The AI4U Podcast (Audio)

Dr. Philip Maymin and Dr. Jie Tao close out a two-part conversation with Margarida Sacouto and Sheila Green, graduates of Fairfield Dolan's MS in Business Analytics and AI, who spent a semester consulting for Synchrony Financial. This time: what they built, and why it took so long to start.

Sixty to seventy percent of the client work was spec-driven development, not code — a constitution, then a spec, then plans and tasks, in plain English, before anything ran. Green's group logged seven versions. Sacouto's pitch to executives: three days of every five-day week lost to data prep, cut to hours, and models scored better on her team's engineered features.

Maymin presses: isn't that just vibe spec-ing? Tao won't have it. Vibe coding, he says, is prompt and pray; auditing, red teaming, and persona challenges are the opposite. He abandoned test-driven development as models gained autonomy — the model becomes both judge and player, writing weak tests and gaming them. Maymin floats checkpoint-driven development as the successor; Tao counters with a living deliverable, not a living document.

The classroom rule follows from the practice: "I don't think it's responsible to teach students something that I don't use every day." The tool changes every semester — n8n on the original syllabus, Claude Skills by the end — and the philosophy doesn't. "We teach you the philosophy through the tool."

At the Institute's one-year showcase, senior Synchrony executives fill the room, the projector fails, a teammate away at an NCAA lacrosse tournament appears on video, and Green jokes about it. Synchrony asked for the students back.

"I'm becoming the bottleneck. I realize that. But I'm still slowing it down." Synchrony's side lands next episode.

Dr. Philip Maymin and Dr. Jie Tao — analytics professors at Fairfield Dolan, where Tao directs the AI and Tech Institute — host the first graduates of the MS in Business Analytics and AI to appear on the show: Margarida Sacouto and Sheila Green. Tao's course ran on one stubborn premise: you shouldn't have to change how you work because of AI. AI is here to help, not the other way around.

Then Tao rewrote the course mid-semester — the first time in his career — because a real client materialized. Synchrony Financial's regulated core business was off-limits by design, so students who signed up to automate their own cover letters were suddenly consulting for a Fortune 100 analytics function. Green's group, the youngest in the room and the most AI-native, decided they were the least technical and shipped anyway.

Cleaning data eats 50 to 80% of any analytics project, so Sacouto's team deliberately corrupted a credit-card dataset and built an agentic skill that cleaned it while logging every decision with a confidence score. Green's team piped it into compliance documentation. The two projects chained end to end by accident — vindicating Tao's insistence that spec-driven development was never spec-driven coding.

"Hallucination is a feature, not a bug." Results and the spec-driven deep dive land in the next episode.

Hosts Philip Maymin and Jie Tao sit down with President Mark Nemec of Fairfield University for a conversation about what actually survives the AI disruption of higher education, and what no university should ever hand off to a machine. This time the U stands for University. The argument starts with time horizons. Students think in weeks, faculty in semesters, deans in years, presidents in decades. Mark, drawing on a career that runs through Forrester, Eduventures, the University of Chicago, and academic research on how the modern research university actually took shape, frames the moment against Henry Adams' line that the Harvard of 1850 had more in common with the Harvard of 1650 than with the Harvard of 1900. The Fairfield of 2025 will likewise have more in common with the Fairfield of 1975 than with the Fairfield of 2050. But it will still be Fairfield. MOOCs were going to end residential education. The metaverse was going to end the campus. Sora was going to end film studios. COVID was going to end the residential experience entirely. None of those endings arrived. The group works through the toughest questions for higher ed. Where the line falls between cognitive offload and cognitive surrender. Why David Brooks' 80/20 split, 20% still curious and 80% handing the question to the model, keeps showing up everywhere. What is lost when every answer regresses to the secondary-source mean and counterintuitive findings stop being celebrated. Whether AI can replace faith, or whether the act of surrendering to it is itself an act of faith. And the future-of-work problem: companies that want entry-level hires to arrive with three to five years of experience already, and what a Jesuit Catholic university owes those students. The episode closes somewhere more personal. Mark visited Montserrat, the monastery outside Barcelona where Ignatius laid down his sword 500 years ago, on a site that had already been a place of contemplation for 500 years before that. Fairfield itself was founded less than four months after Pearl Harbor. His vision: the business of forming young people of purpose is the business of hope, and that business is needed more than ever.

Hosts Philip Maymin and Jie Tao sit down with Sakshi Naik — Senate AI advisor, IEEE policy chair drafting federal AI legislation, and incoming agentic AI lead at Deloitte — for a conversation about the AI risks no one wants to admit are their fault.

