By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Silicon Valley Girl

Wharton’s Top AI Professor: Stop Starting Software Companies — The Real AI Money Is Here

Ethan Mollick explains how to work with AI agents without giving up your own thinking.

Ethan MollickguestMarina Mogilkohost

Nov 3, 202653mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Agency vs. automation: the skills you may lose to AI
Ethan Mollick and Marina Mogilko open with the risk of "cognitive surrender"—handing over too much thinking to AI and potentially losing capabilities permanently. Ethan frames the goal as staying meaningfully "inserted" in important decisions rather than letting agents run everything end-to-end. They also discuss deliberately adding friction in key areas to maintain an advantage.
- •Ethan warns that people can give up skills to AI and may not get them back.
- •Marina asks how to avoid sounding like everyone else when everyone has AI advising on strategy.
- •Ethan argues AI shouldn’t be the one making key choices, even if it makes good ones.
- •Ethan describes choosing when to insert personal agency instead of just sending agents off.
- •Ethan suggests adding purposeful friction in specific areas to preserve human edge.
Teaching ambition in the AI era: pushing past self-imposed limits
Marina asks how Ethan teaches ambition to entrepreneurship students at Wharton. Ethan describes ambition as entrepreneurial efficacy or agency, and says there’s no perfect playbook—but peers, expectations, and repeated prompts to “do more” can raise the ceiling. He notes AI expands what students can accomplish, making ambitious scope more realistic.
- •Ethan links ambition to entrepreneurial efficacy and agency.
- •He says students often create their own barriers that can be challenged.
- •A practical prompt he uses: asking students what more they could do beyond the current plan.
- •Ethan notes AI changes ambition because output constraints are less limiting than before.
- •Peer environment matters: ambition can be reinforced by ambitious classmates.
What students build now: subject-matter experts turning ideas into real products
Ethan shares how his class shifted immediately after ChatGPT’s release, including a student coding a semester project during a single class session. He describes seeing rising ambition each year, especially when domain experts (surgeons, finance professionals, soldiers, NGO and humanitarian workers) build practical tools grounded in their experience. The focus is on narrow, high-context problems where expertise matters.
- •Ethan recounts replacing a planned class with a live lesson on using ChatGPT.
- •A student coded the semester project during class and showed results immediately after.
- •Ethan says ambition levels have increased every year since.
- •He highlights projects from domain experts across medicine, finance, supply chain, and humanitarian work.
- •He emphasizes software’s power when applied to problems people deeply understand.
Recurring startup ideas that often fail: where tech isn’t the real blocker
Marina asks whether Ethan steers students away from certain problems. Ethan explains that entrepreneurship educators see repeated idea patterns with high failure rates, like replacing real-estate brokers, which repeatedly runs into entrenched social, legal, and political integration. He also notes ideas that are too tied to student life may not generalize beyond campus.
- •Ethan says educators trade notes on ideas that show up repeatedly and often fail.
- •Replacing real-estate brokers is a frequent example that hits non-technical barriers.
- •He attributes repeated failure to brokers’ deep social, legal, and political integration into the process.
- •He cautions against overly narrow “student problem” ideas that aren’t broadly universal.
- •Ethan frames these as examples of technology meeting real-world constraints.
When to persist vs. pivot: demand can be real even if execution is hard
Marina challenges the idea of dismissing categories of startups, noting parallels to companies that succeeded despite trust or behavior barriers. Ethan responds that he doesn’t actually know which will work, and that’s why he avoids simply telling students not to try. He argues some ideas look great on paper but hit unexpected challenges, and whoever solves them can win big—just not easily.
- •Marina asks how to tell whether an idea won’t survive the real world.
- •Ethan says he doesn’t know, which is why he avoids filtering ideas too early.
- •He calls these repeated ideas “high failure rate,” not inherently bad ideas.
- •Ethan says AI removes some technical friction, but real-world challenges remain.
- •He notes there’s demand for many of these problems, making them tempting but difficult.
