I’ve seen great results getting an LLM to write thorough and detailed prompts for another LLM. It’s great to see this area explored in a much more thorough and systematic way.

hardmaru@hardmaru.bsky.social · 3mo ago

For the past few years, humans have been doing “prompt engineering” to coax the best performance out of different LLMs. In this work, we explored what happens if we train an AI to do that job instead. Link to our #ICLR2026 paper: arxiv.org/abs/2512.04388 Thread:

I LOVE when science is applied to "toddler questions" like "Why is the sky blue?" or "Why is ice slippery?" Because the answers can be so SUPRISING. Let's dig into recent revelations about how ice skates work & why tires skid across black ice. First, we'll dismiss 3 VERY GOOD hypotheses as wrong.

Jennifer Ouellette@jenlucpiquant.bsky.social · 6mo ago

What makes ice slippery? Scientists may have finally figured it out www.scientificamerican.com/podcast/epis...

I've been in SF for four days and on each day I've had multiple conversations where people (self included) express deep feelings, fear, wonder and worry about their skills being replaced or (this is less named explicitly but an assumed driver I think) no longer valued by others. Big threats to feel

The better LLMs get, the more data scientists are going to find out the truth.

Must read. I’ve spent much of the past year visiting Big Locals, and have never been more certain that its model would deliver significant results if backed at scale (as the government is doing with Pride in Place) It is remarkable what some places have done given just a little money and time.

Sam Freedman@samfr.bsky.social · 7mo ago

New post just out: "Power to the People" Today we have a guest post from @jamestplunkett.bsky.social on how Labour can counter the "Britain is Broken" narrative by investing in civic life and giving people more control over their community. (Free to read) open.substack.com/pub/samf/p/p...

Arthur Ganson’s “Machine with Concrete” The magic of gears: the input shaft is driven at 200 rpm, but the output shaft will take over two trillion years to make one turn. So the final gear is embedded in a block of solid concrete. Via @kottke.org makezine.com/article/craf...

Arthur Ganson's "Machine with Concrete" - Make:

Arthur Gansen's kinetic sculpture includes a gear train that will take well over two trillion years before the final gear makes one turn.

makezine.com

"Synthetic Panels" make no sense. The LLM is trained on data from a particular point in time. This is the same as querying the distribution of responses from an LLM with built-in assumptions about what the distribution can look like. Well done surveys are designed for the opposite.

Sabrina Norwood@sabrinanorwood.bsky.social · 8mo ago

Did you know that from tomorrow, Qualtrics is offering synthetic panels (AI-generated participants)? Follow me down a rabbit hole I'm calling "doing science is tough and I'm so busy, can't we just make up participants?"

Radiolaria are microscopic unicellular sea creatures that create skeletons from silica. Their skeletons are among the most beautiful and complex natural microstructures, resembling elaborate filigree. blog post: n-e-r-v-o-u-s.com/blog/?p=9861

It's nearly that time again! Join us on Tue 4 Nov for the 101st PyData London meetup. Using Hugging Face's smolagents framework to create a Fantasy Premier League AI Agent - @geirfreysson.bsky.social No More Boilerplate: Data-Driven Testing Made Easy - @feregri.no www.meetup.com/pydata-londo...

PyData London - 101st Meetup, Tue, Nov 4, 2025, 7:00 PM | Meetup

**Venue**: Riverbank House, 2 Swan Ln, London EC4R 3AD **Please note**: 1. 🚨🚨🚨A valid photo ID is required by building security. 🚨🚨🚨 2. This event follows the **[Nu

meetup.com

Seeing flowers from my new aster for the first time is a real autumn treat!

I recommend this post from @simonwillison.net: simonwillison.net/2025/Oct/7/v... He lists the engineering skills required to use AI-generation most effectively to build quality software, based on his experience using AI on his projects. In short, we can't replace software engineers with vibe coding

It’s nearly time for our 100th Meetup! 🥳 Looking forward to a full roster of lightning talks, reflections from organisers past and present (including many of the OG crew!) and of course well-deserved celebrations--maybe even with cake! 🎂

PyData Global@pydata.bsky.social · 10mo ago

Our PyData London Meetup chapter is about to celebrate its 100th meetup tomorrow! 💯 We’re beyond proud to see how far our community has come, and can’t wait to hear all about this milestone event! Congrats to our fabulous PyData London Meetup chapter! 🥳

*deep breath* the fact that these models produce plausible output and also hallucinate (e.g. produce completely novel but reasonable output) has always been why they're so exciting for computational social scientists

Prof Mike Yearworth@mikeyearworth.bsky.social · 11mo ago

"In a landmark study, OpenAI researchers reveal that large language models will always produce plausible but false outputs, even with perfect data, due to fundamental statistical and computational limits." www.computerworld.com/article/4059...

