Ambition with Models

Felix Rieseberg (a Claude engineer) turns the AI-practitioner rallying cry “we’re not ambitious enough” into a tutorial: model capability is uneven in the way Terence Tao describes — “astounding in some directions, while inadequate in others” — so the bottleneck in impressive results is usually the scope of the assignment the human gives, and he organizes that scope into four types of ambition: Speed, Altitude, Depth, and Range.¹

The four types

• Speed — the same work in less time. The established economic case for AI assistance; Rieseberg skips it as requiring little creativity in approach.
• Altitude — assign the outcome rather than the task. “Fix this crash” becomes “keep our apps crash-free”; a quarterly review becomes “tell me what changed, why, and what to do about it,” then “work out the three most urgent decisions,” then “design a better mechanism than the quarterly deck and run it beside the current one for a quarter.” The tell for altitude work: if the output will be discussed in meetings, the goal above the task is probably assignable. Durable out-of-the-loop workflows live here, and they usually require reworking the environment first — crash-report fetching, version-to-commit mapping, and machines on all three desktop platforms were prerequisites to assigning crash-freedom.
• Depth — push the task to field-leading practice: insist on the instruments top professionals use (databases and statistical models over markdown tables), set unreasonable targets and treat them as solvable, give the model raw data and ask it to adversarially reproduce your assumptions, ask for new knowledge from inside the data (a customer simulation that finds an early churn signal predicting on held-out data), and validate with an outside judge — in October 2026 self-judging does not hold up.
• Range — pull in fields the task doesn’t obviously need. The noun-to-verb trick (from Kelley and Littman’s The Art of Innovation) exposes which fields apply: a quarterly review is “getting busy people to believe a story about cause and effect, and to decide under uncertainty, in 30 minutes.” Then assemble the crew of fields — analyst, journalist, editor, motion designer, short seller, forensic accountant — and ask the model to do what those roles do rather than to write characters performing them; role-play degrades the transcript without improving the output.

All four assume the model has a computer it can install software on and data access (MCP connectors or equivalent).² The altitude examples assume outcome-controllable work; Rieseberg notes the current limit plainly — “Claude, go make money” does not work for most businesses by itself in 2026 — and says altitude suits creative work less.

The psychology underneath

Rieseberg grounds the tutorial in two research traditions. Self-efficacy (Bandura’s social-cognitive theory): task-specific confidence raises the goals people set for themselves, their commitment to assigned goals, the quality of their task strategies, and their use of feedback. Goal-setting theory (Locke & Latham’s 35-year retrospective): performance rises linearly with goal difficulty until ability limits or lapsed commitment intervene (meta-analytic d .52–.82), and specific difficult goals beat “do your best” exhortations consistently (d .42–.80) — do-your-best goals have no external referent, so people define acceptable performance idiosyncratically. Goals operate through four mechanisms: directing attention and effort toward goal-relevant activity, energizing effort, extending persistence, and driving the search for task-relevant knowledge and strategies. Moderators include goal commitment (raised by importance and by self-efficacy), feedback (goals plus feedback beat goals alone), and task complexity (complex tasks favor learning goals and strategy training over raw performance pressure).³

The essay’s historical cases are demonstrations of these effects: Dantzig solving two unsolved statistics problems he mistook for homework, Bannister’s four-minute mile falling to a second breaker 46 days later (and London 2026’s sub-two-hour marathon producing two under the mark in one race), and The Inner Game of Tennis’ permission-to-fail thesis as the hobbyist posture that let Linux start.

What the page is about

The four-type taxonomy is a prompt-design heuristic tied to current model capabilities and will age with them; the goal-setting and self-efficacy results underneath it are fifty-year-old empirical findings. The stable claim is that assignment scope is a lever with measurable psychological mechanics, and the taxonomy is one practitioner’s current map of that lever. For the adjacent question of how mathematicians are reworking practice around model capability, see ai-mathematical-practice; for the refusal-side response to the same capability moment, see butlerian-jihad-thesis.

Related

• ai-mathematical-practice — Tao’s “Math 2.0” and community norms for AI-assisted proof
• llm-context-engineering — the tooling and data-access preconditions the tutorial assumes
• butlerian-jihad-thesis — the opposite posture toward the same capability curve

Sources

• Felix Rieseberg 2026 — How to be more ambitious: A tutorial for humans with models
• Edwin A. Locke, Gary P. Latham 2002 — Building a Practically Useful Theory of Goal Setting and Task Motivation: A 35-Year Odyssey

Footnotes

1. Felix Rieseberg 2026 — How to be more ambitious: A tutorial for humans with models ↩

2. Some of the workflow-precondition claims are elaborated in llm-context-engineering’s treatment of tool and retrieval channels. ↩

3. Core findings, mechanisms, and moderators from Edwin A. Locke, Gary P. Latham 2002 — Building a Practically Useful Theory of Goal Setting and Task Motivation: A 35-Year Odyssey; the self-efficacy and “hard goals” connections are the essay’s reading of that literature. ↩