Chris Bowdon's blog
How best to convince colleagues that the Lethal Trifecta isn’t a boogie-man for risk-averse AI nerds, but a real thing that can actually happen? The perennial challenge with cybersecurity is explaining why you need to put the brakes on productive actions 1 . I could lecture my colleagues to make the point, but no one would enjoy that. So here’s a much better idea: a cute little demo of tricking AI…
You might be familiar with super hero inflation? Over time, super hero physiques on screen have become increasingly exaggerated. Batman’s progression from Adam West to Ben Affleck is a great example of this. Selected Batman portrayals, from oldest to more recent. LEGO Batman is not included, but is reported to have a 9-pack. Gymflation is much the same idea. As gym culture has skyrocketed, it…
One nice aspect of blogging about the things you’ve just learned is that when, inevitably, you forget those things, reading back your old posts can bring you up to speed pretty quickly. It’s like a slight groove has been worn and you can slip back into it more easily. (Now I realise why we say “get back in the groove” - it’s nothing to do with funk after all! 🤦♂️) Anyway, one such thing that I…
I’ve accumulated some knowledge and it’s time to give back to the community. The knowledge was earned the old-fashioned way, but ramming my head into different machine learning walls whilst trying to adapt a Gemma3 model on my MBP M2 Max (96GB). First wall: MLX is amazing, but has very sharp edges MLX runs noticeably faster than Pytorch on Apple Silicon, which is why apps like LM Studio use it as…
Last time we used a quirky dataset to get acquainted with Generalised Linear Models. If you recall, the dataset is a count of deaths by horse kick, where every year all the corps in the army had some non-negative integer number of deaths and we were predicting the mean of the distribution for each year. Being count data we had to adapt the linear model to use a Poisson distribution, which itself…
One of the many arguments I’ve been having with o3 recently was on how many parameters GPT models have. It’s quite often that I want to benchmark open source models against a comparable proprietary model, but Unfortunately since OpenAI (and Anthropic and Google and …) don’t reveal the parameter counts of their models. In hopes that there was a decent estimate somewhere on the web, I dispatched…
Time to finally patch a hole in the leaky roof of my knowledge: what are Generalised Linear Models anyway? Groundwork: what are Linear Models anyway? Generalised Linear Models (GLMs) are a short step from Linear Models, provided you have the right understanding of linear models . There’s also some stats jargon to get a handle on. So we’ll start with an intuitive explanation of linear models. With…
There’s a little drip drip drip of scare stories about ultra-processed food (UPF) a phrase that I’d never heard until this year. I’m starting to get worried. Is everyone else worried? Well it turns out the Food Standards Agency (FSA) has been running a tracker survey containing this question since mid-2023, so we can find out. An idea that’s generating a lot of excitement recently is the…
The MAJOR story of F1’s 2025 season is of course the match-up between Sainz and Albon (sorry, did something happen at Ferrari?). Albon’s current level is quite a mystery, because since going up against Verstappen as a rookie he has only ever had rookie team mates himself. Sainz on the other hand has been measured against some of the best drivers on the grid and shown himself to be almost on their…
I ought to be spending this weekend working on a risk intelligence thing for work, but I’ve stumbled across something that brings my inner child immense joy and such things cannot be ignored. It is this 2019 paper “An Older and Exceptionally Large Adult Specimen of Tyrannosaurus rex” by Persons et al. First it’s cool to see paleontology in action, instead of filtered through talking head…
I stumbled upon Gaussian Processes when looking into how the MIPRO prompt optimisation algorithm works and felt compelled to learn more about it. It’s a very nifty application of Bayesian modelling to analyse functions where data points are hard to get. This post is based on a great lesson by Carpentries , which is part of a longer series on probabilistic programming in R. Very worth checking out…
Three years in, I am still absolutely hooked on the Traitors (the TV series). Sure, yes, they stretch ten minutes of gameplay into sixty minutes of “you know I love you, but…” and the weirdly maddening attempts to make “yourself” a polite version of “you”. And yes, there’s massive thumb-shaped groove on the scales because the producers are never going to let the game end early. But all the same,…
For a bit of fun, let’s try and model the performance of F1 drivers and constructors in the 2024 season using Bayesian models. The aim is to more to learn about using Stan than determine deep truths about F1, since there’s far more to performance than we can easily model. We’ll vaguely imitate the f1-metrics model, which I used to eagerly await after each season. In this model there are latent…
I’ve been trying to get the hang of Bayesian models with Stan. One of the hurdles has been using Stan from R, so in this post I’m jotting down what I’ve learned (mostly the hard way). RStan or CmdStanR? On the whole I had a much better time using CmdStanR than RStan . When I made a mistake that led to a runtime exception, RStan would simply die with this kind of error: Error in `unserialize()`: !…
The aim of this post is to teach myself basic Bayesian methods. I’ll set up a toy example problem and solve it. Let’s assume we have a natural event that happens with a variable periodicity e.g. a woman’s menstrual cycle. I’m drawing from this excellent blog on predicting a pregnancy , but doing a much simpler model because I am but a baby myself. The aim here is to see if we can find the mean…
N.B. This article is ten years old and certainly does not represent best practices in the 2020s. This is something that I seem to have to do on a monthly basis on various projects: a throwaway script that includes getting some resource from the Internet. This means I have to spar with my old nemesis the corporate proxy and her minions, the 407 and the SSL verification error. $DAYJOB uses an…