Felix Rieseberg, who works on Claude at Anthropic, observes that model capabilities are advancing unevenly, "stunning in some directions, lacking in others" according to Terence Tao. Some users get remarkable results from them; many limit themselves to simple questions. Against the slogan "we're not ambitious enough," he offers a tutorial.
He distinguishes four types of ambition. Speed, doing the same work faster, is already achieved and requires no creativity; he sets it aside. What remains are altitude, depth, and range.
Altitude consists of moving up in abstraction: asking what the goal behind the goal is. The author moved from requests to fix crashes to the instruction to keep his applications crash-free, which forced him to give the model crash reports, the correspondence between versions and commits, and Windows, macOS, and Linux machines to reproduce and verify. Using the example of a quarterly business review at Acme Inc, he grades the approach: explain what changed and why, identify the three most urgent decisions with their arguments, then design a mechanism more effective than the quarterly deck, validate it against the historical record, and run it in parallel. The approach works for targeted outcomes, less so for creative work.
Depth consists of executing the task the way the best expert in the field would. Models readily change vocabulary without changing method: one must demand the instruments of the best, set unreasonable goals while presenting them as reasonable, have hypotheses checked against raw data, ask for new knowledge validated on unseen data, and above all rely on an outside judge, not the model itself. A typical prompt asks the model what a professional would do, what raw material they would start from, with what methods, and what would prove the result, before rewriting the request.
Range consists of calling on other disciplines, which multi-agent systems make economical. Two techniques: replacing nouns with verbs to see which professions are involved, and assembling the ideal team, red team included, by asking the model to do what each profession does rather than imitate its style. All these approaches assume a computer and access to data.
Rieseberg finally ties the problem to self-efficacy: people accomplish more when they believe something is possible, when they have seen someone else succeed at it, or when they allow themselves to fail. In a world where capabilities advance in leaps, he calls for sharing other techniques.