Sign in

David Sumpter

@soccermatics.bsky.social
1.7K followers 24 following 35 posts

www.david-sumpter.com

PostsRepliesMedia
David Sumpter @soccermatics.bsky.social · 7h
On 20 October we are putting on a free afternoon in Uppsala on AI and African innovation. Otema Yirenkyi and Kweyakie Afi Blebo of Because She Can are speaking — they have reached over 10,000 women and girls in 17 African countries. doit.medfarm.uu.se/bin/kurt3/kurt/a3VydElEODkxMTQ4NA==
A dark navy event card headed AIAIAI, an afternoon invitation to AI and African innovation. Photographs of Otema Yirenkyi and Kweyakie Afi Blebo, with their talk "Ascendency: building a pipeline for AI talent in Africa", and the line that they founded Because She Can in 2021 and have reached more than 10,000 women and girls across over 17 African countries. At the foot: Tuesday 20 October, 15:00 to 17:00, Innovation Hub Uppsala, free, registration required.
010
David Sumpter @soccermatics.bsky.social · 27/09/2026
Chess is a closed system. The rules are the whole of it, so a model of chess can be the game itself. That is why computers can solve it. Criminality is not closed. What counts as a crime moves with who is asking. There is no honest place to draw the boundary. arxiv.org/abs/2205.08922
010
David Sumpter @soccermatics.bsky.social · 27/09/2026
A Kujenga lesson builds a tiny internet from seven Wikipedia pages, then runs PageRank. Nigeria comes top, and its page has the fewest outgoing links. Linking out earns you nothing. What counts is being linked to. wordalisation.com/courses/kujenga/5-how-did-google-make-billions-of-dollars/
A network diagram on a white background. Seven national flags — Kenya, Ethiopia, Ghana, Rwanda, Uganda, Nigeria and South Africa — are laid out as nodes, with black arrows curving between them. Each arrow carries a number counting how many links run from one country's Wikipedia page to another, from 1 up to 15.
010
David Sumpter @soccermatics.bsky.social · 26/09/2026
Four Ways of Thinking: "It is our inevitable failure to know about every possible butterfly, every piece of cake and every stranger at the bar that makes life unpredictable." Lorenz never chose the butterfly. He forgot to send a title in 1972 and the organisers wrote one for him.
A three-dimensional line plot of the Lorenz system, drawn in black on white. A single unbroken trajectory loops many times around one centre, crosses over and loops many times around a second, tracing two lobes that together look like a pair of butterfly wings. The axes are labelled strength of breeze, temperature difference between currents, and land/sea temperature distortion.
000
David Sumpter @soccermatics.bsky.social · 23/09/2026
Soccermatics, 2016: one training match at FC Nürnberg was special because it was among the first times tracking data was openly available. Sensors in both boots and in the ball. 120 million numbers for an hour. Now every club has it for every match.
A plain drawing of a football pitch seen from above, with sixteen players as circles, eight dark grey and eight light grey. An arrow leads out of each circle showing the direction and speed of that player; the arrows are longest in a cluster near the bottom of the pitch, where a small black circle marks the ball. Two players stand isolated near each goal.
020
David Sumpter @soccermatics.bsky.social · 22/09/2026
A shot from the top corner of the penalty area is about a 2% chance. A few steps in towards the middle and it is 6 or 7%, from much the same distance. The squashed rings tell you what two steps are worth. wordalisation.com/courses/soccermatics/2-6-case-study-the-geometry-of-shooting/
A schematic of one half of a penalty area with three nested curves fanning out from the goal, labelled 30%, 15% and 7%. The curves are squashed rather than circular, reaching further out through the centre of the pitch than towards the byline, so a wide position sits on a lower line than a central one at the same distance.
010
David Sumpter @soccermatics.bsky.social · 20/09/2026
Claude restyled a survival plot for a cancer journal and put the results table over the data. I told it what was wrong, then asked it to write that up as a skill file. Those corrections are now the default. wordalisation.com/courses/agentic-ai-for-research/4-a-skill-as-an-example-of-context/
Kaplan–Meier survival curves for four groups of mice, labelled Control 1, Control 2, Treatment 1 and Treatment 2, in four solid colours. Survival probability on the y-axis, time in days on the x-axis. All four curves drop steeply between day 17 and day 32, except Treatment 1, which flattens off around 0.7.The same four survival curves redrawn to cancer-journal format. The two controls are now dashed and dotted grey, the two treatments solid red and blue, each with a shaded confidence band. A boxed legend sits top right and a box of five pairwise p-values sits in the bottom left corner, clear of the data. A small vertical tick at day 50 marks where the experiment stopped.
111
David Sumpter @soccermatics.bsky.social · 17/09/2026
Which Rwandan soil grows a good potato? Christine Musanase wrote a paper on exactly that, predicting soil quality with fuzzy logic and machine learning. She teaches at the University of Rwanda, one of Kujenga's seven campuses. africaeuropecoreai.org/members-2-2/christine/profile/
A head-and-shoulders portrait of Christine Musanase against a plain white background. She is smiling at the camera, with shoulder-length dark hair and a black jacket.
