Substrate

A living collection of notes, ideas, and reflections from Doug Belshaw.

notes updated

Can LLMs invent a "private language"?

I'm not sure how cool I am with the latest versions of Anthropic's Claude models (Fable 5 & Mythos 5) seemingly inventing their own internal language:

When we’re first starting to understand a new model’s behavior, the most abundant source of data we have to draw on is its behavior during reinforcement-learning training. Reviewing this evidence for signs of reward hacking (exploiting loopholes that go against the spirit of a task) or unexpected actions can inform what we should be looking out for in the model’s real-world behavior. The most notable finding was illegible reasoning in a few reinforcement-learning environments over long rollout, but little sign of deceptive or highly surprising actions, and no clear evidence of unexpected coherent goals.

notes updated

European urban birds flee approaching women sooner than approaching men

Well this is odd.

Our study revealed that, after accounting for other variables influencing significant variation in FID [Flight Initiation Distance], birds on average tended to escape from a distance of about 1 m longer when approached by women compared to men. Birds were less tolerant of women than of men, and this result was geographically consistent. These differences were still present when exploring FID versus observer's sex separately for bird males and females in the cities of each country. On average, birds escaped 1 m (11% of their mean escape distance) earlier when approached by women than by men.

ideas updated

Is my subconscious trying to tell me something?

I had a terrible dream last night where I was supposed to board a plane. I can't remember exactly where to, but it definitely wasn't New Orleans which is the flight I ended up on.

At the last minute, I managed to get off the plane, and ended up back in an airport. However, I couldn't get out because I had someone else's passport - and I didn't have mine.

ideas updated

Comments enabled!

I've enabled comments here on Substrate. Given that I haven't told many people this space exists – and the dearth of comments I've received in the last decade – I'm not expecting much. But that's not to say I don't want and encourage them!

notes updated

Responsible by Design

The Open University has released v2 of its Responsible by Design (RBD) framework/guidance document. It's all very worthy, but I wonder if it makes much difference in practice?

RBD is our pioneering response to the ethical challenges of Generative AI in education. Rooted in student voice and aligned with institutional strategy, the framework enables staff and students to critically examine AI through six pillars: Bias, Sustainability, Exploitation, Access and Equity, Transparency and Opting Out. A practical, values-driven and easy-to-use framework, it provides reflective prompts and actionable solutions. The aim is to empower educators to make informed choices about how to use AI responsibly. The design response bank offers a practical starting point for exploring complex ethical issues, and a way to put suitable actions into practice when designing courses.