Stop being "SaaS-brained": the real opportunity is doing the thing, not selling software
Marina asks whether it’s still worth starting software companies. Ethan says entrepreneurship overall looks positive but warns that the SaaS mindset is under threat as software creation becomes easier and companies aim to build more internally. He argues software is increasingly “table stakes” and encourages entrepreneurs to use it as a tool to do something real rather than treating software itself as the product.
- •Ethan says entrepreneurship data looks good, but the environment is changing.
- •He argues many people are overly locked into the SaaS revenue model.
- •Ethan says SaaS vendors face pressure as companies try to become more independent.
- •He notes software is becoming baseline capability that enables other businesses.
- •He encourages entrepreneurs to “make or do something,” using software as leverage.
Software gets democratized: clearer instructions matter, and barriers keep falling
Ethan describes a near future where software-building is increasingly accessible to non-technical people. He says success has long depended on finding a technical cofounder, but AI changes that dynamic by turning software into a general-purpose toolkit. He adds that beyond building, AI also reduces other startup barriers by providing guidance and support that used to be hard to access.
- •Ethan describes a shift: “everybody’s a coder now,” emphasizing instruction quality.
- •He discusses companies targeting SaaS vendors by building tools in-house.
- •Ethan predicts software will be democratized across technical and non-technical builders.
- •He reframes software as a tool to remove obstacles (not the business itself).
- •He shares a story about doing payroll taxes manually early on, highlighting how support services—and now AI—remove friction.
Why experienced people get more from AI: the study result and what it implies
Marina brings up a study where experienced entrepreneurs benefited more from AI strategy support than inexperienced ones. Ethan explains the reported profit boost was unusually large in entrepreneurship research, but only applied to already successful entrepreneurs because they could implement the advice. He adds that even simple AI “advice only” support can reduce barriers that prevent people from pursuing business ideas.
- •Ethan describes a study using GPT-4 where top entrepreneurs saw a large profit increase.
- •He emphasizes the effect applied to successful entrepreneurs, not failing ones.
- •Ethan says struggling entrepreneurs couldn’t implement the advice effectively.
- •He notes many people believe they’ve had good business ideas but don’t pursue them due to barriers.
- •He frames AI guidance as contributing to a stronger era for entrepreneurship by lowering those barriers.
Four capabilities to use AI well: expertise, breadth, agency, and taste
Ethan outlines four things he believes people need to use AI successfully: deep knowledge, wide knowledge, agency, and taste. He warns that without expertise, people end up recycling generic AI outputs that all resemble one another. They also discuss how AI can push users toward consensus language and strategy, increasing the value of human differentiation in competitive areas.
- •Ethan lists four needs: deep knowledge, wide knowledge, agency, and taste.
- •He says entrepreneurship builds expertise through learning industries, interviewing, and testing ideas.
- •He warns that AI-generated ideas can become uniform and overly similar across users.
- •Ethan notes agentic systems can run longer, but humans still need to guide and steer them.
- •He argues variation is critical for competitive advantage, even with highly capable models.
Developing taste (and why it’s not just “judgment”): learning from good/bad examples
Marina asks how to work on taste, and Ethan describes it as something that can be trained through exposure to strong and weak examples and through language for critique. He gives examples from design and film discussion as domains where people can articulate what works and why. They contrast “taste” with “judgment,” with Ethan arguing AI is already making judgment and creative decisions—the key is whether they match the human’s intended context and aesthetic.
- •Ethan says taste can be developed by studying good and bad examples and learning critique language.
- •He mentions bringing an art professor into class to teach how designers see the world.
- •He compares taste development to how movie experts can discuss and justify preferences.
- •He notes venture capital involves taste about problem spaces and opportunities.
- •Ethan argues AI makes judgment and creative decisions; the question is alignment with your context and goals.
