Weeknotes: Week 3 - Q1 - 2025
This is going to be another short one. With it still being school holidays i'm still spending more time with the kids than I am working. The weather has been abysmal too which hasn't helped. So what have I been up to over the last week? Well, I've started doing a fair bit of boring admin stuff, set up an MLX fine-tuning pipeline for agdata, tweaked the MTM site, started running again and started looking into doing something meaningful with open agdata.
Andrew ❤️ Xero
I'm a reluctant Xero user and to date i've used, well basically none of the features that it offers. I put in what we earn and we spend and that's about it so this week i've been reading a bit more about what you can do with it. More Than Machines is a small business so our accounting needs are pretty limited but there's some interesting stuff in there. One of the things first things I am thinking of doing is automating getting my receipts in there. I'm terrible at claiming expenses as i'm really rubbish with receipts so most of my expenses never make it back to my personal credit card. Yes there are 1000 apps out there for recording receipts but they're all too much for me and the one one that comes with Xero - hubspot I think it's called, is not good. I want to be able to take a photo on my phone and then run a script at the end of the week that pulls the info from the photos library I put them in, sticks it in Xero and then tells me how much money to claim. Now that local vision models are getting better, I think it's about time to give this another go.
I have also been looking into the feasibility of another idea i've had for using Xero. More on that in a later blog post.
Fine-tuning LLama 3.2 3B
At a high level, Fine tuning is the process of getting a large language model to focus. Large Language Models are trained on vast amounts of data, the vast majority of which has nothing to do with the problem domain in which you want to use one. Fine tuning - in this case using lora, uses training data to adjust a model's weights and help it focus using the data it already has knowledge of, this is different to RAG which primarily solves the problem of giving a model access to information that it does not already know. MLX is a Python and C++ API for efficient and flexible machine learning on Apple silicon - and given More Than Machines runs on Macs i've been meaning to pull together a fine tuning pipeline for a while. Now it's done. I've been specifically looking to finetune some of the smaller LLama models and so far the results are pretty good. I'm writing this up in another blog post with a lot more detail and a few benchmarks so will share soon.
MTM website tweaks
I mentioned in last week's weeknotes that I've been making some updates to the More Than Machines website. By the time you read this, I should have pushed them up onto the site. Nothing earth shattering. I've just added some information about our priorities, some of our recent work and also polished up a bit of the copy. The work i've added to the site isn't particularly inspiring but i'll also give that a shakeup in the coming weeks. We've done so much great stuff. I've also given the blog section of the site a bit of an updo and will start posting a bit more there too. I've dialled the More Than Machines social footprint right back to just LinkedIn now (i'm also thinking about Bluesky), this means that the blog is going to become a much more important way to communicate. While updating the blog and thinking about social media, I also made a little script that automates the creation of images to use on linkedin and other sites. It integrates well with my other AI workflows for automating content creation. You can see an example of it in action here.
Exercise is go!
One thing I haven't mentioned in my weeknotes so far is that I am trying to get back to running at least 50kms a month. It's been a slow start so far, basically because of illness and very wet weather but I managed to get out for a run last week. Hopefully this coming week i'll be able to go a couple more times. My pace is way down and i've a bit of a hill to climb to get back to where I was. It doesn't help that during the hiatus, the sole has fallen off my running shoe so i'm running with one show held together with gaffer tape at the moment.
Finally
One of the things I put on my todo list for this year is to get better engaged with open data, especially in agriculture but not only there. I've been scheming a bit this week about what that might look like, especially in light of what i've been reading about the lack of available data to support those affected by the recent LA fires. Despite all the talk, there's still not a lot of good open datasets in Australian agriculture, or if there ios they're not easy to find. I'd like to do something about that. I'm not sure what yet but I have some ideas and I have a plan for trying to rope some others into the cause.
Well that's it for another weeknote, next week i'm going to be starting to ramp up my time in the studio a little more so should have some more interesting things to talk about.