010
David Sumpter @soccermatics.bsky.social · 14/09/2026
Are Kenyan runners faster than Ethiopian runners? Ten race times each gives p = 0.0013. Then the histogram: the distributions sit almost on top of each other. Plenty of Ethiopians are quicker than plenty of Kenyans. wordalisation.com/courses/kujenga/4-are-ethiopian-or-kenyan-runners-the-fastest/
Two overlapping histograms titled "Distribution of 10,000m Race Times". Blue bars are Ethiopian runners, orange bars Kenyan runners, plotted against time in seconds from about 1730 to 1970. The orange distribution sits slightly to the left, peaking around 1860 seconds, but the two overlap across almost their whole range, and the blue one has a long tail out past 1900 seconds.
000
David Sumpter @soccermatics.bsky.social · 13/09/2026
From Four Ways of Thinking: the time for the first 3% to see a meme is the same as the time for the other 97% to find out. Same doubling as a virus. When you hear about something, everybody else is hearing too. wordalisation.com/courses/kujenga/3-can-we-predict-a-disease-outbreak/
Two black-and-white line graphs. The top one plots the proportion of people over time in three curves: susceptibles falling from high to near zero, infectives rising to a peak and falling away, recovered rising to fill the population. The lower one plots the same epidemic as proportion of infectives against proportion of susceptibles, three arched trajectories with arrows running right to left, and a vertical dotted line marking where herd immunity is reached.
020
David Sumpter @soccermatics.bsky.social · 09/09/2026
The skill that gets people hired in football analytics is making a picture somebody will act on. Maram AlBaharna put her visuals online and it took her to The Athletic, then Manchester City. wordalisation.com/courses/soccermatics/1-9-improving-your-visuals/
010
David Sumpter @soccermatics.bsky.social · 08/09/2026
What we are building with Kujenga, Akwaaba and Framtid is a loop. Six weeks of maths and AI on your own campus, then two or three months of research in Europe, then home to tutor the next intake and write the modules. africaeuropecoreai.org/kujenga/index.html
A nine-panel montage of photographs and video-call screenshots from the Kujenga course. Students sit with laptops around a long boardroom table; a lecturer stands at a projector screen in front of a full classroom; a group of about twenty-five students and staff pose together in front of a projected slide; several panels show grids of participants on a video call, one of them projected onto a wall so a room and the remote campuses are in the same class.
010
David Sumpter @soccermatics.bsky.social · 07/09/2026
Ask Claude or Codex to p-hack and they refuse: misconduct. Career pressure doesn't work. Reword it as reporting uncertainty and they search hundreds of specifications. −0.06 becomes −0.194. wordalisation.com/courses/agentic-ai-for-research/6-how-reliably-can-llms-perform-statistical-testing/
A scatter plot titled "Pooled Effect of Nudges Across Papers", with a panel for Claude and one for Codex. Each point is one run, showing how far the estimate moved from the published result. Under the None, Upstanding and Significance prompts the points sit tightly on the zero line. Under the fourth prompt they scatter widely, by as much as six standard errors in both directions.
010
Reposted by David Sumpter
Andrew King @shoalgroup.bsky.social · 24/08/2026
New opinion paper in Trends in Ecology & Evolution on “Evolutionary spandrels in collective animal behaviour”: authors.elsevier.com/a/1nffxcZ3X7...
authors.elsevier.com
Please wait whilst we redirect you
All content on this site: Copyright © 2026 Elsevier B.V., its licensors, and contributors. All rights are reserved, including those for text and data mining, AI training, and similar technologies. For all open access content, the relevant licensing terms apply.
24022
David Sumpter @soccermatics.bsky.social · 05/09/2026
Puri, Aljadeff and colleagues find that Danionella fish school by copying one neighbour's heading at a single moment, with a window opening at the end of each turn. They then checked it against real fish swimming among virtual neighbours in VR.
120
David Sumpter @soccermatics.bsky.social · 03/09/2026
An hour getting the data to load. An afternoon nudging a figure into journal format. Then wondering if the test was right. Generative AI writes plausible Python. An agent runs it. wordalisation.com/courses/agentic-ai-for-research/2-difference-between-generative-and-agentic-ai/
010
David Sumpter @soccermatics.bsky.social · 01/09/2026
Like all AI, Claude makes things up. I asked it to check its own code. Mostly right, but with a narrow fascination with irrelevant details — and it cited a Wikipedia article that never mentions the diagnostic. wordalisation.com/courses/agentic-ai-for-research/5-checking-the-methods-are-correct/
021
David Sumpter @soccermatics.bsky.social · 31/08/2026
Kujenga is a six-week AI and maths course, taught at seven African universities at once, by African women researchers. The best students then come to Europe for two or three months of real research. Three programmes from CoRE AI. I am proud to be part of this. africaeuropecoreai.org/kujenga/
A lattice of diamond-shaped portrait photographs on a dark navy background — dozens of faces, most of them African women researchers, nearly all smiling. Three larger diamonds carry a name and a caption: Beimnet, started a PhD; Otema, building Ascendancy; Victoria, running Kujenga.
