Summary
A technical talk with Anthropic's Thariq Shihipar aimed at builders shipping real LLM systems—where routing, evaluation, and reliability start to matter more than model vibes.
Why it matters
Provider-adjacent technical talks are useful because they surface recurring failure modes seen across many teams: tooling brittleness, safety boundaries, and the operational costs of agentic workflows.
Tokenmaxxing angle
If you're going to use a lot of AI, you need constraints. This is the mindset shift from tokenmaxxing as a scoreboard to tokenmaxxing as operations: measure outcomes per token and enforce budgets, eval gates, and safe tool access.
Source takeaway
Good for updating your mental model of what "serious" LLM engineering looks like right now—how practitioners think about reliability, routing, and cost as first-class design constraints.