How to be more ambitious: A tutorial for humans with models

Felix Rieseberg works on Claude at Anthropic, and he thinks many of us are still using increasingly capable AI models far below what they can actually do. His new essay tries to turn the familiar advice to “be more ambitious” with AI into something more concrete.¹

Rieseberg divides that ambition into four directions: speed, altitude, depth, and range. Speed is the familiar one, doing the same work faster. The more interesting ideas involve asking for the outcome beyond the immediate task, letting the model work directly with raw data and professional tools, and bringing techniques from other disciplines into problems where they normally would have been too expensive or time-consuming to involve.¹

One of the essay’s stronger points is that this often requires changing the environment around the model, not merely writing a cleverer prompt. Rieseberg describes giving Claude access to crash reports, source history, and machines running multiple operating systems so it can move from fixing individual crashes toward independently keeping an application crash-free.¹ His argument also overlaps with decades of goal-setting research: specific, difficult goals have repeatedly produced better performance than vague instructions to simply “do your best.”²

The article gets increasingly interesting as it moves from prompting advice into a broader question about human expectations. Rieseberg uses examples from psychology, mathematics, engineering, and history to argue that what people believe is possible can constrain what they attempt in the first place. One of those stories, George Dantzig unknowingly solving two unsolved statistics problems after mistaking them for homework, is documented by UC Berkeley and fits the essay’s larger theme unusually well.³

There is considerably more here than a list of AI prompting tricks. Rieseberg is really proposing a way to reconsider the scale of the work we hand to models, and several of his examples are worth seeing in full. If you regularly use Claude, ChatGPT, Codex, or another capable model for serious work, the original essay is worth reading from beginning to end.