Wat

Such a good piece today from @jburnmurdoch.ft.com which shows that the declining graduate premium is very much a UK problem rather than a general (or inevevitable) consequence of more people going to uni www.ft.com/content/649d...

somewhat terrifying chatjimmy.ai

Hypothesis: AI writing is bad primarily because in general what makes writing good is the specificity of what it communicates and its information density, so the prompt that produced good AI writing would be as long as the output.

This is a rare case where if you give a screenshot of this post to ChatGPT/Claude and ask it to explain the joke, if any, it doesn't get it

sick claude is rather tragic

Oh this is what it feels like to be a reactionary

Here is a very rough version of the For You Stats dashboard: linklonk.com/foryou It shows you 3 things: - how much you've been using For You - how many views and likes your posts got in For You - how your likes/reposts helped surface posts to other people in For You Feedback welcome

seems bad

Ok, clicking through to the blog post that this blog post is referencing, the author makes this point directly: theshamblog.com/an-ai-agent-...

Genuinely very impressed by the SVG of a pelican riding a bicycle I just got out of Google's new Gemini 3 Deep Think model simonwillison.net/2026/Feb/12/...

Gemini Deep Think Oneshot! Note the Oxpecker bird on the hippo! gist.github.com/theosanderso...

See paper in Nature, looks interesting, read it, the latest model they evaluated was o1 preview. This needs to stop. I understand publishing timelines are long but at least change your headline or the abstract.

ARTIC 2 measles resources are live! As part of the Wellcome Trust funded ARTIC 2 project we have developed a collection of measles virus resources that are now openly available at artic.network/viruses/mev

what a time to be alive

My latest, on a town's escalating fight to fix a broken road, in which become accidentally radicalised by Nick Offerman. m.youtube.com/watch?v=RKiV...

So many low hanging fruit now due to agentic development, but hard to decide whether they are worth picking when anyone can do so

New Claude bug spotted (reproducible), any theories?

ahaha, I'm regenerating the map labels to reflect the fact that they are landmarks on a landscape rather than classifications of individuals, but I might need to do another pass (Sonnet may be less diplomatic than Haiku)

A long time ago in a galaxy far away, there was a SARS-CoV-2 pandemic. Our paper, led by @martibartfast.bsky.social a) correcting errors in 4.5 million genomes & their phylogeny b) improving representation of the Global South in public data www.nature.com/articles/s41... (thread 1/n)

I made a map of 3.4 million Bluesky users - see if you can find yourself! bluesky-map.theo.io I've seen some similar projects, but IMO this seems to better capture some of the fine-grained detail

TIL of this page to tell you which posts you liked to cause For You (bsky.app/profile/spac...) to recommend a particular post: linklonk.com/bluesky

I've peaked (thanks to a rainy day and a nicher course)

opus 4.6 is a significant step up at my openscad benchmarks

This is the actual author of this video btw

--dangerously-skip-understanding

OOOOH Researchfish Is going extinct 🐡🐟🐠 help.researchfish.com/en_US/resear...

Finally got around to digging out my old desktop getting it set up with tailscale and dangerclaude. Has definitely been worth the faff to be able to leave claude to do its thing in a relatively safe environment (just w/ screen/tmux mostly)

wild that the future has arrived and it uses ANSI text graphics

Made a quick thing to generate these test prompts: theosanderson.github.io/wordblanker/ dang impressive ime

Really excited our paper about how H5N1 rapidly adapted to cattle (and how these adaptations also increased its ability to infect cells from the human respiratory tract) is now published! www.nature.com/articles/s41...

need an aurora app that suppresses notifications when it's cloudy to avoid fomo :/

This photo was taken in 1911 using glass plate technology by Herbert Ponting who was part of Scott's Antarctic expedition, The composition and detail are exquisite with the band of white snow/ice creating a perfect frame around the two people and the ship in the distance Iconic imo

my new job: life coach for llms

New post, on how AI is coming for open source software: kucharski.substack.com/p/will-your-...

Want to get the data out of a PDF figure? As in, the actual data – not a rough trace-along-the-lines version? I made an app you might like: adamkucharski.github.io/pdf2plot/ It all started a few years ago... 🧵

gemini deep research is the most pretentious model out there

A policy @ncsc.gov.uk specifically advise against www.ncsc.gov.uk/blog-post/pr...

Anyone have a good GHA yaml for making claude's code review more like codex? (Less fluff, less compliments!) The below doesn't seem to have done the trick

Following accounts like this is probably quite an effective way to convince others to move

What proportion of new production code for commercial software in 2028 will have been read by a human? 1️⃣ 0-10% 2️⃣ 10-50% 3️⃣ 50%-90% 4️⃣ >90% 📊 Show results

i do feel a bit sad that i can't tell if this is real or not

"She and Chen sent GISAID three emails in the weeks after, but never received any response. (GISAID says the emails came from Chen, whereas its agreement was with Stadler.)" www.science.org/content/arti...

I also got this, but I thought it was kind of interesting, and did their survey to see how well it did (spoiler: it was bad). As I mention in a reply to the OP, I don't see an obvious problem with the research (other than that their AI is bad at doing the thing they want it to do).

This seems positive?

Global cough radar sleepcycle.com/coughradar?l...

On the first day of Christmas (in about 2014), @duckswabber.bsky.social gave to me - a partridge in a pear tree. Happy Christmas everyone!

Bike-couriered by the editor himself. An illustrious publication!

'Beer stein icons indicate roughly five-day dosing windows' www.sciencenews.org/article/vacc...

This smells distinctly like collider bias and/or selection bias and/or regression to the mean... You simply can't select teen prodigies, and world class athletes rom databases, and go run regressions without serious consideration of the selection process!