If you are interested in the hype about AI recently, this is a very relevant result.

The authors have shown that a 14-line piece of code using gzip can be used to 'understand' similarities between sentences.

This illustrates how a culture builds up around what is considered good research.

Hype builds up about certain benchmark tests and methods, and we forget that (some of) the reason we are making rapid progress is that it is only now that we consider the question interesting.

Couldn't resist (finally) asking chatGPT a few things about myself. Was very enjoyable to find out that I had helped Brentford to success in 2014/15😮(I wasn't even working with football then 🤣 ). I went on to help Midtyllland, signing Pione Sisto.

The Future of Life Institute is a problem.

Being in same age-group (lower end maybe😃)/cultural background (8-bit programming📼⌨️) as these men, makes me feel uncomfortable and embarrassed for them.

Here they (Musk, Hassabis, Tegmark, Bostrom etc.) are in 2017. I'm embarrassed because part of me wouldn't mind idling an afternoon away talking about what the mind is and what AI has to hold.

But another part knows it is wrong. They want to think in the loosest possible terms, while insist on being listened to by world leaders.

Expected Threat (xT) is one of the most important (but least well-understood) statistics in football. ⚽️⚽️

It was used by Liverpool (they call it Goals Added) to do some of their best scouting. 📈

And now, YOU can learn all about it...🧵 The basic idea of xT is to assign a value to a position on the pitch or an action based on the probability it will lead to a goal.

A thread on Liverpool using numbers.

Six games is great for using expected goals and expected threat to find out what is going on.

First of all they have 'won' all six on xG. They are best in league at getting the ball in to the final third and in to the box per match.

How should we measure the performance of a machine learning model? 😕

When I was teaching ML for the first time last year, I was surprised to find there was no agreed upon single number which measures model performance. 🤯

So I decided to look at the question myself... 🧵 Here is the question. Imagine you have created an algorithm which assigns a score to an image based on how likely it is to contain a cat. 🐈‍⬛

The pictures and the scores from your algorithm might look something like the following.

In the new spirit of me writing about what I am working on each day, here is a random thread about the 'bespoke football analytics' which we are working on at @twelve_football and featured in the @AnalyticsFC podcast. The image below is a player radar for Adama Traore from this season in PL before he went to Barcelona. It shows that he is top rated for high-speed dribbles, creating his own chances and maintaining possession. He is also good at cutbacks and through balls.

I have spent much of today studying the Stochastic Parrots paper by @emilymbender, @timnitGebru, @mcmillan_majora and @mmitchell_ai. (faculty.washington.edu/ebender/papers…) This paper is important for many reasons... It explains why language models like GPT3 are in large part gimmicks: stochastic parrots. They are just generating randomish sequences of sentences from the data put in to them. (1/n)