Stewart Squared
Episode #40: Probabilistic Machines: Living with the Illusion of Control
- AI & Agents
- vibe coding
- Claude code
- Descript
- Riverside
- Duolingo
- LLMs
- Google Live Transcribe
- Google Translate
Listen
Download MP3About this episode
Welcome to Stewart Squared podcast with the two Stewart Alsops, where this episode takes you on a ride through vibe coding experiments, AI-powered doom loops, and the fading utility of language learning apps like Duolingo in a world of real-time translation glasses. Stewart Alsop shares how he replaced Descript with Claude-generated code, while Stewart II unpacks the uncanny valley between needing to understand code and getting the machine to do it for you. They riff on the evolution of Apple’s infrastructure, Unix origins, the role of kernels, and Microsoft’s unlikely embrace of open source. There’s also a tribute to Cursor, the AI-infused IDE built on VS Code, and talk of enterprise LLMs like McKinsey’s internal model. Expect a whirlwind of anecdotes from student visa bureaucracy in Buenos Aires to early software packaging in Ziploc bags.
Timestamps
01:28 Stewart II discusses Google's real-time transcription and translation technologies and their potential impact on language learning.
04:24 Stewart Alsop explains his probabilistic "vibe coding" workflow using Claude and Gemini to build applications.
06:19 The role of Cursor IDE in providing visibility into AI-generated code and the dilemma of learning versus relying on AI.
12:50 Discussing the fundamental shift from deterministic to probabilistic approaches in computer science due to LLMs.
18:48 Tracing the history of Unix, Apple, and Microsoft operating systems and their respective kernel developments.
45:03 How AI might fulfill the promise of integrating siloed enterprise data, a concept Ray Ozzie explored with Lotus Notes.
47:57 Examining Apple's highly integrated IT system as a model for enterprise efficiency and control.
56:35 The potential impact of tariffs on global manufacturing, supply chains, and the economy.
Key Insights
- Probabilistic "Vibe Coding": AI-driven coding is inherently probabilistic. Unlike traditional deterministic programming, using LLMs like Claude for coding, or "vibe coding" as I call it, means outcomes aren't guaranteed. Prompts can yield perfect results, better-than-expected innovations, or frustrating "doom loops" of errors, making it a psychologically unique experience.
- The Uncanny Valley of AI Skill: Navigating AI coding puts users in an "uncanny valley" of knowledge. One needs enough understanding to craft effective prompts and debug AI-generated code (e.g., using tools like Cursor for visibility), yet deep traditional coding expertise might become less critical as AI improves, creating a difficult balance.
- Legacy Tech's LLM Adaptation Challenge: Established companies may struggle to adapt to the LLM revolution. Businesses like Descript or Duolingo, with set processes and products, might find it hard to pivot quickly and fully leverage LLMs, potentially falling behind more agile or AI-native solutions.
- Real-Time AI's Transformative Potential: Real-time AI for tasks like transcription and translation is becoming highly effective. Tools like Google Live Transcribe and Translate demonstrate near-perfect capabilities, which could fundamentally change the necessity and approach to learning foreign languages for purely functional communication.
- The Shift from Deterministic to Probabilistic Computing: LLMs signify a major paradigm shift in computer science. We're moving from an era dominated by deterministic logic, where inputs predictably produce specific outputs, to a probabilistic one where AI generates responses based on likelihood, requiring new ways of thinking and working.
- Enterprise AI for Data Integration: AI holds significant promise for solving enterprise data silos. Just as Ray Ozzie envisioned with Lotus Notes, modern AI, especially custom-trained LLMs, could enable companies to integrate vast, disparate datasets, unlocking new insights and efficiencies, though this remains a complex challenge.
- Apple's Vertically Integrated IT Prowess: Apple's sophisticated, vertically integrated IT system is a masterclass in operational control. Their ability to manage the entire chain from silicon design to manufacturing, software, and customer delivery through tightly integrated systems showcases a level of control and efficiency few other companies achieve.
Episode transcript
Everyone is wrong about AI, but if you know the history of technology, you can be wrong and still be right.
I know the history of technology since I started with the personal computer in 1980. So listen in while we explore what is right about AI.
Welcome to the Stuart Squared podcast. We are going to go into Vibe coding and all the things that I am learning about in terms of vibe coding as I just ask Claude code to build me things and I feel like I'm in a stock machine or I feel like I'm in a slot machine casino where sometimes they get something, you know, I've now stopped paying for Descript, although it's funny because now we're recording this on Riverside and River. Riverside does the same thing as Descript. But before I go into that, I've now stopped paying $60 a month for my Descript because I was able to use the Claude code to build a solution that does what Descript does. So I was able to stop paying that and now work.
I think they're now at the stage where they're too big to understand it. Actually, one of my friends is the CFO of that company and I think they're just like, you know, in the same position as Duolingo. Same position as all of, you know, all of these companies that didn't don't really understand LLMs and don't have the time to understand it because their business process is already stuck. So I think they're going to fail to productize Duolingo. Really weird. The CEO of Duolingo recently said we're now an AI first company when it comes to building the product. But no mention of how awesome LLMs are for learning languages and how that could be built into the product itself.
Your first full transcript is free. After that, an email opens every transcript in the index — a list of readers we can write to, not a guest book.