Each second thousands of planes send messages about their location and flight data - in ADS-B messages. These are intended to prevent mid-air collisions, but we can exploit the aggregated data to map out atmospheric conditions. By deriving the wind speed from the broadcast flight data, we can model wind speed and direction.
This post is a little history lesson in the world of natural language processing. It should give a pretty practical overview of techniques (Markov Chains, BPE, GPTs) used to generate text since 1948 (Claude Shannon’s seminal paper) to today’s LLMs.
Instead of simply enjoying a cool board game, I dissect and explore the different approaches to solving the imperfect information game Take-It-Easy. I compare basic heuristics to a more sophisticated reinforcement learning approach. What’s the best a computer can do?
You often don’t need to **know** (be sure) to be **right**. To be right more often, you need to know more, which helps improve your heuristics. And there are some bits of knowledge which can help you be right way more often than others…
While reading through the fantastic book *[The Lost Art of Logarithms](https://www.lostartoflogarithms.com/)* by [Charles Petzold](https://www.charlespetzold.com/) I was nerd-sniped by a [simple method](https://www.lostartoflogarithms.com/chapter04/) of estimating the logarithm of any number base 10. According to the book, it was developed by [John…
Inspired by multiple posts on dithering, I set out to create a correct implementation of Atkinson dithering with support for RGB colour palettes. The post also outlines implementing linearising the colour-space and correcting for human perception.
Have you ever wondered how radio works, and why different radio broadcasts don’t interfere? Why you can listen to classical and your neighbour to rock, at the same time? The answer is FM - frequency modulation - those magic letters on the radio set dials. The post describes simulating FM radio and my own stations to understand how broadcasting and tuning in works.
Looking back at my performance in my first every Metaculus’ Quarterly Cup (Q4 2024), there’s a lot of room left for improvement. This post compares the different forecasting methods I used and what their results were. Thanks to some analysis, I then figure out what I can improve: stick to reasonable probabilities; use fat-tailed distributions to forecast; avoid the trap of nowcasting; update my…
I’m not sure you can really call someone rating their least favourite dish with 1/5 in an informal poll on my website voter fraud, but I sure felt like a detective uncovering a crime while looking through my logs and analytics. It also makes me look better than saying: “I allowed voting multiple times because of lacking security.”
Phishers tried to steal my Instagram credentials. They also forget to properly secure their database…