Please, no more papers with cartoon 'team members' to anthropomorphise slightly different prompts to the same LLM.

Easy version of 9-9-6: Do 12 hours work (or ‘work’), 6 days per week. Hard version: Your young children only go to bed at 9pm, then they wake up 9 times between them during the night, then are up and ready to go at 6am.

Just over an hour a way! 👇

To paraphrase Max Roser: AI is awful. AI is much better. AI can be much better.

Looking forward to chatting with Eugenia Cheng about maths, proof and our notions of what 'equal' means live at 1pm ET/6pm UK time today: open.substack.com/live-stream/...

I am (self-evidently) very interested in language. And also in technology. From that perspective, the rise of LLMs has created many new things for me to think about. One is: why is it that certain very small clusters of words are *clearly written by an LLM*? What is the quality of that writing?

Really enjoyed this piece from @adamjkucharski.bsky.social on the culture war between broad social improvements and targeted scientific understanding as ways to improve health. It’s frustrating because obviously both are important—but the former is often portrayed as somehow more worthy.

Turns out the rise of germ theory was more complicated than the one liners we usually hear about Snow and Koch ... of course. The travails are interesting and relevant.

“In an 1855 editorial, Wakley had dismissed Snow’s theory: ‘In riding his hobby very hard, he has fallen down through a gully-hole and has never since been able to get out again’.”

Von Pettenkoffer comes out of this episode looking like an idiot. And yet he was the public health scientist who argued that poor ventilation contributed to tuberculosis transmission (he was right) and identified CO2 measurement as a useful surrogate for ventilation

When can a 10% improvement be worse than a 5% one? New post on vibe coding and 'volatility tax': kucharski.substack.com/p/the-hidden...

Was chatting about that 2004 Nature 100m extrapolation paper the other day (pmc.ncbi.nlm.nih.gov/articles/PMC...) and decided to see how current results tracked. Post 2004 points are above fitted lines (and line for women now shallower because of retrospective drug DSQ for winning 2000 time).

New post on the original culture war: kucharski.substack.com/p/the-origin...

Want a large language model that never hallucinates? Easy: just get it to always print “I don’t know”, regardless of the input. If you find this solution is a bit unsatisfying, you’re not alone. But it hits on a crucial trade-off with learning machines and decision-making more generally. New post:

Recently came across the really useful term 'Moat of Low Status' from Cate Hall. It describes why we find it so difficult to take steps required to fulfil our ambitions - if we're already good at one thing, it's hard to endure being bad at something else.

Obviously I am not the first person to coin this term but it’s a thing right? Just recently without being conscious of it I found I’d started emailing old friends. LONG email. Cos we’re old and busy, replies come weeks or months later, they’re a nice surprise, I have to put aside time to read them!!

This astonishing feat has been dismissed across the political spectrum Some are upset about things the administration did/did not do Warp speed still delivered and deserves recognition Others applaud things the administration did/did not do Warp speed still delivered and deserves recognition

Would need to dig more into methods, but very curious about the process that has generated this clustering in ranking uncertainty. Would understand if it looked 'boxy', i.e. high ranked places that couldn't be distinguished. But seems uncertain about 15-25 range, and confident of those in 30-35?

Operation Warp Speed was arguably one of the most successful policy initiatives in recent history, accelerating vaccine delivery that saved millions of lives. It’s remarkable to see some now urging the leaders who backed it to disown the achievement.

Thanks to Braver Angels & David Zweig for engaging in this debate on the primary effects of the first year of the pandemic (540,000 deaths), the secondary effects (one year of learning loss for kids, unemployment, isolation), and things that we did right and wrong 1/ www.youtube.com/watch?v=59Gg...

A lot of unevidenced claims about vaccines are getting picked up and amplified on X by journalists who should probably know better. So if your aim is to use social media to reduce impact of these claims, then it does feel like you should be tackling the flawed-claim-to-media-uptake step on X...

“The public don’t trust vaccines”, say the people who’ve spent years trying to undermine public trust in vaccines.

mRNA vaccines are one of if not the most studied human inventions in history at this point. It's madness to say that they are somehow not well-understood!

A bizarre dogma has emerged among the anti-mRNA crowd that the only data we have on vaccines is from the original trial papers, rather than the HUGE volume of studies that came out subsequently...

