Tactic 01
Andrej Karpathy, Lex Fridman
Treat Neural Nets As Alien Artifacts, Not Brain Simulations
The play
Describe the model by its training objective, not by comparing it to a brain.
Andrej Karpathy says
Karpathy is much more hesitant with the analogies to the brain than is typical in the field. A trained neural network is like a very complicated alien artifact. There was no multi-agent, self-play setup and evolution for it, just an optimization that amounts to a compression objective on a mass amount of data. Biological neural networks, by contrast, are trying to survive.
Try this
When analyzing a model's behavior this week, describe it in terms of what its training objective required rather than brain analogies.