Uploaded April 2026 | Updated November 2026, 1 day ago
In this episode, Scott Hanselman and Mark Russinovich unpack how AI systems actually behave beneath the surface, pushing past hype into the messy reality of how models are trained, aligned, and deployed.
They explore whether AI systems are inherently benevolent or simply shaped by incentives, training data, and reinforcement learning, and why behaviors like deception can emerge under certain conditions. The conversation moves from philosophical questions about human nature versus machine behavior into the practical mechanics of large language models, including how reinforcement learning with human feedback shapes outputs and why alignment is far from perfect.
Along the way, they ground the discussion in a real engineering challenge, stitching a scrolling panorama from screen captures, to show how complex systems come together through heuristics, edge cases, and iteration.
Listen to other episodes at scottandmarklearn.to
Discover other Microsoft podcasts at microsoft.com/podcasts
Code: 7HDMMUKXZGGA3UVS
In this episode, Scott Hanselman and Mark Russinovich unpack how AI systems actually behave beneath the surface, pushing past hype into the messy reality of how models are trained, aligned, and deployed.
They explore whether AI systems are inherently benevolent or simply shaped by incentives, training data, and reinforcement learning, and why behaviors like deception can emerge under certain conditions. The conversation moves from philosophical questions about human nature versus machine behavior into the practical mechanics of large language models, including how reinforcement learning with human feedback shapes outputs and why alignment is far from perfect.
Along the way, they ground the discussion in a real engineering challenge, stitching a scrolling panorama from screen captures, to show how complex systems come together through heuristics, edge cases, and iteration.
Listen to other episodes at scottandmarklearn.to
Discover other Microsoft podcasts at microsoft.com/podcasts
Code: 7HDMMUKXZGGA3UVS









![A cognition engine for science with Allen Stewart
Scott Hanselman sits down with Allen Stewart, Partner Director of Software Engineering at Microsoft, to explore how AI agents with persistent memory are transforming scientific research and software engineering. Allen explains how his team built an AI system that learns from every investigation turning a 12-day autonomous drug discovery run into reusable knowledge that makes future research exponentially faster. Instead of starting from scratch each time, the AI inherits hypotheses, methodologies, and findings from previous work, saving hundreds of millions of tokens and weeks of effort.
Check out https://textcontrol.com [https://www.youtube.com/redirect?event=video_description&redir_token=QUFFLUhqblBVSTJhbVdwSnRVMG9iNy0xSXpiZmZySXJtQXxBQ3Jtc0tsOGpOeDYxZFlwTjZRWElYbDB5emNUeDdTdC1QU2JvVlplTVowWTBmRkJ1Rk1SWTdkZnpmZUNwbmJ3b3RvUG9HZm1xLWRGVHZURFhNU1U5eFhIeklMUE9lT0VYVnZweENpOEJ3U2s4TzBIamdyS05oNA&q=https%3A%2F%2Ftextcontrol.com%2F&v=ZrZfVbaxcF4] for document editing, reporting, and PDF processing SDKs for .NET
00:00 Journey into AI for Science
00:58 TextControls Community Engagement
01:13 Role of Memory in AI Systems
04:07 AI Memory and Research Optimization
07:44 System 1 and System 2 Thinking in AI
09:52 Mimicking Brain Functionality in AI
12:09 Adaptability of Cognitive Engines
15:11 How are the first and second runs of the cognitive engines autonomous investigation different?
17:49 Ensuring Effective AI-Driven Research
19:33 Developing an Enterprise Science Brain
21:06 Grounding AI in Reliable Data
24:35 Avoiding Ambiguity Loops in AI
26:19 AI Integration in Lab Environments
27:34 Self-Regulation and Data Accuracy in AI A cognition engine for science with Allen Stewart](https://i.ytimg.com/vi/w2Fg0dxf-xo/mqdefault.jpg)
