Handling black boxes and debugging LLMs

In this episode of Crazy Wisdom, host Stewart Alsop talks with Jared Zoneraich, CEO and co-founder of PromptLayer, about how AI is reshaping the craft of software building. The conversation covers PromptLayer’s role as an AI engineering workbench, the evolving art of prompting and evals, the tension between implicit and explicit knowledge, and how probabilistic systems are changing what it means to “code.” Stewart and Jared also explore vibe coding, AI reasoning, the black-box nature of large models, and what accelerationism means in today’s fast-moving AI culture. You can find Jared on X @imjaredz and learn more or sign up for PromptLayer at PromptLayer.com.

Check out this GPT we trained on the conversation

Timestamps

00:00 – Stewart Alsop opens with Jared Zoneraich, who explains PromptLayer as an AI engineering workbench and discusses reasoning, prompting, and Codex.
05:00 – They explore implicit vs. explicit knowledge, how subject matter experts shape prompts, and why evals matter for scaling AI workflows.
10:00 – Jared explains eval methodologies, backtesting, hallucination checks, and the difference between rigorous testing and iterative sprint-based prompting.
15:00 – Discussion turns to observability, debugging, and the shift from deterministic to probabilistic systems, highlighting skill issues in prompting.
20:00 – Jared introduces “LM idioms,” vibe coding, and context versus content—how syntax, tone, and vibe shape AI reasoning.
25:00 – They dive into vibe coding as a company practice, cloud code automation, and prompt versioning for building scalable AI infrastructure.
30:00 – Stewart reflects on coding through meditation, architecture planning, and how tools like Cursor and Claude Code are shaping AGI development.
35:00 – Conversation expands into AI’s cultural effects, optimism versus doom, and critical thinking in the age of AI companions.
40:00 – They discuss philosophy, history, social fragmentation, and the possible decline of social media and liberal democracy.
45:00 – Jared predicts a fragmented but resilient future shaped by agents and decentralized media.
50:00 – Closing thoughts on AI-driven markets, polytheistic model ecosystems, and where innovation will thrive next.

Key Insights

1. PromptLayer as AI Infrastructure – Jared Zoneraich presents PromptLayer as an AI engineering workbench—a platform designed for builders, not researchers. It provides tools for prompt versioning, evaluation, and observability so that teams can treat AI workflows with the same rigor as traditional software engineering while keeping flexibility for creative, probabilistic systems.
2. Implicit vs. Explicit Knowledge – The conversation highlights a critical divide between what AI can learn (explicit knowledge) and what remains uniquely human (implicit understanding or “taste”). Jared explains that subject matter experts act as the bridge, embedding human nuance into prompts and workflows that LLMs alone can’t replicate.
3. Evals and Backtesting – Rigorous evaluation is essential for maintaining AI product quality. Jared explains that evals serve as sanity checks and regression tests, ensuring that new prompts don’t degrade performance. He describes two modes of testing: formal, repeatable evals and more experimental sprint-based iterations used to solve specific production issues.
4. Deterministic vs. Probabilistic Thinking – Jared contrasts the old, deterministic world of coding—predictable input-output logic—with the new probabilistic world of LLMs, where results vary and control lies in testing inputs rather than debugging outputs. This shift demands a new mindset: builders must embrace uncertainty instead of trying to eliminate it.
5. The Rise of Vibe Coding – Stewart and Jared explore vibe coding as a cultural and practical movement. It emphasizes creativity, intuition, and context-awareness over strict syntax. Tools like Claude Code, Codex, and Cursor let engineers and non-engineers alike “feel” their way through building, merging programming with design thinking.
6. AI Culture and Human Adaptation – Jared predicts that AI will both empower and endanger human cognition. He warns of overreliance on LLMs for decision-making and the coming wave of “AI psychosis,” yet remains optimistic that humans will adapt, using AI to amplify rather than atrophy critical thinking.
7. A Fragmented but Resilient Future – The episode closes with reflections on the social and political consequences of AI. Jared foresees the decline of centralized social media and the rise of fragmented digital cultures mediated by agents. Despite risks of isolation, he remains confident that optimism, adaptability, and pluralism will define the next AI era.