Unsecurity Episode 242: AI Evolution, Application, & The Future w/Jim Wilt

# Unsecurity Episode 242: AI Evolution, Application & The Future w/Jim Wilt

## Overview Hosts Megan Lkins and Brad Nye interview Jim Wilt, a Minnesota-based AI educator and consultant, to explore the evolution, practical applications, and future of artificial intelligence. Wilt brings three decades of experience spanning aerospace, medical diagnostics, machine learning, and generative AI, emphasizing that AI should be approached as an augmentation tool rather than a job replacement technology.

## Speaker Background Jim Wilt is introduced as a key figure in the Twin Cities AI conversation, actively involved with Applied AI—a bi-weekly meetup of local AI practitioners and technologists. He positions himself within a collaborative brain trust rather than claiming expertise above peers.

## Key Topics and Technical Evolution

### AI History (1990s–Present) Wilt traces his AI journey starting in the mid-1990s in aerospace manufacturing, where neural networks controlled laser heating of composite materials for stealth aircraft (B-2 bomber, F-18, F-35, 787). These networks learned in real-time, becoming smarter with use despite requiring complex mathematics.

By 2005, he worked on NexOpinion—an early "cognitive" AI (not machine learning) diagnostic software using Bayesian logic that achieved 80% confidence in narrowing 20,000 possible medical conditions to 20 likely candidates. Though designed never to diagnose patients independently, it found wide adoption in teaching hospitals and internationally.

Machine learning and predictive analytics emerged over the following 15–20 years, using massive datasets to identify patterns and make consistent decisions—though biased decisions persist if training data is flawed (e.g., Amazon's hiring AI that favored white males from certain universities).

Generative AI and large language models represent a recent breakthrough, diverging dramatically from the incremental improvements of prior decades.

### Prompting and Model Interaction A critical technical insight concerns prompting. Wilt recounts a kidney stone video analysis case in which a researcher spent a month trying to prompt an AI to identify kidney stones with only 50% accuracy. After stepping back and simply asking "What do you see?"—receiving the response "peanuts"—he reformatted the prompt to "Find the peanuts in this video," achieving 98% accuracy. This demonstrates that how queries are framed determines AI output quality far more than retraining models.

Wilt recommends interactive prompting for personal exploration: asking AI to analyze your LinkedIn profile, recommend ice cream brands, or plan travel itineraries through back-and-forth dialogue rather than one-shot commands.

### Bias and Continuous Improvement Bias in AI models reflects biased training data and historical patterns. While models include bias checks (some toggleable via APIs), Wilt argues humans often exhibit stronger bias than modern AI systems—citing ChatGPT's balanced discussion of Star Wars films versus polarized human opinions. Algorithms exist to detect and mitigate bias, and vendors are making conscious efforts to reduce it, though complete elimination is impossible given biased data sources. Notably, within two weeks of Wilt's London/Dublin conference trip, ChatGPT improved enough to solve a previously failing synopsis-writing task—illustrating continuous monthly improvements in capability.

## Enterprise Adoption Strategy

### The Three Zones Wilt identifies a phased adoption model: 1. **Learning Zone**: Play with public data on free or API-based models without business risk. 2. **Enterprise Zone**: When deploying internal organizational data, implement security, regulatory compliance, and governance frameworks. 3. **Public Announcement Zone**: Only announce AI usage publicly after internal mastery.

### Common Pitfalls Hype-driven adoption often fails. When a board pressures a CIO to "bolt on AI" without prior exploration, teams may pursue unrealistic projects (e.g., medical diagnosis for a non-medical company). Failing to achieve promised results, organizations lose funding and develop AI skepticism. Conversely, companies like Thomson Reuters and Capital One quietly implement AI internally, improving products without premature announcements.

### Governance and Education Bala Krishna's HIKUA organization in Ireland plans to educate constituents on AI before enforcing EU AI Act compliance, recognizing that informed users adopt AI more safely and effectively. Wilt advises leaders demanding immediate AI deployment to commit instead to a three-week exploration period with public data, then present four validated use cases—often originating from non-IT departments (accounting, customer support, care services).

## Real-World Applications

**Manufacturing**: Companies Wilt declines to name by industry (but "you would never think would need AI") are achieving groundbreaking results in material selection and manufacturing process optimization.

**Medical Diagnosis**: AI should listen to patients (sentiment analysis, frustration identification) and recommend local experts by analyzing publications—augmenting rather than replacing human expertise.

**Travel Planning**: Wilt used LLMs to design a Europe itinerary during anniversary travel, with AI asking seven clarifying questions about budget, dining preferences, and special occasions, producing an accurate, detailed plan.

## Key Principles and Warnings

1. **Augmentation, Not Replacement**: Properly deployed AI makes humans more effective. Companies attempting to eliminate roles via AI typically rehire those positions within 2–3 years. 2. **Play First**: Users unfamiliar with AI's capabilities and limitations shouldn't deploy it in business. Hands-on exploration with public data is essential. 3. **Beware Hype Layers**: Over-engineered solutions (MCPs, custom wrappers, consulting-heavy frameworks) dilute value. Raw, simple use is often preferable. 4. **Know Your Prompts**: Understanding how to ask questions determines success—poor prompting yields poor results regardless of model sophistication. 5. **Educate Before Regulating**: Governance succeeds when users understand AI's mechanics and responsible application. 6. **Continuous Learning**: Monthly improvements mean strategies must evolve. Stale knowledge becomes liability.

## Broader Context Wilt frames public reception on an emotional spectrum: fear and rejection at one extreme (relatives avoiding AI discussions), unbridled enthusiasm at the other (attendees at Applied AI seeking rapid adoption without grounding). Both extremes risk poor outcomes. The optimal path combines measured exploration, realistic expectation-setting, and continuous education.

The conversation underscores that AI's revolutionary character stems not from replacing human judgment but amplifying it—provided organizations approach adoption with curiosity, rigor, and humility about what the technology can and cannot accomplish.

## Acknowledgments Wilt acknowledges local leaders Justin Graham, Steve Peterson, Stephen Kaufman, and Michael Olufo for advancing Twin Cities AI education. He notes an upcoming healthcare AI discussion with Olufo at University of St. Thomas (July 30th). Hosts thank their incident response team and welcome feedback and guest suggestions via unsecurityforcure.com.