This article explains how to move beyond using AI merely to finish tasks faster and instead set higher goals, investigate more deeply, and connect more fields. The author organizes this approach around Altitude, Depth, and Range, and argues that self-efficacy, our belief in our own abilities, matters even more in the AI era. The point is not to give models more work, but to change the level at which we frame problems.

1. Remarkable Capabilities, Insufficient Ambition

Working on Claude, the author encounters uses he had never imagined almost every day. Claude may have helped an amateur astronomer discover a new planet; he has seen the first AI music video he actually likes; and discussion continues about a new era in mathematics. But model capabilities do not advance evenly in every direction.

As mathematician Terence Tao puts it, models can be astonishing in some areas while still falling short in others.

"Marvelous in some directions, inadequate in others."

Some people are already steering these capabilities to achieve results that feel like flying. Many others do not know how to use their new wings. They still ask only questions such as "What is the capital of France?" or "Make this email sound less angry."

The author agrees with the AI industry's frequent refrain: "We need to be more ambitious." People repeat it whenever they see an impressive example, but putting that advice into practice is difficult. This article offers practical ways to become more ambitious with AI.

He identifies four directions:

• Speed: Doing the same work in less time, such as writing code faster. This familiar, established source of AI's economic value is not discussed at length here.
• Altitude: Looking beyond individual tasks to the outcome you ultimately want.
• Depth: Reaching into primary sources, tools, and methods to tackle problems as a leading expert would.
• Range: Drawing on multiple disciplines and skills instead of staying within the one field the problem appears to require.

He mostly uses Claude, but emphasizes that the advice is not limited to one model.

2. Altitude: Ask for the Outcome, Not the Task

Altitude means entrusting the model with the ultimate goal at a higher level of abstraction rather than giving it a small task. When assigning tasks, we have usually already guessed how to reach the desired outcome. Raising altitude starts by asking:

"What do I actually want to happen afterward?"
"What is the real goal beyond this goal?"

The author used to ask Claude to fix individual app crashes. Over time, his request changed to "Keep our app crash-free." He delegated the outcome of app stability, rather than the task of fixing a bug.

Three Stages of Preparing a Quarterly Business Review

He uses a hypothetical Acme Inc. quarterly business review, or QBR, deck as an example. Filling numbers into a slide template seems mundane compared with discovering a planet. But raising the approach step by step can make that same work much more ambitious.

A typical review asks what changed this quarter, why it changed, and what to do. Resist the temptation merely to update numbers in an old template. Give the model as much data as possible so it understands the business, and ask:

"Tell me what changed this quarter, why it changed, and what we should do."

Next, recognize that a good meeting should end in decisions. Instead of simply making slides, ask Claude to identify the most urgent decisions.

"Based on everything you have learned, identify the three most urgent decisions and explore possible directions for each, including arguments for and against."

Finally, reconsider why the QBR exists. It is a format designed for people to examine data, diagnose problems, and allocate money and staff. Is a quarterly deck really the best way to do that?

"This review exists to help executives detect problems early and move money and people to the right places. Design a better approach than a quarterly deck, show using our historical data that it would have caught problems earlier, and run it alongside the existing process for one quarter."

Raising altitude is especially powerful when there is a clear ultimate outcome. It may fit purely creative work less well. If you feel you need a meeting to discuss the work you are doing, that may signal a need for greater ambition in altitude.

There are limits. Even in 2026, "Claude, go make money" is not enough to run most businesses. Still, growing capabilities are expanding the range in which models can work at higher levels of abstraction.

Altitude also reveals durable automation workflows that do not require constant detailed instructions. Keeping an app crash-free takes more than granting permissions. The entire workflow must let the model retrieve crash reports, understand how app versions map to Git commits, reproduce bugs on Windows, macOS, and Linux, and verify fixes.

3. Depth: Reach the Primary Sources and Verification Methods Experts Use

Depth means handling a task as a leading expert would, beyond merely doing it well. Go beneath the surface to the tools, techniques, media, knowledge, and raw data experts use. Ultimately, aim for results that could teach even an expert something new.

Returning to Acme's QBR, suppose the company struggles with churn, customers leaving the service over time. The author suggests several ways to investigate deeply.

Make It Actually Use the Best Tools

Telling a model to "act like an expert" often changes its vocabulary without changing its methods. Require the actual tools and analyses used by leading data scientists and statisticians, rather than a professional tone.

If Claude plans to produce a Markdown table, for example, require a database and statistical models instead. Users need not know every specialized tool: models generally know what experts use.

Set Unreasonably High Goals

It can help to set a much higher goal than seems realistic and tell the model it is solvable. If you want website load time reduced from ten seconds to five, ask for 200 milliseconds instead. Even if the model judges it impossible, reiterate that the problem is solvable and it should keep trying.

This is not simply about making unreasonable demands. It seeks to overcome both the low expectations people unconsciously accept and models' tendency to stop too soon.

Challenge Existing Beliefs With Raw Data

Many beliefs, especially in business, are only rough hypotheses about reality. Provide raw information and actual data, not polished reports, and have the model test assumptions adversarially.

If the company believes churn is an onboarding problem, do not have the model accept that belief automatically.

"We believe churn is an onboarding problem. Use the raw data to reproduce or refute that claim."

Ask It to Discover New Knowledge

Once a model is deep in the data, it can go beyond testing hypotheses to discover blind spots and new signals. The author suggests asking for a customer simulation replaying a year of logs, then finding the earliest unknown signs of churn and testing their predictive power on unseen data.