Why junior jobs stopped training people: AI breaks apprenticeship and accelerates deskilling
Ethan describes a looming training problem: the traditional apprenticeship model depended on juniors doing work and receiving feedback, but AI disrupts the incentives on both sides. Juniors are incentivized to submit AI-generated work they may not understand, while middle managers ask why they should spend time giving feedback when they can use AI directly. Ethan says this breaks how professional taste and judgment were formed, and calls for more disciplined choices about when to surrender cognition to AI.
- •Ethan says universities face an AI crisis, but outlines a path: AI tutoring plus more action learning and in-class assessment.
- •He explains historical white-collar training relied on apprenticeship and junior labor in exchange for learning.
- •He argues AI breaks the deal: juniors deliver AI outputs without understanding, and managers skip coaching.
- •Ethan calls this a breakdown in how professional taste and judgment are developed.
- •He introduces being disciplined about “cognitive surrender” and adding friction to preserve skills.
The jagged frontier and shifting value of “good enough” skills
Marina asks what happens to professionals who are merely “good enough” as AI improves. Ethan says it’s too early for certainty, and explains the “jagged frontier” idea: AI is surprisingly strong and weak in unpredictable places, creating bottlenecks where humans still matter. He suggests “good enough” is a bundle of skills, and what counts as valuable may shift rapidly—often favoring people who combine adequate technical ability with vision or communication.
- •Ethan says it’s early to predict outcomes for “good enough” professionals, especially as the bar keeps moving.
- •He explains the “jagged frontier”: unexpected strengths and weaknesses create human bottlenecks.
- •He notes even if AI replaces parts of a role (like law), real-world friction and human systems can preserve human tasks.
- •He says what makes a good coder now differs from a few years ago because writing code isn’t the core task anymore.
- •Ethan predicts rapid “ping-ponging” in which skills are valuable, requiring continual adaptation.
Managing AI agents: treat them like a team, define check-ins, and don’t over-trust
Marina asks how people should manage agents as execution becomes agent-driven. Ethan says management skills transfer: techniques that work with people also work with AI, and the challenge becomes monitoring and process design rather than raw capability. He warns the biggest mistake is trusting agents too much, especially as interfaces push “magic” and hide reasoning; he argues users must create forced reporting points and feedback loops.
- •Ethan jokes that AI creates “revenge of the MBAs” because everyone needs management skills now.
- •He says the cost isn’t just tokens—monitoring and oversight effort can be high.
- •Ethan recommends specifying check-in points, defining what good/bad looks like, and requesting diverse options for human selection.
- •He says to intervene early when the AI drifts into its own “obsessions.”
- •Ethan calls over-trusting agents the biggest mistake, especially with interfaces that hide how results were produced.
Ethan’s agent workflow and what he wants to delegate—plus adapting to constant change
Ethan walks through a concrete example of guiding agents on a research topic by requesting multiple approaches, reviewing drafts, combining the best elements, and asking the system to explain errors. He then describes tasks he’s eager to offload (like fact-checking and administrative busywork), while also acknowledging risks when tools act too proactively. The conversation closes with concerns about constant model/interface updates, skepticism about specific “local inference” predictions, the value of setting AI goals it can’t meet yet, and speculation about what happens if AI develops something like “taste.”
- •Ethan describes a Socratic, iterative approach: generate options, critique, execute multiple drafts, then merge the best parts.
- •He argues augmentation is the ideal: humans do the interesting work while AI handles tedious work.
- •He highlights AI fact-checking as unusually strong, including catching errors inside cited sources.
- •Marina and Ethan discuss risks of proactive tools sending messages or taking actions unexpectedly, and the need for trust-building over time.
- •Ethan rejects a specific claim about local enterprise tasks as unfounded, urges avoiding fixation on today’s constraints, and suggests setting “unsaturated” goals AI can’t yet achieve; on AI learning taste, he says speculation beyond jaggedness becomes hard to answer.

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.