I found out the other day, courtesy of @mattyglesias.bsky.social, that after WW2 Christie had her literary agent remove some of the anti-semitism from the older books:
forward.com/culture/4580...
Open call for papers, The Rate of Return to R&D Investments. Conference to be held in Washington, DC on October 2, 2026. Submit papers by 11:59pm EDT on June 17, 2026. More information: www.nber.org/calls-papers...
A small side street near Victoria contains a mosaic that people will tell you is an old advert for the Victor Talking Machine Company. They are wrong.
www.ianvisits.co.uk/articles/the...
It’s nice round there! If you get the chance I cannot recommend enough going to the Louisiana, great art and the grounds are stunning. louisiana.dk/en/plan-your...
open.substack.com/pub/paulkrug...
Useful post with useful model. Being great at something that's collapsing in price doesn't necessarily advantage you
The answer there, iirc, is about funding security for the sector wrt to management of department budgets
Super interesting
Locations of new homes built in Switzerland in 2018
Country-wide effects of new housing supply: Evidence from moving chains, by Lukas Hauck and Frederic Kluser
Another new paper on housebuilding and vacancy chains, this time with data on every Swiss resident & housing unit! An interesting context given Switzerland's high immigration, very large rented sector and strong tenancy rent controls... frederickluser.github.io/files/Moving...
A drawing about vector databases, describing first vectors in AI, then how we map them in a high dimensional coordinate system to derive semantic similarities, and then how we can perform fast searches using ANN algorithms, and HNSW to traverse a heirachal graph structure to find the nearest neighbor
💥 I made a new drawing in my AI series, this time about Vector Databases, ANN, and HNSW. I hope it's useful!
It might be good to look at the Transformers one in the series before this one, for additional background.
An economic model of "Artificial Jagged Intelligence"—where AI systems perform well on some tasks but fail on similar ones—showing how users' inability to predict local reliability creates adoption challenges that persist even as models improve, from @joshgans.bsky.social www.nber.org/papers/w34712
3) Accurate classification: CAT performs comparably to human analysts when classifying themes to individual free-text responses, enabling robust estimates of theme prevalence rates to be produced.
2) Accurate theme extraction: CAT identified between 75% (blind) and 90% (non-blind) of the themes identified by human analysts. With our human-in-the-loop review process, this increases to 100%, while allowing human experts to apply their policy judgement to ensure no key themes are missed.
*Key headlines*
1) Potential benefits: If scaled across all DfT consultations, CAT could save ~£1.5-4m/yr, while significantly reducing resource pressure, with time saved in the policy development cycle helping us meet the CO recommendation to publish consultation responses within 12 weeks
CAT was developed and evaluated by the DfT AI team, in collaboration with a team at the Alan Turing Institute, using LLMs with human oversight to analyse free-text responses from public consultations. This is the first evaluation we've published for one of our AI tools in DfT, hopefully more to come
I wanted to start the new year by celebrating the work of my team in DfT. We just published the evaluation of our Consultation Analysis Tool (CAT): www.gov.uk/government/p...
CAT is one of our main internal AI projects at DfT, and has shown the strongest internal demand and clearest benefits.
This is a nice trip down memory lane
In the early 1990s, before the vaccines were introduced, the virus caused around 4 million cases of disease, 12,000 hospitalisations and 100 deaths each year in the United States. Children were most affected, making up >90% of cases, >60% of hospitalisations and around 40% of deaths. I’ve used past tense, because here’s what happened after that.
My blogpost on the success of universal chickenpox vaccination in the US, and why the UK should have made the move sooner:
www.scientificdiscovery.dev/p/13-the-suc...
Chickenpox vaccines are really effective!
Woohoo! Chickenpox vaccines are on the menu in the UK - finally!!
www.theguardian.com/commentisfre...
If you don’t mind an old season style, then places like brandalley could be a good option for saving £ once you’ve picked a brand
I put together a detailed collection of useful patterns I've collected after vibe-coding 150 different single-file HTML tools over the past couple of years simonwillison.net/2025/Dec/10/...
We tested one of the most common prompting techniques: giving the AI a persona to make it more accurate
We found that telling the AI "you are a great physicist" doesn't make it significantly more accurate at answering physics questions, nor does "you are a lawyer" make it worse.
Another example of the bitter lesson?
I'm afraid we've been at it again
‘Super sewer’ chief bows out to the sweet smell of success on.ft.com/4gdIXWU
😭
Tempting fate? No, the second scoop is an earl grey sorbet!
See also: the phenomena of second scoop regret with gelato h/t @scientificdiscovery.dev
New method detects inflation turning points using disaggregated price data, revealing early signals in Argentina's 2024 disinflation and US tariffs in 2025, from Alberto Cavallo and Gaston Garcia Zavaleta https://www.nber.org/papers/w34102
I think this way of doing is exactly the right one for figuring out AI tool applications and trying to improve internal processes (albeit it’s not my kind of language). The discovery/alpha/beta cycle or framework can become a straight jacket. Or trap. Or whatever the right metaphor is.
