I’m running the Hackney half for charity again, this time WWF. www.justgiving.com/fundraising/...

Fiction: “I have instructed my AI to remove all unnecessary politesse and verbiage from all incoming messages. There will thus be no manipulative value to you in not getting to the point. Just say what you want without the schmooze.”

Delighted to join today’s UCL Policy Lab Roundtable on AI, the future of Work and the Role of the University. There was a broad panel of participants with distinctive & fascinating perspectives. HE mist be relevant to student needs, but what is ‘relevant’ increasingly is difficult to predict

We have funded 10 ambitious Challenge-led Awards through the @UCL Grand Challenge Data Empowered Societies, University College London www.ucl.ac.uk/grand-challe...

This is a classic format. The last time I saw this, the subject was signal-to-noise ratio. It is interesting to consider that math is the signal, and 'magic' is the noise.

At the SC25 trusted research environments BOF. Sadaf Alam presenting our work on Federated Research Infrastructures through Data Governance Extension (FRIDGE). With @bristoluni.bsky.social @ucl-arc.bsky.social Turing Institute and Cambridge

Come to UK Universities booth at #SC25 to find out more about the UK's #DRI programme and pick up some merchandise!

Extremely impressed by what U Illinois have achieved with in house LLM services using their local HPC investments and open source models and tools. This is what I’ve been advocating for UCL and they’ve proved it can be done.

Please visit us later this week at the UK Universities booth to learn more about our research and activities.

It’s fascinating how hard we had to work to get computers to be bad at maths.

Causing this to occur has been an ambition of mine since 2019: We did it! A joint uk universities booth at Supercomputing! Congrats to the team.

Suppose cost of software does fall 20x because of LLMs. (jury is out, but let’s explore). What does that do? It might make bespoke software for each org competitive with platform solutions. The value of better org-sol fit might be bigger than cost of build. Makes IT orgs builders not buyers.

Arsenal WFC vs Chelsea. #coyg

Uk hpc, what is your contingency planning around US ATC issues for SC25? I’m wondering if I can rebook through Toronto as that would be more robust?

We have people booked to bring broken food processors, favourite jumpers with damaged zips, and ailing plants. Don't miss this opportunity to repair your stuff with the help of 40 technicians and makers. Repair Café | Sustainable UCL - share.google/2qFlWYPHRPLs...

arxiv.org/pdf/2510.22254 “Ten Simple rules for Ai assisted coding in science” is very good. poldracklab.org/10sr_ai_assi... .

I am the cowpox of grumpy professors. Work with me so it’s safe to work with the real ones.

Data is the new sewage.

Double jabbed! Flu paid by UCL, covid privately. Would recommend.

Popes hate this one weird trick!

Research Infrastructure Engineers (RIE's) are the technical staff who build and maintain accelerated compute clusters. RIEs are invited to join the ACIT-Hub Training CoDesign Day on Oct 28 in Edinburgh to help define the training it offers: edin.ac/43f5VHR

caPitalising rAndom letters IN the title to Form the word yoU want is not cLever - PAINFUL. #reviewing

You know when you’re reviewing grants, and it’s like meh, meh, meh, WOW THIS IS THE BEST PROPOSAL EVER WISH ID THOUGHT OF IT, meh, meh, meh.

⏳ The DisCouRSE Network+ has launched its first Flexible Fund, offering up to £10,000 to support projects that strengthen leadership training and career pathways for digital Research Technical Professionals (dRTPs). Find out more at www.software.ac.uk/news/discour...

Let’s just go over again why low/no code platforms for data analytics and computational modelling are Wrong. I do this because I’ve just seen another advert for a tool “so all your colleagues can get insight from data regardless of technical ability”, and I need to call bullshit again. 1/N

Fermats margin proof was ‘sorry’ : #Lean4

What is the fractal dimension of the Jagged Frontier?

For advanced knowledge workers, there’s an assumption the job is 10% explicit learning, (self taught, conferences and formal courses), 20% on job learning, 20% faff, meetings and admin, and 50% productive. This will change. Maybe 75% explicit learning, 5% faff and 20% productivity time.

We’re excited to announce the International Research Software Conference (IRSC)! 7–8 September 2026 | Sheffield, UK & online | Co-located with #RSECon26 @society-rse.org Learn more: www.researchsoft.org/irsc/ For updates, join: preview.mailerlite.io/forms/778129... Image by @danielskatz.org

“Midjourney on the holodeck”. Star Trek TNG anticipated recent advances in generative AI, from autogenerated UX in the bridge consoles to narrative generation via prompt in the holodeck, to deep research with the ship’s computer in episodes such as Remember Me.

Beyond ‘we used ChatGPT’: a new way to declare AI in research. Recording which tasks were done when builds trust and transparency, say Yana Suchikova and colleagues. www.researchprofessionalnews.com/rr-news-euro...

Glorifying violence or harm violates Bluesky’s Community Guidelines. We review reports and take action on content that celebrates harm against anyone. Violence has no place in healthy public discourse, and we’re committed to fostering healthy, open conversations

Using a lot of Just Block today. The Bluesky social protocol is sound.

Just a reminder that despite the horror the world and all the people in it are also really interesting valuable and worth our love.

Extra tables being brought in to the room for our discussion on GenAI and research software engineering at RSECon25. It’s going to be interesting…

Is there an alternative to Conways law regarding how software architecture follows business model rather than organisational structure? (Good idea from Nicholas del Grosso #RSE25)

I’ve decided I don’t like the slido for all questions, no-hands-up thing. I know upsides: democratic vote of question rather than status, limits scope for people droning on. But the point of in person meetings is human engagement. Slido good for remote, social warmth cost too much IRL.

Never ask a man his age, a woman her salary, or GPT-5 whether a seahorse emoji exists

That‘s #bookhistory and #urbanhistory: exploring text and image communication in a social setting

(B)R&D and Circuses betterscienceproject.substack.com/p/the-counte... @ersatzben.bsky.social

There might be some weird biomodal effects from genAI on how deep education needs to go. Being moderately educated might be less valuable, but genuine deep expertise might be more valuable. What if it’s either stop at GCSE or carry on to PhD, with little benefit in the middle? Just thinking etc…

This made me laugh: "The AI industry is decades away, not years away, from achieving the level of safety, assurance, understanding, and professionalism that existed in the Chernobyl control room the night their reactor exploded anyway.” — Eliezer Yudkowksy

It is possible for these things to be simultaneously true: 1. The things GenAI can and will do well are a small subset of human potential, and improvements will saturate 2. The things it does well overlap with much of the current human labour tasks in the middle classes in the developed world

Not having to do the same information work twice is a powerful aspect of information as a reproducible good. Our improved ability to discover what has already been done and not have to do it again is a beautiful liberator. We just have to get used to the new economics carefully and kindly.

You know those moments when you’re trying to prompt the LLM to write some prose or code and it sucks, and it’s frustrating and it turns out faster to do it yourself? That’s the signal you’re doing real new intellectual work, not sampling in “already done by someone but you haven’t met them” space.

Our worry about AI in labour economics is really the discovery of how much of the knowledge economy was duplicative waste. We’ve been wasting minds and lives on duplicative information busywork.

Humans populated a very high dimensional space called Art. Economics drove the distribution in that space. Generative models can now sample within the more populated parts of that space. Can social systems help humans to generate art in the corners of the space where samples don’t yet exist?

Teaching is understanding: to teach something is the best way to really know it. Understanding for humans is “rules of thumb”: higher order concepts that help structure complexity. AIs aren’t as good as we’d like them to be at extracting and explaining such structures…

I’m fed up of reading papers where people say the reals when they mean the floats.