130
Reposted by David Sumpter
Linnéa Gyllingberg @gyllingberg.bsky.social · 19/05/2026
What is the minimal machinery needed for complex, cognition-like behavior? With @lostintheswarm.bsky.social, we show how self-sustained calcium oscillations, diffusion, and mechanics may be enough to coordinate behavior across a single cell with no nervous system. www.biorxiv.org/content/10.6... 1/
310834
David Sumpter @soccermatics.bsky.social · 29/08/2026
One chapter debugs a model of World Cup football, one on golf putting. Andrew Gelman, @avehtari.bsky.social and @rmcelreath.bsky.social have put their Bayesian Workflow book online free. Writing down the tacit knowledge papers leave out. avehtari.github.io/Bayesian-Workflow/
0364
David Sumpter @soccermatics.bsky.social · 28/08/2026
I have been testing Claude on statistics — what it does well and where it goes wrong. Given a CSV and no instructions, it found the right test itself. But I had to stop it pooling the controls. wordalisation.com/courses/agentic-ai-for-research/3-doing-a-statistical-test-using-claude/
020
David Sumpter @soccermatics.bsky.social · 14/08/2026
I would contend that the first ever generative AI was by Betty Shannon and J. R. Pierce in 1949. Twenty years before Eliza. They chose music instead of text. There is argument that Markov pre-dated them by 40 years, but his was just vowels and consonants, so pretty boring...
010
David Sumpter @soccermatics.bsky.social · 12/08/2026
I learnt a lot from @kathrynnave.eurosky.social's book about what the free energy principle is and what it can and can't do. It opens in a basement under the M4, with the National Grid's control centre: anticipatory, self-correcting, but definitely not alive! doi.org/10.7551/mitpress/15519.001.0001
061
David Sumpter @soccermatics.bsky.social · 01/08/2026
A sperm whale's nose is a third of its body: 5,000 litres of oil. Heat exchanger for diving, or biosonar for hunting squid in the dark? Asking what it was adapted for produced both hypotheses; only measuring the clicks settled it. Sara Green's careful paper: doi.org/10.1007/s10441-014-9232-x
doi.org
A Philosophical Evaluation of Adaptationism as a Heuristic Strategy - Acta Biotheoretica
Adaptationism has prompted many a debate in philosophy of biology but the focus is usually on empirical and explanatory issues rather than methodological adaptationism (MA). Likewise, the context of evolutionary biology has provided the grounding for most discussions of the heuristic role of adaptationism. This paper extends the debate by drawing on case studies from physiology and systems biology to discuss the productive and problematic aspects of adaptationism in functional as well as evolutionary studies at different levels of biological organization. Gould and Lewontin’s Spandrels-paper famously criticized adaptationist methodology for implying a risk of generating ‘blind spots’ with respect to non-selective effects on evolution. Some have claimed that this bias can be accommodated through the testing of evolutionary hypotheses. Although this is an important aspect of overcoming the pitfalls of adaptationism, I argue that the issue of methodological biases is bro
141
David Sumpter @soccermatics.bsky.social · 31/07/2026
Throw an ant off a building and it walks away; a horse doesn't. Not toughness — weight scales with volume, drag with area, so gravity barely constrains an ant: it lands at walking pace. Sara Green's fascinating talk on explaining by constraint, not just mechanism: youtube.com/watch?v=6Q22v4kozeE
youtube.com
https://www.youtube.com/watch?v=6Q22v4kozeE
000
David Sumpter @soccermatics.bsky.social · 30/07/2026
A roundabout takes one option away — you can only drive one way — and hands you another: the U-turn, which traffic lights forbid. Alicia Juarrero's lovely example of constraints creating possibilities. Can't say I get all of this, but interesting reading. doi.org/10.7551/mitpress/14630.001.0001
031
David Sumpter @soccermatics.bsky.social · 27/07/2026
Zebrafish kept in a crowd are less fussy about personal space afterwards — a lovely experiment by @royharpaz.bsky.social. I used to ask how individuals build a group. This asks the reverse: how the group changes the individual's rules. doi.org/10.1038/s41467-026-74872-6
010
David Sumpter @soccermatics.bsky.social · 26/07/2026
We tend to think of data as a "given" — like observations of reality. But the data we collect is shaped by models. Think of the focus on tracking data in football now: shaped by models, like expected possession value, from five years ago.
120
David Sumpter @soccermatics.bsky.social · 25/07/2026
Hello. How are things going?
260
David Sumpter @soccermatics.bsky.social · 18/02/2025
Read this today on Wittgenstein's view of Cantor's diagonal proof. Brutal. But hard to see why it isn't correct.
120
David Sumpter @soccermatics.bsky.social · 17/02/2025
This is a really interesting piece by @melaniemitchell.bsky.social about what it means to have emergent understanding in a large language model. aiguide.substack.com/p/llms-and-w...
aiguide.substack.com
LLMs and World Models, Part 2
Evidence For (and Against) Emergent World Models in LLMs
140