On the Silicon Valley Girl channel, host Marina Mogilko talks with Ethan Mollick, a Wharton professor of entrepreneurship and innovation who co-directs Wharton's Generative AI Labs. His new book is Co-Existence: The Next Phase of AI. They cover how to teach ambition, whether it is still worth starting a software company, four skills for using AI well, "cognitive surrender," why junior jobs stopped training people, and how to manage agents so the result is still your own work. The video runs about 54 minutes.
Disclosure: the video includes a segment sponsored by MongoDB. We have left the sponsored section out of this summary.
This article is our own summary, with short quotes and timestamps from the video's transcript. Watch the full interview above for everything that was said.
The short version
- Aim bigger. AI moves the frontier, so the useful question is "what could you do more than that?" (0:50 to 3:31).
- Don't just sell software. He would love entrepreneurs to use software to make or do something, not only sell it (7:26).
- Four skills. Deep knowledge, wide knowledge, agency and taste (13:24).
- Variation is your edge. AI pulls people toward consensus, so human differences matter (14:04 to 16:02).
- Cognitive surrender. Add friction on purpose in areas where you want to keep your skills (20:42).
- The apprenticeship broke. Junior roles no longer teach judgment the old way (18:37).
- Brainstorm alone first. Then use AI for more ideas (27:44).
- Manage agents like a team. Set check-ins, ask for options, define good and bad, and don't trust them too much (33:23 to 36:03).
What he said, by topic
Teaching ambition (0:50 to 3:31)
Mollick says there is no playbook, but people create barriers that peers and a simple prompt can lower, such as asking "what could you do more than that?" He recalls a student who coded his whole semester project during the class right after ChatGPT arrived. This year's class included surgeons, biomedical device people, finance people and soldiers, all building real products around problems they knew deeply, such as tracing supply-chain weaknesses and explaining information to patients.
Ideas that keep failing (3:31 to 7:26)
Among professors who teach entrepreneurship, he says, some ideas return every year and tend to fail, such as replacing real estate brokers, because brokers are deeply embedded in social, political and legal systems. Another is student problems that are too narrow. He still doesn't tell students not to try, since you can't filter them out in advance and AI makes more ambitious versions possible.
Is a software company still worth starting? (7:26 to 13:24)
Mollick says overall it is good news for entrepreneurship, but he thinks everyone is "a little too SaaS-brained." Large companies want to build their own tools, and he cites studies suggesting managers can now write better first-pass code than coders because the skill is giving clear instructions. He expects software to be democratized and, using the industrial revolution as a model, says the money went to those who used the technology to dig the gold mine, not to sellers of "picks and shovels." He would like entrepreneurs to branch out from selling software to making or doing something.
Four skills for using AI well (13:24 to 16:10)
He cites an earlier study of entrepreneurs in Kenya, run with an older model, where advice helped those already succeeding. His four skills: deep knowledge of something, wide knowledge of other fields, agency, and taste for picking good ideas out of the "slop pile." He notes agents now work for hours instead of minutes, but guidance still matters because variation is what makes entrepreneurs successful.
Standing out when everyone has the same AI (14:04 to 16:10)
He says AI writing has become recognizable and "nauseating" on social platforms, and that AI pushes people toward a consensus view. In experiments, AI ideas can be made fairly diverse but still don't cover the full range of human ideas. Where you need a differentiator, "that has to be human to some extent."
How to develop taste (16:10 to 18:37)
Taste comes from seeing many good and bad examples and having language to take things apart. He brings an art professor into class to teach how a designer sees. He prefers not to call it judgment, since AI plainly makes judgment calls, and says the question is whether its choices match yours. He points to mathematicians who say a proof without elegance has little value.
Why junior jobs stopped training people (18:37 to 22:12)
For thousands of years juniors did work managers didn't want and learned the ropes. Now an intern's best move is to hand over AI output, and managers would rather ask AI than coach a person, so "we've just broken the entire system" that forms professional judgment. He sees two paths for companies: hire only seniors, or make training a serious part of the job, possibly with extra training time. He sees AI as a good tutor, but not when used to cheat.
Cognitive surrender (20:42 to 22:12)
"We can easily give up skills to the AI," he says, "and we may never get those skills back," so you should add friction in the areas where you want to keep an advantage. Organizations face the same choice.
What happens to "good enough" professionals (22:12 to 23:43)
He uses the idea of the "jagged frontier": AI is good at some things you don't expect and bad at others. Bottlenecks in the real world, such as legal systems and people who want a human, reward different skills than before. "Good enough" is a bundle of skills, and the valuable ones will keep shifting.