Incidentally, my personal favorite interpretation of advancements in computer science is that most of the progress we've made has been in defining the problems themselves. Many successes correlate to clearer definitions of what "good" means or abstracting the task away from weak definitions.

Great to hear Customs House will be redeveloped! Such a grand building in an amazing location by the Thames, but today you have to follow a rickety walkway and peer at it through railings. Big improvement for the Thames Path there. customhouse2024.co.uk www.ianvisits.co.uk/articles/his...

Python: The Documentary | An origin story What began as a side project in the 1990s now powers AI, data science and some of the world's biggest companies. But Python almost disappeared at one point. Featuring Guido van Rossum, Travis Oliphant, Barry Warsaw, and many more! youtu.be/GfH4QL4VqJ0?...

Python: The Documentary | An origin story

YouTube video by CultRepo

youtu.be

I was particularly pleased to hear Ruby Waller tell us about apprenticeships in data. I had no idea this was a possible route into DS until she told me at the conference this year. So glad she agreed to share her journey!

PyData London@pydatalondon.bsky.social · 11mo ago

Last but definitely not least we have Ruby Waller on Apprenticeships in Data: a practical pathway into the profession.

Had a great PyData London tonight! Was a real treat to hear from @gaelvaroquaux.bsky.social on @skrub-data.bsky.social and the real world data pains its solving. (Try it if you haven’t already; super easy to get going!)

PyData London@pydatalondon.bsky.social · 11mo ago

Our first talk tonight is from @gaelvaroquaux.bsky.social on @skrub-data.bsky.social. Real tables are too messy for sklearn - skrub preprocesses them for you.

PyData London 2025 keynote: Tony Mears -- Innovation is Dead Innovation gets stifled in the public sector by risk aversion and punishment for failure. But there is hope: Tony shares a way forward, drawing on his experience directing operations in the NHS and UK Space Agency. youtu.be/XMLbJYM46xo

Keynote- Tony Mears | PyData London 25

YouTube video by PyData

youtu.be

on.ft.com/4oVBaBe My day trip to medieval England Jousting is back in vogue after a 500 year absence. Fantastic pictures too!

My day trip to medieval England

At the ‘Wimbledon of jousting’, Simon Usborne finds historical re-enactment is becoming big business — and meets the knights and tourists partying like it’s 1499

on.ft.com

Fascinating interview from Odd Lots on what happens when a startup dies. Not often formal bankruptcy. Finding buyers for some company assets can be challenging—like human skulls! podcasts.apple.com/gb/podcast/o...

This Is What Happens When a Startup Dies

Podcast Episode · Odd Lots · 20/08/2025 · 47m

podcasts.apple.com

Excited about this one!

PyData London@pydatalondon.bsky.social · 12mo ago

Join us at the 99th PyData London Meetup — with a talk from a genuine PyData celebrity! 🤩 Tue 2 September 19:00 Skrub: machine-learning with dataframes — @gaelvaroquaux.bsky.social Model Context Protocol: is refactoring tool calls worth your time? — Victor Zommers meetu.ps/e/Pt20x/5p7c...

Today at #EuroScipy2025, @glemaitre58.bsky.social and I presented a tutorial on pitfalls of machine learning for imbalanced classification problems. We discussed what (not) to do when fitting a classifier and obtaining degenerate precision or recall values. probabl-ai.github.io/calibration-...

I'll admit, I was skeptical when they said Gemini was just like a bunch of PhDs. But I gotta admit they nailed it.

Fusion energy start-up claims to have cracked alchemy This is really cool: modifying the fusion process could produce gold from mercury almost as a side effect, worth almost as much as the electricity produced each year. on.ft.com/3IXXlWD

Fusion energy start-up claims to have cracked alchemy

San Francisco start-up says modified fusion process could be used to produce gold from mercury

on.ft.com

Whisper bug report: 'Complete silence is always hallucinated as "ترجمة نانسي قنقر" in Arabic which translates as "Translation by Nancy Qunqar"' social.coop/@brainwane/1... It's fun, but as the thread points out not a new problem, e.g. fanfare -> "subscribe"—and would be fixed in English already.

Visited a few Sculpture in the City exhibits on my afternoon cycle today. Will see if I can find the rest over the next few weeks! www.sculptureinthecity.org.uk

Lovely evening for a cycle yesterday.

Great panel on driving meaningful change to developer experience. "We need better measures than satisfaction that connect directly to things *under our responsibility*, like are people able to access their best problem solving selves, is wellbeing being protected, and quality of problem solving"

Cat Hicks@grimalkina.bsky.social · last yr.

Revealed myself as a "Satisfaction" skeptic in this panel 😂 it's just a very messy and noisy human measure and we have other ones!