The Shift in Teaching AI

Ideas Are Cheap and Execution is Everything? Think Twice

After over a decade of teaching AI at Carnegie Mellon University’s Integrated Innovation Institute, this year felt different. My students built production-grade AI agents in two weeks. Some built VC-ready AI products in under 10 hours. That kind of speed forced me to rethink one of the oldest beliefs in tech. We used to say that ideas are cheap and execution is everything, but AI is changing that equation. A growing share of execution is becoming faster, cheaper, and easier to access. The hard part is moving upstream: deciding what is worth building in the first place.

Organizations can build faster, but must choose what’s worth building.

AI has not made product building easy, but it has made the first draft dramatically cheaper, collapsing the distance between idea and prototype. Code, UI, research, testing, and workflow automation can all move faster now. That does not remove the hard work of reliability, security, user trust, evaluation, or distribution; instead, it changes where the bottleneck sits. In many cases, the bottleneck is no longer “Can we build this?” It is “Should we build this, and what would make it actually useful?”

The broader data points in the same direction. GitHub reported that developers pushed nearly 1 billion commits in 2025, up 25.1% year over year, and merged 43.2 million pull requests per month, up 23% (GitHub, 2026). Stack Overflow’s 2025 survey found that 84% of respondents are using or planning to use AI tools in development, and 51% of professional developers use them daily (Stack Overflow, 2025a, 2025b). Google Cloud’s 2025 DORA research found that 90% of technology professionals use AI at work and more than 80% say it improves productivity (Harvey & DeBellis, 2025). But DORA also found the catch: 30% report little or no trust in AI-generated code, and Google’s March 2026 follow-up notes that time saved in creation is often re-spent on auditing and verification (Baolin & Harvey, 2026). The first version is cheaper. The last mile is still hard.

Building a prototype is cheap, but establishing trust is hard.

That distinction matters. AI has commoditized a meaningful slice of implementation, not outcomes. A demo or prototype is now easier to build, but a product that people trust is not. McKinsey’s November 2025 global AI survey found that nearly two-thirds of respondents say their organizations have not yet begun scaling AI across the enterprise, even though AI use is now widespread. One of its clearest findings is that redesigning workflows matters more than simply adding tools. In other words, AI creates leverage, but only when people understand the workflow, the constraints, and the standard for success.

That is why I changed how I teach. I moved the Applied AI course toward case studies, hands-on building, and real-world tradeoffs. Less lecture. More judgment. Less time on material AI can explain on demand. 

More time on questions AI cannot answer well on its own: What is the real bottleneck? Where does accuracy matter most? What still needs human review? What separates a real product from a polished demo? In some ways, this is an extension of how I have always approached Applied AI at CMU — connecting technical depth with practical use cases. AI just made that need much more urgent.

Teach for decisions, not just delivery.

The classroom is changing because the students already have. HEPI’s 2025 survey found that 92% of undergraduates use AI in some form and 88% use generative AI for assessments, yet only 36% say they have received AI-skills training from their institution. 

UNESCO’s 2025 higher-education survey, based on 400 responses from 90 countries, found that 9 in 10 respondents use AI tools in their professional work. Adoption is no longer the main question. The real gap is fluency: using AI with judgment, context, ethics, and discipline.

This also matches UNESCO’s broader direction for education. Its AI competency framework for teachers emphasizes a human-centered mindset, ethics, AI pedagogy, and professional learning. That is exactly where I think teaching has to move. When execution gets cheaper, education should spend more time on what is harder to automate: framing, critique, judgment, communication, and responsible decision-making.

What the guest speakers reinforced

This year’s guest speakers made that shift real. The class was shaped by insights from James Xian, Liz Wang, Mark Hull, Peizhao Zhang, Ryan Hanrui Wang, and Ted Nyman. 

They came from different corners of AI — trading systems, life sciences, engineering intelligence, frontier research, infrastructure, and DevOps — but they pointed to the same conclusion: the bottleneck has moved.

James Xian of Metabit wrote that his goal is to give “every engineer a professional AI partner,” with AI spanning requirements, code, tests, and pull requests. That is a good description of what students now experience: AI is no longer a helper for one task. It is moving across the full development loop. 

Liz Wang of Deffai named a different kind of bottleneck with one sharp line: “Science isn’t the bottleneck. The regulatory maze is.” That is a lesson I want students to internalize. Better execution matters little if you are solving the wrong constraint.

Mark Hull of Exceeds AI put the human side well when he wrote that “coding tells a story,” because it reveals how people think and architect solutions. In another public post, he described building three AI-assisted products in ten days for about $1,270 in AI costs, versus what he estimated as a traditional engineering equivalent of $36,000 to $60,000. Even as a single founder’s anecdote, it captures something important: getting to version one is getting radically cheaper. Ryan Wang of Eigen AI framed the next phase of AI as a race for “speed, efficiency, reliability, and real-world performance,” not just bigger models. That is exactly the shift students need to understand.

Ted Nyman of Cased offered the right counterweight to the hype. In serious incidents, he wrote, “smart, experienced humans” are still “golden.” That line stays with me because it captures the future I see most clearly. AI will take more of the repetitive work. Human value rises in the novel, ambiguous and high-stakes moments. 

Peizhao Zhang, formerly a Senior Staff Research Scientist at Meta, added another important lens: frontier AI progress still depends on deep research into efficient models and systems, not just product-layer speed.

What students need now

The students who stand out now are not simply the fastest builders, they are the best choosers. They can identify the real problem, see the hidden bottleneck, test assumptions quickly, and decide where AI should help and where it should not. They understand that a prototype can be cheap, but customer trust is earned. They know that shipping more output is not the same as creating more value.

I still believe execution matters. But I define it differently now. In the AI era, execution is no longer just producing every artifact by hand. It is directing human and machine intelligence toward the right problem, under real constraints, with a high enough standard of trust for the real world.

Execution is getting cheaper.
Judgment is getting more valuable.
That is why my teaching changed.