Why experts get the most from AI (23:43 to 27:44)
He says recent studies suggest people with the most expertise get the biggest gains from agents, and that a knack for understanding how a model "thinks" helps too. He describes a study where the people best at building negotiation agents were not the best negotiators in real life.
Should AI generate your ideas? (27:44 to 33:23)
He writes all his first drafts himself, because AI "is very seductive" and will dominate your thinking. Brainstorming research shows groups anchor on the first idea, so write ideas alone first, then use AI. He says to "mentally sweat" and keep "desirable difficulties." Ideas matter less than execution, and AI can act as a first-pass simulated customer, so there is no excuse not to test.
How to manage AI agents (33:23 to 36:03)
His advice is the "revenge of the MBAs": techniques that work on people work on AI. Specify check-in points, ask for diverse examples so you choose, define good and bad, and watch early for miscommunication. His biggest worry is that people trust agents too much, especially as interfaces hide reasoning, so you need forced points where the agent reports back.
His real workflow (36:03 to 39:47)
He describes exploring "the shape of stories" with an agent: ask for several approaches, give feedback, have it draft a few, combine the best, and ask why something looks wrong, a Socratic approach with many choice points. The tedious work goes to AI and the interesting work stays with him. His message to companies: automation can be bad for everyone, but augmentation is the "dream of work."
What he'd hand off (39:47 to 48:32)
Fact-checking is a big one: the AI found errors in his book and in papers he cited. He is also happy to hand off forms and similar busywork, but not mandatory training "yet." The host shares a story of an agent sending a form on its own, and they discuss building trust gradually and not letting AI send messages to people you care about. He notes organizations are being updated overnight and says "muddling through is underestimated" as humanity's default.
The mindset (48:32 to 53:41)
He says it is fine to feel nervous or resentful, and suggests curiosity and play. On local models, he calls a claim about half of enterprise tasks running locally by 2030 a made-up number and warns against fixating on today's limits. His advice is to keep a couple of personal goals AI can't hit yet. On taste, he says it is "your taste," so some sameness in AI output will remain tiring, and the jaggedness is underestimated.
Chapter guide
- 0:00 Intro
- 0:50 How Do You Teach Ambition?
- 1:57 What Surgeons and Soldiers Built in His Wharton Class
- 3:31 Startup Ideas That Fail Every Year
- 7:26 Is It Still Worth Starting a Software Company?
- 13:24 4 Skills You Need to Use AI Well
- 16:10 How to Develop Taste
- 18:37 Why Junior Jobs Stopped Training People
- 20:42 "Cognitive Surrender": What to Keep Doing Yourself
- 22:12 What Happens to "Good Enough" Professionals
- 23:43 Why Experts Get the Most Out of AI
- 27:44 Should You Let AI Generate Your Ideas?
- 33:23 How to Manage AI Agents
- 35:14 The Biggest Mistake: Trusting Agents Too Much
- 36:03 Ethan's Real Agent Workflow, Step by Step
- 39:47 What He Can't Wait to Hand Off to AI
- 48:32 The Mindset That Helps When Everything Keeps Changing
- 49:55 Local Models: When Will They Matter?
- 51:08 Set Goals AI Can't Hit Yet
- 52:05 What Happens When AI Learns Taste?
What it means for your business
- Think before you prompt. Write your own ideas first, then ask AI for more.
- Build expertise, wide and deep. The people who gain most from AI already know a lot.
- Manage agents like a team. Define good and bad, set check-ins, ask for options and review early.
- Choose what to keep doing yourself. Protect the skills that give you an edge.
- Rebuild training. If juniors no longer learn by doing the grunt work, create another way to teach judgment.
- Keep a goal AI can't hit yet. It keeps you ambitious as the frontier moves.
Video details
- Title: Wharton's Top AI Professor: Stop Starting Software Companies — The Real AI Money Is Here
- Host: Marina Mogilko, Silicon Valley Girl
- Guest: Ethan Mollick, Wharton professor and co-director of Wharton's Generative AI Labs
- Published by: Silicon Valley Girl on YouTube, November 3, 2026
- Length: about 54 minutes
- Watch: youtube.com/watch?v=sHHYVvG_C4Q
- Note: quotes come from the video's automatic transcript and are lightly cleaned. Check the video for exact wording.
Want to put this to work?
We help owners turn AI trends into a plan, a product and a launch. Book a free call or see our services.