On the Generative AI in the Real World podcast, independent consultant and longtime software engineer Duncan Davidson joined host Ben Lorica to explore how personal AI could change if people could carry their context between services and decide what each service may access. The conversation touched on developer workflows, open source review challenges, privacy and trust, and a model for device‑based personal memory.
Podcast context
Generative AI in the Real World examines how the momentum from 2023’s ChatGPT wave can be turned into practical solutions by 2026. In the show, Ben Lorica interviews leaders building with AI to surface lessons organizations can apply.
Developers’ roles are shifting toward outcomes (06:31)
Davidson compared AI‑augmented coding to desktop publishing: automation removed much of the physical assembly work but left designers responsible for the final appearance. At Webflow he observed teams using AI to build websites far faster than with earlier tools. For developers, the emphasis has moved from typing speed to questions like “What outcome do you want? What are you trying to accomplish with this code, and how can it be improved?” People must still judge whether an implementation fits the broader system.
How will new developers gain judgment? (10:41)
Computer science fundamentals remain important because when an agent proposes several plausible fixes, a human must choose the one that integrates with the larger system. Ben asked where future engineers will develop that judgment if AI takes on entry‑level tasks. Davidson likened early career work to an apprenticeship in the arts: even menial assignments expose newcomers to how the whole operation functions. He warned that firms focusing only on throughput may leave too little room for on‑the‑job learning.
AI‑assisted changes strain open source review (23:07)
Davidson described using a library in one of his apps and noted that his fixes, combined with performance issues flagged by Claude, left his branch roughly 25 pull requests ahead of the project’s main branch. Maintainers were also using AI, he said, and the challenge was returning useful changes to the upstream project without overwhelming its review process. He concluded that older open source collaboration practices need a “serious retrofit” to handle AI‑assisted contributions.
Personal memory is fragmented across tools (28:27)
Davidson initially had his coding agents store useful information in Markdown files. After commercial agents added memory features, he found that “Claude knew me one way and ChatGPT knows me another way.” He could push context back to each tool using instructions and periodic reminders, but that approach requires more tinkering than most people want. His aim is a small set of shared mechanisms enabling a person to carry context across services.
Portable memory needs scoped boundaries (34:01)
Davidson argued that prompt context, an agent’s short‑term memory of interactions, and a person’s longer‑term knowledge should be portable, but users must control access by scope. He keeps general notes alongside family and financial information and tests systems by asking, “Give Claude access to these scopes of my knowledge, but not everything.” The problem is more complex when personal agents encounter workplace knowledge; he left open where an employee’s expertise ends and an employer’s information begins.
Trust and security will shape what people share (36:44)
Ben said he is not yet comfortable letting a consumer agent access his finances. Davidson observed that trust will need to build through experience, similar to working with a human agent. AI agents already find vulnerabilities faster than established disclosure practices can handle. Davidson expects maintainers to face more pressure as agents surface latent bugs, while Ben questioned how quickly organizations depending on vulnerable code would learn about those findings. Users need control over an agent’s access and confidence in the systems that surround it.
A personal memory that answers requests with permission (49:52)
Davidson envisioned keeping his memory on his phone and letting a new service request specific information: a car dealership’s agent, for example, could ask for details, but the phone would prompt him for approval before sharing. In his words, a service could be given a handle to his memory so its agents could query him: “Hey, we’d like to know these things about you,” and the phone would not necessarily respond automatically.
Conclusions
The episode raised technical and human questions about portable personal memory: how to carry context while preserving user control and security; how future engineers will learn judgment if AI assumes entry‑level tasks; and how open source collaboration must adapt to an influx of AI‑assisted contributions. Davidson’s proposal — local memory plus scope‑based permissioning — aims to balance privacy with useful cross‑service context sharing.
The full conversation and other episodes are available on the O’Reilly Learning platform and through the podcast on YouTube, Spotify, and Apple.
