What’s on my desk right now
Now
This is a snapshot of what I'm doing, learning and trying to get working at the moment. I update it when something important changes. Once there are older updates, you'll be able to read them in the archive.
On this page
What am I building?
One of my main projects is AI Interview Trainer, a job interview simulator that should be able to hold a natural practice interview for a job chosen by the user.
I'm also rebuilding this personal website. The aim is to make a place where I can talk about my projects, thoughts and activities without trying to fit everything into a CV or a professional introduction.
Other projects in progress, or waiting their turn, include making music, local AI experiments and cryptocurrency market analysis. I'm not moving them all forward at the same time.
What am I learning?
Local language models, building AI systems and how to use AI in software development while still getting a reliable result.
I'm also following developments in quantum computing and looking for uses of AI that could lead to something useful—and possibly commercially interesting too.
I'm particularly interested in how far AI can take one person trying to build something that would previously have needed a whole development team.
What isn't working yet?
The interview simulator's local AI still doesn't work as well as I'd like. Language models don't always understand the interview situation correctly, responses can be inconsistent, and Finnish has caused problems. I've also been investigating cases where the interview gets stuck or the AI's response isn't accepted by the program.
I have more interesting projects than time to carry them out. A new idea can seem so appealing that the previous one is left waiting. This isn't a new problem, and I haven't found a particularly good solution yet.
AI can already do astonishing things, but sometimes sorting out its mistakes takes longer than doing the work itself. I'm especially interested in how to make it handle long, complex tasks reliably without constant supervision.
What have I learned lately?
Working with local language models has shown me that a model's size or speed alone doesn't tell you very much about how useful it is in a real application.
A model can give a perfectly sensible answer to one question and then fail in a longer conversation, forget what was said earlier or misread the situation.
I've also noticed that the hardest part of AI-assisted software development isn't always getting AI to write code. It's making sure all the pieces work together and the program actually does what it's meant to do.
What would I like to finish next?
This website. And the interview simulator, to the point where it can actually be used.