@gonzalo__nunez Large orgs might actually crack this through sheer volume. Run enough similar tasks & start to build patterns "code reviews on repos this size," "summarizing this class of document," etc. You build cost archetypes, apply a p90 buffer, can kind of get a rough estimate?
@gonzalo__nunez I tried building this in CC, gating execution for a task behind users approval of an approximate token budget. technically possible as an MCP server (agents can call), w/ a set of hook scripts.
Problem is that it's a ~blackbox. Anthropic's Interpretability team is our only hope
@nabeelqu How do you think they could use AIs to gain an edge?
Alireza was prob the one, but everyone's training engine-assisted
One way might be using AI to train suboptimal lines & beating players in blitz by pushing them out of book. Pragg might've done this in 2024
creative chaos rests on order
low-entropy carriers bear high-surprise information
"Steve was preoccupied with the nature & quality of his own thinking" said Ive.
Obsessed with the quality of your thinking, you outgrow displays of taste & only care to Make Something Wonderful.
@internetvin Interesting way to do this would be for them to talk about what they've been writing about recently i.e open up their Obsidian/Notion/whatever. People are routinely writing interesting things informally & don't present it but a conversation/podcast format is perfect to explore it