In one sentence
Use capable models more ambitiously by moving beyond faster task completion toward higher-level outcomes, expert methods, external validation, and useful techniques borrowed from other fields.
Overview
Felix Rieseberg argues that telling people to be more ambitious with AI is too vague to be useful unless it comes with concrete methods. He divides ambition into four dimensions: speed, altitude, depth, and range. Speed is familiar—doing the same work faster—so the article concentrates on the other three. Altitude means asking for the outcome beyond the immediate deliverable. Depth means working from raw material with the tools and standards of leading practitioners. Range means bringing relevant methods from several disciplines into the task. Rieseberg illustrates each dimension with a quarterly business review, then connects the framework to self-efficacy: people attempt and achieve more when difficult goals feel possible. The article’s practical message is that better model use depends as much on widening the human idea of the task as on improving the model itself.
Core ideas
Speed is the least imaginative form of ambition
Using a model to write code, analyze material, or prepare a document faster can be valuable, but it leaves the shape of the work unchanged. The more interesting gains come from reconsidering the requested outcome, the standard of execution, and the range of knowledge brought to bear.
Altitude starts with the goal beyond the goal
A task is usually one proposed route to a broader result. Instead of merely updating a quarterly deck, ask what leadership needs to notice and decide, then consider whether a different system could surface problems earlier. Higher-altitude work shifts attention from producing the requested artifact to improving the outcome it exists to support.
Durable automation requires changing the surrounding workflow
Asking a model to keep an application crash-free is more ambitious than asking it to fix one crash, but the higher-level instruction only becomes practical when the model can retrieve crash reports, connect releases to commits, reproduce failures, and verify fixes across platforms. Ambitious prompts need suitable data, tools, permissions, and feedback loops.
Depth means adopting expert methods rather than expert vocabulary
Models can sound professional while using shallow methods. Rieseberg recommends starting from raw data, using the instruments a strong practitioner would use, testing assumptions adversarially, setting unusually demanding targets, and asking the model to discover new predictive signals rather than merely summarize known ones.
External judgment is part of serious model work
The model and its operator should not be the only judges of success. Tests, measurements, held-out data, reproducible experiments, and domain review provide evidence that a result is correct. The deeper the task, the more important an independent feedback mechanism becomes.
Range comes from translating artifacts into actions
Replacing nouns with verbs reveals the real work. A deck becomes getting busy people to understand cause and effect and make decisions under uncertainty. Once the task is described that way, useful fields become easier to identify: analysis, journalism, editing, motion design, forensic accounting, or adversarial research.
Multidisciplinary work should use real techniques, not role-play
Assembling an imagined team is useful for discovering methods and perspectives. It becomes empty when the model merely adopts a persona. The point is to perform the research, tests, editing, or production practices used by those disciplines, not to imitate how their practitioners sound.
Self-efficacy sets the ceiling on attempted work
People often limit model use to what they already believe is feasible. Rieseberg connects greater ambition to research on hard goals, examples of previously impossible achievements, and permission to fail. Models expand capability quickly, but people still need ways to revise their sense of what can be attempted.
Practical takeaways
- Before prompting, state the result that should happen after the immediate deliverable is finished.
- Ask whether the requested artifact is the best mechanism for producing that result.
- Give the model raw evidence and access to the tools required to test its conclusions.
- Name the methods a leading practitioner would use, then require those methods rather than a professional tone.
- Set a measurable target that forces a different approach instead of accepting incremental improvement.
- Use independent tests, held-out data, instruments, or qualified reviewers as the judge of correctness.
- Describe the task with verbs to uncover relevant fields and assemble a useful mix of techniques.
- Build repeatable systems around successful prompts so the model can pursue outcomes without constant task-by-task direction.
Caveats and counterpoints
- Higher-altitude delegation can hide important choices if the desired outcome and constraints are vague. Human judgment remains necessary when values, priorities, or trade-offs are contested.
- Unreasonable targets can stimulate new approaches, but they can also encourage misleading claims or wasted effort unless success is measured independently.
- Giving models broad access to data and computers increases their usefulness and also increases the need for careful permissions, privacy controls, and review.
- Multidisciplinary range can become superficial if the operator collects perspectives without understanding when their methods apply.
- Some tasks genuinely require a bounded deliverable. Reframing every small request as a larger system can add unnecessary complexity.
- The self-efficacy examples are motivational illustrations rather than proof that confidence alone produces exceptional outcomes.
Questions worth revisiting
- Which recurring tasks would benefit most from moving from individual fixes to outcome-level ownership?
- What raw data or tools are currently missing from the workflows where a model performs poorly?
- Which external tests could replace subjective model self-evaluation?
- When does a deliberately extreme target reveal a better method, and when does it distort priorities?
- Which fields have techniques that would improve the work, beyond merely giving it a different voice?
- What safeguards should accompany broader computer and data access for increasingly autonomous workflows?
Return to this when…
Return to this article when AI use feels efficient but untransformative, when a model keeps producing polished versions of the same artifact, or when you need to turn a one-off prompt into a durable workflow. Its most useful diagnostic is to ask whether the task needs more altitude, more depth, more range, or simply more speed.