Day 2: builders wire Claude into their workflows while interpretability researchers crack open arithmetic

From the editor

Two days in, the Claude community is already splitting into two productive camps. The builders are squeezing latency out of their day: a Korean instructor wraps a quiz tool in MCP, and practitioners are mapping out what 200K context actually buys you beyond marketing slides. Meanwhile, the researchers are going the other direction, using neural language autoencoders to show Qwen solves multiplication by analogy, not arithmetic. The meta-pattern: 'next token prediction' is a training story, not a behavior story. Use the tools accordingly, and stop trusting your intuitions about what is happening inside.