Notes from actually building the things
Applied AI: what works, what doesn't, and what matters in the real world.
serverless is cheap until it isn't ... so design for the price list
Serverless compute is a wonderful deal when you're small. You only pay for what you use, it all scales to zero when not in use, and the free tiers are generous enough that a side project can live on them indefinitely. The catch is that the bill scales...
can a frontier model beat a trained specialist on a "reasoning" task?
Frontier models are going into jobs that used to need a purpose-built system. The pitch is attractive: just describe the task and plug in reasoning. This is obviously much easier than developing your own systems and models from scratch and drops the cost...
can you spot the firms with the most to gain from the outside?
AI is going into all sorts of business processes at the moment, with results that range from
using a benchmark on a specific task
Second of three posts. Part 1 covered building an honest benchmark honest enough to trust. This post is the reference companion: each retrieval method in turn - how it works, where it came from, what it typically scores in the literature, how I implemented...
building a benchmark you can actually trust
This is the first in a series of posts about improving search results using a real-world scenario. Inevitably, as it’s written in 2026, it’s also a bit about the process of using LLMs. This part covers the work that underpins the rest - getting a solid...