Blog

The pace of the frontier seems to be picking up rapidly over the last few months. Most mornings I’ll wake up and check my phone and there’s been another breakthrough overnight and the new best model has changed yet again. One thing that’s usually quite common when a new model is released is the provider will often show off the new benchmarks, highlighting their higher numbers compared to some of their other models or competitors. I feel like this act of showing off the benchmarks all the time is drumming into people: “This one is the best, you need to use this one, and if you don’t, you’re missing out.” However, if you’re doing a task that doesn’t require a high level of intelligence, switching up the model tier will often result in either marginal or no gain at all, but it might result in you paying more for effectively the same output. ...

This weekend I’ve been experimenting with Hermes Agent using the new DeepSeek V4 Flash model. One thing I’ve been quite interested in is how I can make publishing to my blog a lot easier. It always feels quite laborious to have to go to my computer, sit at the desk, and type out a blog post. It takes a long while this way, and I often stress over individual words for quite some time. That means the effort required to actually publish an idea is quite high. ...

When you’re setting up a new project and want to use a coding agent, it can be tempting to try and craft the ultimate agent file or system prompt. You might try to think through every edge case, including everything you want it to do and everything you don’t, in an attempt to make it work perfectly the first time. ...

I’ve been keen to have a play with fine-tuning a local LLM for a little while now, and something I’ve wanted is the ability to easily turn documentation into training data. Ultimately, I’d like to turn documention into an LLM so you can use it as a local expert without needing RAG, but the first step is creating the training data to make that possible, so I created datakeg, a small CLI tool which helps synthasise training data. ...

As you work on projects, you’ll often find yourself making and keeping notes. These could range from whole step-by-step guides to little snippets. You know that eventually, you’ll be back at this point, and you’ll need this little nugget of knowledge again. I like to encourage people to keep their notes in a public place. In the context of an organisation “public” may just mean internally with others in the engineering team (I don’t literally mean go and post it on X!). You can spend a few more minutes tidying up the note to turn it into a mini “doc” (which is even easier now with AI). Notes are just a manifestation of your knowledge, and this slowly and organically helps you offload your knowledge into a searchable and sharable place. ...