In this episode of the Crazy Wisdom Podcast, I, Stewart Alsop, talk with AJ Beckner about how AI is reshaping the terrain of software development—from the quiet ritual of mate-fueled coding sessions to the radical shift in DevOps, tool use, and what it means to build software in a post-LLM world. AJ shares insights from his time in the Gauntlet AI program, reflecting on how platforms like Cursor, Lovable, and Supabase are changing what’s possible for both seasoned engineers and newcomers alike. We also explore the nuanced barbell dynamic of skill disruption, the philosophical limits of current AI tooling, and how rapid prototyping has morphed from a fringe craft into a mainstream practice. You can find more about AJ on Twitter at @thisistheaj.

Check out this GPT we trained on the conversation!

Timestamps

00:00 — Stewart and AJ kick off with mate culture, using it as a metaphor for vibe coding and discussing AJ’s caffeine stack of coffee and yerba mate.
05:00 — They explore how AI coding reshaped AJ’s perspective, from URBIT and functional languages to embracing JavaScript due to LLMs’ strength in common corpuses.
10:00 — AJ breaks down why DevOps remains difficult even as AI accelerates coding, comparing deployment friction across tools like Cursor, Replit, and Lovable.
15:00 — They outline the barbell effect of AI disruption—how seasoned engineers and non-technical users thrive while the middle gets squeezed—and highlight Supabase’s role in streamlining backends.
20:00 — AJ dives into context windows, memory limits, and the UX framing of AI’s intelligence. Cursor becomes a metaphor for tooling that “gets it right” through interaction.
25:00 — Stewart reflects on metadata, chunking, and structuring his 450+ podcast archive for AI. AJ proposes strategic summary hierarchies and cascading summaries.
30:00 — The Gauntlet AI program emerges as a case study in training high-openness engineers for applied AI work, replacing skeptical Stanford CS grads with practical builders.
35:00 — AJ outlines his background in rapid prototyping and how AI has supercharged that capacity.
40:00 — The conversation shifts to microservices, scale, and why shipping a monolith is still the right first move.
45:00 — They close with reflections on sovereignty, URBIT, and how AI may have functionally solved the UX problems URBIT originally aimed to address.

Key Insights

1. AI reshapes the value of programming languages and stacks: AJ Beckner reflects on how large language models have flipped the script on what makes a programming language or stack valuable. In a pre-LLM world, developers working with niche, complex systems like URBIT sought to simplify infrastructure to make app development manageable. But now, the vast and chaotic ecosystem of mainstream tools like JavaScript becomes a feature rather than a bug—LLMs thrive on the density of tokens and the volume of shared patterns. What was once seen as a messy stack becomes a fertile ground for AI-assisted development.
2. DevOps remains the bottleneck even in an AI-accelerated workflow: Despite the dramatic speedups in building features with AI tools, deployment still presents significant friction. Connecting authentication, deploying to cloud services, or managing infrastructure introduces real delays. Platforms like Replit and Lovable attempt to solve this by integrating backend services like Supabase directly into their stack, but even then, complexity lingers. DevOps hasn’t disappeared—it’s just more concentrated and increasingly a specialization distinct from traditional app development.
3. The 'barbell effect' splits value between experts and DIY builders: AI tools are democratizing access to software creation, allowing non-technical users to build functional apps. Yet this doesn’t flatten the playing field—it polarizes it. There’s growing value in both high-skill engineers who can build and maintain scalable, AI-enhanced systems, and low-skill users leveraging AI to extend their reach. But the middle—average developers doing routine app-building—risks being squeezed out. Success now hinges on openness, adaptability, and knowing where to play on the spectrum.
4. Context isn’t about memory—it’s about tooling: A key insight AJ offers is that context window size isn’t the real limiter for LLMs; it’s a UX issue. Tools like Cursor demonstrate that with the right infrastructure—error tracking, smart diffs, and contextual awareness—you don’t need infinite memory, just the right slices of relevance. It mirrors human memory: we don’t recall everything, just the crucial bits. Designing around that constraint is more impactful than waiting for models with bigger context windows.
5. The future belongs to those who prototype with conviction: AJ emphasizes that the ability to spin up ideas rapidly with AI is more than a technical advantage—it’s a philosophical one. People coming from a culture of iteration and lightweight experimentation are uniquely positioned to capitalize on this moment. He contrasts this with legacy engineers and institutions still approaching AI with skepticism or nostalgia. The ones who win will be those willing to reimagine their workflows, not retrofit AI into old mental models.
6. Sovereignty and convenience sit in quiet tension: While platforms like URBIT aim to give users sovereign control over their compute, AJ points out that ease-of-use platforms like Lovable and Replit are delivering on the practical dream of custom software for all. There’s a moral ideal to sovereignty, but in practice, many people just want software that adapts to their needs, not servers they can control. The deeper question isn’t just about owning your stack, but about where ownership and usability intersect.
7. AI doesn’t eliminate engineering—it clarifies its edges: Far from replacing engineers wholesale, AI sharpens the distinction between development and engineering. The role of site reliability engineers (SREs), DevOps specialists, and system architects remains crucial—especially in scaling, monitoring, and maintaining complex systems. Meanwhile, traditional “web dev” becomes increasingly automated. The craft of engineering doesn’t vanish; it migrates to the edges, where problems are less about writing code and more about ensuring resilience, scalability, and coherence.