Thinking Straight

Kevin Lotto is a volunteer with the Torchbearer Community, a regular AI safety speaker with the AI Safety Awareness Project in Wisconsin, and co-founder of the Control AI Wisconsin Action Group. He is also a daily power user of the same AI tools he advocates regulating, which puts him squarely in the tension this show is built around.

We start with what people get wrong about AI company economics, and why the profitability debate may be a sideshow to a bigger race. From there we get into why modern LLMs are grown more than engineered, the Bitter Lesson, and the difference between AGI and ASI. We take on the “just build an off switch” argument, the Moloch trap that pushes businesses to adopt AI faster than they’d like, and the multitasking trap of running too many agents at once. We also cover how business leaders can roll out AI with guardrails, how listeners can engage their lawmakers, the argument that China makes slowing down impossible, and the case for narrow AI as an alternative to the race toward general intelligence.

We close on whether there’s cause for optimism, and on a question the show keeps coming back to: AI isn’t going away, so how do we steer it, and is the real problem less the technology itself than the speed of the race around it?

🔗 Episode Links:
- Kevin Lotto on LinkedIn: https://www.linkedin.com/in/kevinlotto/
- Torchbearer Community: https://www.torchbearer.community/
- ControlAI: https://controlai.com
- The Bitter Lesson (Richard Sutton): http://www.incompleteideas.net/IncIdeas/BitterLesson.html
- Meditations on Moloch (Scott Alexander): https://slatestarcodex.com/2014/07/30/meditations-on-moloch/
- Max Tegmark: https://en.wikipedia.org/wiki/Max_Tegmark
- Instadocs: AI Gone Wild (Netflix): https://www.netflix.com/tudum/articles/instadocs-ai-gone-wild-release-date-news

Mike Maniscalco is the CEO of QumulusAI, an AI infrastructure provider built to deliver high-performance computing for AI natives and inference providers. Mike is a serial entrepreneur with multiple successful exits who previously served as CTO of Applied Digital.

Most of us experience AI as software, but every prompt runs on chips, power, cooling, and buildings that take years to bring online. Mike spends his days on that physical layer, and this conversation provides a fascinating glimpse into the infrastructure that makes it all possible.

We start with what a neocloud is and why the ChatGPT moment opened the door for a new kind of cloud provider to compete with traditional hyperscalers like AWS, Google, and Microsoft. From there, we unpack the challenges around scaling AI infra, how demand for compute keeps outrunning the data center capacity needed to serve it, and how advancements in chips and cooling technology are reshaping the way data centers are built.

That leads to what all of this means for businesses building with AI: why more companies are moving to open source models and inference providers as their AI spend grows, how routing each task to the right model can cut costs without sacrificing quality, and why trust and data isolation matter when deciding where your workloads run.

Finally, we zoom out to get Mike's take on the broad implications of AI on topics like job displacement, healthcare, and the future of work, and why he remains an optimist about where it's all headed.

🔗 Episode Links:
- Mike Maniscalco on LinkedIn: https://www.linkedin.com/in/michaelmaniscalco/
- QumulusAI: https://www.qumulusai.com
- QumulusAI on LinkedIn: https://www.linkedin.com/company/qumulusai
- TOP500 list of the world's fastest supercomputers: https://www.top500.org

🔖 Chapters:
00:00 Intro
03:26 What are Neoclouds?
07:08 Comparing Neo Clouds with Traditional Cloud Providers
09:52 The Intersection of AI and Cryptocurrency Infrastructure
12:24 QumulusAI Overview
15:52 The Mixture of Experts: Enhancing AI Models
22:48 The Supply-Demand Imbalance
26:45 Who are Neocloud Customers?
29:15 Business implications of open source and self-hosting
36:49 Trust and Flexibility in AI Solutions
39:06 Challenges in the Neo Cloud Space
41:54 Innovations in Data Center Cooling Technologies
48:52 The Future Impact of AI on Society
57:45 AI's Impact on Healthcare and Quality of Life
01:02:56 Wrapping Up

Melissa Tondi is an agentic AI and quality engineering leader with deep experience in process engineering across the software development lifecycle. She has deployed AI at scale inside real organizations, and in this conversation we dig into the traps companies fall into and how leaders can use proven change management to avoid them.

We start with why AI is an operating model change, not a tool rollout, and why buying licenses is the easy part. Melissa shares a candid story about a company-wide adoption metric that measured activity instead of value, and how it sparked fear, shadow AI, and eroded trust.

We then get into what worked instead: letting department leaders own their outcome metrics, one-on-one conversations with respected detractors, and an AI ambassador program built on peer credibility.

From there, we make AI governance concrete, covering approved tools, data classification, human review by risk level, and customer disclosure. We close with agentic engineering, including how a risk-tiered approach to testing and autonomy keeps quality built in rather than inspected at the end.

If you're a leader trying to move your organization from tossing out AI licenses and hoping for the best to generating real results, this one is for you.

🔗 Episode Links:
Melissa Tondi on LinkedIn: https://www.linkedin.com/in/melissa-tondi-186551/?isSelfProfile=false

🔖 Chapters:
00:00 - Intro
00:52 - Introduction to AI and Operating Models
03:37 - Common Traps in AI Rollouts
12:43 - Navigating AI Mistrust and Detractors in Organizations
28:54 - Building an Ambassador Program for AI Engagement
36:44 - Establishing Governance in AI Implementation
42:03 - Understanding Agentic Engineering and Its Impact
54:02 - Balancing Quality and Flexibility in Software Development

Rob Carpenter is the managing member of Frost River Capital, where he's actively acquiring and growing a portfolio of residential service businesses. Before that, he founded and led Valyant AI, a conversational AI platform for the drive-thru food industry that was acquired in 2024. He holds five patents in conversational AI and has been featured by Fox Business, The Washington Post, The New York Times, the BBC, and TechCrunch.

We cover: Grok Bot, cybersecurity in the age of AI, robotics, and AI Infrastructure

🔖 Chapters:
00:09 Grok Bot
00:29 Intro
02:07 The Age of the Agent
06:47 Grok Bot
18:09 Context Windows
21:33 Security and Computer Control
27:03 Grok Bot & Image Creation
28:36 Is Grok Bot the Real Deal?
34:58 The Broader Cybersecurity Landscape
40:26 Robotics
44:59 Infrastructure Build-Out
49:31 How Optimistic Are You?
52:00 Getting in Touch

Is AI making everyone in the room dumber, one polished answer at a time?

In this episode of Thinking Straight, we talk with Erika Lenz, founder of KeepYourMind.ai and Kinetic Change, about the human side of AI adoption: the messy, cognitive, and organizational work that most companies skip on their way to "just go use AI."

We explore what "human-centered AI" actually means in practice, why 95% of AI adoptions are failing to deliver expected ROI, and why the real bottleneck in software isn't engineering anymore; it's distribution, customer attention, and the humans still stuck in every workflow.

We dig into the theory of constraints, the danger of frameworks that automate an entire product lifecycle while skipping the ends that matter, and why senior judgment and taste are becoming more valuable, not less, even as AI writes the code.

We also get into AI critical literacy: the discipline of asking not just how to use AI, but whether and when to use it at all. That includes the sycophancy baked into these tools, the practice of demanding the "bull and bear case" before trusting an output, the real research on cognitive de-skilling, and simple habits like journaling decisions, drafting without AI once a month, and capturing a first take before handing it to the model, for keeping your own thinking sharp.