My third year Ethics and AI students are a mixed bunch (philosophy, law, psychology, teaching, literature, business, cyber security). It's a philosophy, ethics, and philosophy of information course, and so the idea is to use analytic philosophy, formal and philosophical logic, and normative ethical concepts and theories to ask the right questions and formulate solutions for AI-related ethical problems. One of the challenges is characterisation of the nature of transformers and their functions. I use a mix of the method of levels of abstraction, mechanistic philosophy, and functionalist philosophy of science. This presentation really helps the students.
1. Hardware: the physical 'layer' and its programmatic elements.
2. Simulation of Mechanism: The LLMs and their architecture simulating the brain's neural networks (Mechanistic interpretation.)
3. Emulation of Function: The semantic and polysemantic (multimodal) vector spaces and their concept mappings, emergent logic in induction heads, and emergent reasoning linked to linear concept vectors.