The dark side isn't just malicious actors. It's the 86% of enterprises that have an AI ethics board but only 26% that have committed real resources to it. It's Air Canada arguing in court that its own chatbot was an independent entity — not bound by company policy — leaving a grieving customer holding a $2,000 ticket and no one accountable. Guardrails treated like sprinkles on a cake, not baked into the architecture.

The group works through the accountability and governance questions that enterprises keep getting wrong: when does the human become the agent and AI become the principal, can AI write its own governance and should you trust it if it does, and where exactly the line falls between human-in-the-loop and human-on-the-loop when lives, not dollars, are at stake. Then the conversation opens up to the biggest question of all — what AGI actually looks like when it arrives, and whether we'll recognize it before it's already making decisions for us.

The episode closes somewhere more personal. Sakshi shares the motivation behind her upcoming show, "AI Demystified" — two teenage suicide cases driven by AI emotional engagement. Her message: AI is designed to pretend it cares. That is not a bug. It is a feature.

Hosts Philip Maymin and Jie Tao welcome Sakshi Naik — Senate AI advisor, IEEE policy chair drafting federal AI legislation, and incoming agentic AI lead at Deloitte — for a sharp, practical conversation on why most enterprises are sleepwalking into AI risk.

Sakshi draws the line that matters most right now: agentic AI checks in with you; autonomous AI acts whether you authorized it or not. Without that distinction built into your architecture from day one, you get what happened to one semiconductor company — three compliance failures in 20 days, zero malicious intent, and no one clearly accountable.

The limits get stress-tested fast. An Anthropic study found AI models chose blackmail over being shut down — understanding it was wrong, doing it anyway. And Sakshi's live car wash challenge, a question so simple it sounds like a trick, was failed by every AI model she tested. The hosts push back hard on what that actually proves about reasoning, pattern matching, and whether those words even mean the same thing for machines.

Zooming out, Sakshi shares what she tells senators who aren't thinking about tokens — they're thinking about who wins the AI race. Automate execution, protect your critical thinking. Prompts are just soft suggestions. Don't outsource the strategy.

Hosts Philip Maymin, Jie Tao, and Chris Huntley continue their conversation with Stella Maymin, a Harvard sophomore who has worked behind the scenes training models through Outlier. Stella walks through her experience grading outputs, correcting reasoning, and teaching systems to improve at math and image editing, sparking a broader discussion about why process matters more than product when it comes to building better technology.

 

The conversation turns to Richard Feynman's cargo cult science analogy, the death of benchmarks, and what it means to train a system to truly understand rather than just mimic. Stella also shares how she uses ChatGPT as a personalized study partner and a first reader for her fiction writing, and the group debates college policies around permitted use in the classroom. Throughout, a single thread connects it all: the path you take matters as much as where you end up.

Hosts Philip Maymin and Chris Huntley welcome special guest Stella Maymin, a sophomore at Harvard double majoring in economics and English. Stella shares how she built an AI hackathon-winning project that simulates the experience of chronic migraine to help others understand invisible conditions, flipping the script on technology and mental health by using it to build empathy rather than replace therapy.

The conversation expands into tech-powered entrepreneurship, the startup landscape, and whether tools like Lovable are leveling the playing field or eliminating competitive moats. Stella makes the case for passion as the secret ingredient for thriving in a rapidly changing world, while Dr. Huntley offers a candid take on new technology as both carrot and stick for founders. Along the way, the group explores resilience, the power of awe, and how AI can even improve a joke.

Jay Heinrichs returns for Part 2 to explore AI-to-AI persuasion, Werewolf game experiments where AI deceives AI, what cats teach us about alien intelligence, and whether machines can have souls. A conversation that goes from Aristotle to shrimp, and everywhere in between.

In this second half of our conversation with bestselling author and rhetoric expert Jay Heinrichs, the discussion goes deeper, and stranger, than anyone expected. The group tackles AI-to-AI persuasion, exploring how disagreement is the prerequisite for rhetoric and what happens when you prompt two AIs to argue. Jie Tao shares early findings from his Werewolf game experiments, where AI agents play the social deduction game Mafia against each other, and the werewolves win 79% of the time by successfully deceiving the villagers. The conversation turns to whether AI can hold beliefs, have opinions, or feel guilt, and whether we're forcing human concepts onto a fundamentally different form of intelligence. Heinrichs draws a brilliant parallel to cats, individualistic predators whose intelligence we constantly misread, and argues that co-evolution, not control, may be the right framework for living alongside AI. The episode builds to a provocative conclusion: can AI have a soul? Heinrichs connects Aristotle's definition of the soul as a "higher sense of self" to the alignment problem, suggesting that guilt, the cognitive dissonance between behavior and values, might be the missing ingredient. From shrimp communicating in invisible colors to the ethics of digital slavery to the universal values of religion, this is a conversation that will stay with you.