Self-driving cars have two dominant architectures and neither works well enough. The modular approach, dominant in the 2010s, splits the problem into Perception, Localization, Planning, and Control. End-to-end learning collapses all four into a single neural network predicting steering and acceleration directly, but produces an uninterpretable black box. Both camps are stuck.
The question this piece raises is specific: can Large Language Models break the deadlock. The author, a self-driving systems practitioner at ThinkAutonomous, walks through exactly how LLMs process information, starting with tokenization (converting words to integer tokens like 'a' equals 1, 'abracadabra' equals 121), through Transformer encoder-decoder architectures built on multi-head attention blocks, to next-word prediction as the core output mechanism. The LLM primer is not padding. It is the setup for understanding why language model reasoning might substitute for the planning module specifically.
The original article goes further into how LLMs could interface with real driving scenarios, which is where the argument either holds or collapses. The Fleming-penicillin framing is heavy-handed, but the engineering question underneath it is legitimate. Read it to evaluate whether the proposed LLM integration point is the planning layer or something more ambitious.
[READ ORIGINAL →]