"Build a customer simulation that replays the past year using our logs. Then find the earliest churn signal we do not yet know about, and show that it predicts churn on data you have not seen."

Do Not Let the Model Grade Itself

The author particularly stresses external verification. A model should not evaluate its own output. Greater intelligence may eventually improve this, but as of October 2026, he considers external verification essential.

Ideally, the model can run the verification tool itself and receive immediate feedback: automated tests for code, held-out datasets for analysis, or real user behavior and independent measurement for products.

He offers this prompt for deeper work:

"I want a deeper version of this task, not a bigger one. Before rewriting my request, tell me:

1. What would a leading expert in this field do that I have not asked for?
2. What are the rawest sources they would start from, rather than processed information?
3. Which specific methods and tools would they use?
4. What, other than you or me, could demonstrate that the result is correct?
   Then rewrite my prompt to ask for exactly that."

4. Range: Bring Multiple Disciplines to One Problem

Range means drawing on perspectives and techniques from multiple fields rather than remaining within the single specialty the problem seems to require. Hiring many experts used to cost too much. AI products now make multi-agent systems, parallel workflows, orchestrators, and subagents easier to assemble, making this approach more practical.

The author notes that the idea is hardly new. Many heist movies spend their first half hour assembling people with different talents. AI likewise allows combinations of the expertise a problem needs.

Expanding range still takes creativity. Until he saw colleague Alex Albert render interiors in Blender, he had not thought of using a 3D environment on his own homepage, despite having built video games. Yet he thinks the greatest magic happens precisely where disciplines meet. ✨

Turn Nouns Into Verbs

He introduces an old method from Tom Kelley and Jonathan Littman's The Art of Innovation: turn nouns into verbs when defining a problem.

A QBR is not simply a "deck." It can be reframed as:

"Getting busy people to believe a causal story and make decisions under uncertainty within 30 minutes."

A verb-centered framing makes it easier to identify disciplines with useful, distinctive perspectives. "Making a deck" suggests designers and analysts. Shaping beliefs and decisions also suggests journalism, psychology, narrative design, communication, sound, and video production.

Assemble a Problem-Solving Team

Imagine, he suggests, that aliens will destroy Earth unless the deck is perfect. What team would humanity assemble?

It might include:

• An analyst to handle data
• A journalist to construct the central story
• An editor to sharpen the message
• A motion designer and sound designer if presentation is included
• A red team to attack weaknesses
• A short seller to question the investment thesis
• A forensic accountant to find anomalies in numbers and records

This must not become mere role-play. Asking the model to "write like a short seller" may only make the conversation strange. Instead, make it perform the research, evidence gathering, and challenges a short seller would use to test a thesis.

"Do not ask the model to write in the role of a short seller. Have it do what a short seller actually does to investigate a thesis."

These ambitious approaches work best when models can use a computer, install needed software, and access relevant data. The author recommends an environment connected to data and tools through MCP connectors and similar mechanisms.

5. Why People Struggle to Imagine Bigger Goals

That concludes the practical advice, but the author goes further to explain why it is needed. Humans have always been limited by what we believe is possible, and often trust ourselves only within narrow specialties.

Michelangelo is a classic example. He reportedly did not want to paint the Sistine Chapel ceiling because he regarded himself as a sculptor, not a painter.

Investigating this led the author to the psychological concept of self-efficacy. People with greater self-efficacy believe more in their abilities, and that belief can in turn lead to achievement.

Self-efficacy can be deliberately increased to some degree. Locke and Latham's 35 years of goal-setting research found that people performed better with specific, difficult goals than when told to "do your best." Stripe founder Patrick Collison's "Fast" list likewise records how quickly people have accomplished things when given bold goals and tight deadlines.

He outlines several familiar ways people achieve difficult things.

People Do Hard Things When They Believe They Can

In 1939, George Dantzig arrived late to a statistics class at Berkeley. He copied two problems from the board, assumed they were homework, solved them, and submitted them. They were actually two unsolved problems in statistics.

Not knowing their status allowed him to attempt them without the usual psychological barrier.

People Do Hard Things When They See Someone Else Do Them

People had long tried to break the four-minute mile. Roger Bannister first did so in May 1954, but remained its sole achiever for only 46 days. Once someone demonstrated the possibility, others followed.

The author says something similar is happening with the two-hour marathon. In April 2026, Sebastian Sawe recorded the first official sub-two-hour marathon in London, and the runner-up in the same race also broke the barrier.

People Do Hard Things When They Are Allowed to Fail

The author recently read The Inner Game of Tennis. He describes its central argument as the idea that trying too hard prevents people from playing their best. Allowing room to experiment without excessive fear of failure matters.

Engineers can think of Linux's beginnings. It was initially introduced as:

"Just a hobby, won't be big and professional like GNU."

That "hobby" later became one of the operating systems underpinning the world's computing infrastructure.

6. Closing: Expand the Boundaries of Possibility in the AI Era

Self-efficacy is not a new challenge. But it becomes more powerful when AI capabilities leap forward within weeks. If our abilities expand rapidly while we remain bound by old assumptions and limited imagination, we may fail to use those abilities fully.

The author is not advocating indiscriminately giving AI more work. He recommends demanding higher outcomes, reaching down to rawer data and expert methods, and connecting a wider range of disciplines.

• Raise altitude by delegating ultimate outcomes instead of individual tasks.
• Add depth through expert tools and external verification.
• Expand range by combining methods from multiple specialties.

Doing more ambitious work with AI ultimately requires expanding what we believe is possible, alongside understanding model limitations. He closes by asking readers to share other techniques they discover for enabling greater ambition. 🚀