Results so stunningly clear they inspired this classic xkcd (xkcd.com/2400/):

Good decision. I've always found the evidence for the idea that we need chickenpox to circulate to boost immunity to shingles quite weak. (And in reality, what's happened in UK is that people who can afford the vaccine get it anyway.)

This new analysis of the spread of France's 1789 Great Fear in Nature is a very nice dataset (www.nature.com/articles/s41...), but a few things look off with the epidemiological modelling... 1/

“Very nice… a breath of fresh air” Great to see Proof is one of the staff recommended reads by The Intercept: theintercept.com/2025/08/27/b...

Unlike a public health emergency of international concern, "pandemic" isn't a particularly well-defined term, which can cause problems, e.g. several legal cases around COVID hinge on when the pandemic "started" vs was "declared"... which is basically unanswerable.

New pre-print looking at social contacts and mixing by age, ethnicity, and socioeconomic status in the UK. We found that contacts have decreased since before COVID, but are up from figures found in 2022. We also show what this means for infectious disease dynamics in a novel pathogen outbreak 👇

Pre-pandemic, people in the UK on average made 10-11 social contacts per day (according to a large study we ran with the BBC in 2017/18). In late 2022, when there were no COVID restrictions still in place, it was around 6-7 in the UK. Have a guess what the number was in late 2024/early 2025....

Fascinating post on the risk of using LLMs for anomaly detection - very important for anyone thinking they’d be great for infectious disease or bioevent surveillance. What IS the tolerance to false positives, in this context? How many times can a detection system cry wolf before it’s ignored?

When Max Planck edited Annalen der Physik (which published Einstein's big papers), he said he aimed "to shun much more the reproach of having suppressed strange opinions than that of having been too gentle in evaluating them" I.e. setting a low bar was better than rejecting important-but-odd ideas.

TIL the original paper describing CRISPR, by Francisco Mojica, was rejected by 4 journals and took 2 years to be published

Screenshot of cases from the U.S. CDC’s website showing the number of national measles cases indexed by rash onset date. Are they going down?

New post, on why a 1% hallucination risk is terrible if you’re using large language models to look for unusual events: kucharski.substack.com/p/small-hall...

Enjoyed my Sunday afternoon visit to X dot com

Has anyone done a version of this plot, but for 'AGI is going to take all the jobs next year'?

'An expansive exploration of certainty in science, from geometry and jurisprudence to randomized trials and neural networks.' Great to see a thoughtful review of Proof in Science: www.science.org/doi/10.1126/...

Most are familiar with p-hacking, but I suspect there's a new trend in AI: demo-hacking. In other words, I wonder how many failed attempts were involved in selecting the remarkably impressive demo examples that end up in product announcements...

Great to see the positive response to Proof over the past 5 months! It's striking that I've had more e-mails about this book than my previous ones - based on readers' thoughtful messages, it's clearly sparked some new curiosity and highlighted stories they weren't aware of...

Really good piece. Well worth your time. 👇👇

And here's my longer piece on what makes a 'winter virus' a winter virus: kucharski.substack.com/p/is-covid-n...

"Whereas a small change in transmissibility conditions is enough to stop flu in summer, Sars-CoV-2 can basically evolve its way over this barrier" – I spoke to The Guardian about why COVID doesn't seem that constrained by season: www.theguardian.com/world/2025/a...

“Confidence intervals aren’t just a textbook method to be rote-learned. They are part of a historical story, which spans everything from economic strife to philosophical arguments about what is ‘true’, what can be imagined & measured. We can learn a lot by viewing the method in this broader context”

AGI = ability to produce a well-designed animated R plot. We're a long way off...

“Dr. Bhattacharya might not realize that many human vaccines are based on live attenuated viruses, where we similarly cannot definitively quantify antigen amounts produced in recipients. The mRNA and live attenuated vaccine platforms have both proven to be safe and effective in clinical studies.”

Want to get deeper intuition for a mathematical or statistical concept? I find that one of the best ways is to dig into its history. New post: kucharski.substack.com/p/to-underst...

"Ultimately, science is about helping us make better decisions—not perfect ones. And that’s especially true for doctors working on the frontlines of uncertainty." @adamjkucharski.bsky.social, author of the book Proof, explains how doctors can tackle misinformation www.bmj.com/content/390/...

My favorite variation of this theme are studies that say that future experimental studies are needed when the cause is something that cannot be experimentally manipulated, at least not in a targeted manner according to our current state of knowledge.