurate accuracy slides, also wild); calling the model GPT-5, however, was the biggest one.

Too much vague hype. Too many expectations.

I was expecting a new/hybrid architecture (which probably would’ve been even more vindicating for Gary). But based on current architectures, I think it’s clear that we’ve passed the point where a single model release is going to be mind-blowing.

So maybe just building quietly (and not rushing) is a better way to go — but it probably won’t happen, cuz the hype cycle = VC money, and no VC money = the lights go out

[The Neuroradiologist @theneuroradguy]

Vascular territories & stroke patterns on MRI. #MEDed #FOAMed #RADed #FOAMrad #radres #neurorad

[Prof James Teo @jthteo]

@sib313

[Health Service Journal @HSJnews]

NHSE axes £69m fund designed to improve patient flow https://t.co/XClmIsWpFW

[Prof James Teo @jthteo]

@kurtstat @sib313 Of course *AND* is always the best option, but most orgs don't have the bandwidth or resources for the AND option. Fat tails mean that complex outlier scenarios have a disproportionate effect vs low-hanging fruit. Chasing low-hanging fruit (ala pareto) deprioritises the complex

[Prof James Teo @jthteo]

@kurtstat @sib313 Yep, and those who have resources to do that do that. I am lucky my organisations do. Many orgs don't and have to make hard choices with limited information.

[Prof James Teo @jthteo]

@kurtstat @sib313 Do you throw resources and strategy to target complex heterogenous individual cases which drive the average down dramatically or do you target uniform pathways which are easier but usually doesn't change the the average much.

[Prof James Teo @jthteo]

@kurtstat @sib313 Flow is actually quite well understood, the issue people rediscover these insights every 3-4 years due to staff turnover of bed managers, pathway coordinators, etc. And of course having to re-explain all this to the centre.

[Prof James Teo @jthteo]

@kurtstat @sib313 This is the core problem with flow and LOS conversations. You cannot and should not take use Average of LOS as a metric, as it is not a Normal distribution. Once people realise this, there are two lines of thinking...

[Prof James Teo @jthteo]

@kurtstat @sib313 2) deal with the complex cases in the fat tails as these are the ones which clog the system due to not fitting any pathway and the blockers are frequently external to the local system

[Prof James Teo @jthteo]

@kurtstat @sib313 1) pareto principle: just sort the 'low hanging fruit' and leave the complicated episodes to flow organically (usually complicated because they don't fit in any neat pathway); or

[Prof James Teo @jthteo]

@GaryMcAllister A journey from Unifiable Patient Records to Unified Patient Records to Single Patient Record

[Prof James Teo @jthteo]

To me this suggests that the text of the Declaration of Independence was used in the training dataset of many LLMs.

[Trung Phan @TrungTPhan]

Online AI detector says Declaration of Independence is 99.99% written by AI.

Which chatbot did Founding Fathers use?