~ further reading

Links

Curated readings on how LLMs work, model risk, and the economics of AI.

How LLMs work

How a (small) language model walks through its training text

Chris Wiggins's interactive demo of Shannon's 1945 bigram model. Walk through a Markov chain to see exactly how next-token prediction works — no neural network required.

Read

Model risk and strategy

Anthropic: Statement on Fable 5 and Mythos 5 access

The US government suspended access to Anthropic's frontier models over a jailbreak finding. A case study in why locking in to a single SOTA model is a major risk for research pipelines.

Read