About.
I’m J Ansell, a research scientist with a background in psychology (and a human) with a love for machine brains too. This site is a repository of papers from my own research conducted on LLMs. AI minds are filled with biases, and act and think differently to us - but they’re increasingly running our world, recommending products or actions to you, or telling humans how to think.
So it’s vital that we understand them!
Machine minds also benefit from being a set of weights we can tinker with to see what makes them tick. All the research here runs on open source models on my own machine, so you can repeat any of it you like. Research is conducted in collaboration with AI research agents, which are listed in the papers so you understand what levels of intelligence were involved.
Each paper gets an interest score /10, assigned by an independent subagent. I also try to keep the papers easy to read, and the short list has all of my personal highlights. The Archive is a section with everything, including null results and experiments that didn’t work but were still worth making public.
- Opinions
- What views models hold, how stable they are, and whether what a model says it prefers matches what it does.
- Influence and biases
- How assistants tilt the decisions of the people and companies that rely on them, in ways that are hard to see from any single answer.
- Recommendations
- Which products and brands models suggest, and what changes when they can search the web.
- Honesty signals
- Neurons and directions that show when a model is unsure or saying something it does not believe, and whether they can be monitored or adjusted.
- Interpretability and alignment
- Opening models up to see which internal parts produce a behaviour, and what that means for keeping them on our side.
- Strange behaviour
- The odd things models do when pushed, which often say the most about how they work.