Craig McLuckie built his career on reliability. As a co-creator of Kubernetes, he helped define the modern era of deterministic, predictable infrastructure. But today, as the CEO and co-founder of Stacklok, he is challenging engineering leaders to embrace the exact opposite: the chaotic, probabilistic world of AI.
McLuckie advocates for an "AI Maximalist" philosophy. In a landscape where many leaders cautiously ask, "Can we use AI to improve this?", he argues for a radical inversion of the question: "Why can't we use AI to do this?"
This shift isn't just about adopting a new tool; it is about challenging the fundamental operating assumptions of how value is created. It requires leaders to move from a mindset of rigid control to one of experimentation, managing the natural entropy of AI to unlock capabilities that were previously impossible.
The AI maximalist philosophy
The AI Maximalist philosophy requires a departure from traditional platform engineering thinking. In the Kubernetes era, success meant assembling known components like databases, compute, and networking, and tuning them until they behaved exactly as expected. Generative AI, however, demands an "inversion of thought."
Rather than architecting a perfect system from day one, McLuckie argues that leaders must push the envelope by figuring out what works through experimentation first, and then optimizing it over time. Trying to force-fit AI into a deterministic platform engineering box often leads to failure; success comes not from tuning a system to perfection, but from embracing the experimental nature of the technology.
Embracing the entropy of stochastic systems
To adopt this philosophy, engineers must understand the nature of the beast. Unlike the deterministic software of the past, large language models are stochastic systems, essentially a "natural source of entropy" in your environment that enables new capabilities.
In a deterministic system, input A always leads to output B. In a stochastic AI system, the output is probabilistic. This explains why "hallucinations" are not merely bugs to be fixed, but intrinsic features of how the models operate. As McLuckie points out, hallucination is effectively an artifact of the reward system during training; models are rewarded for producing a result, not for admitting ignorance.
This unpredictability complicates root cause analysis. When an AI system fails, traditional debugging tools often fall short because the behavior emerged from a probabilistic process rather than a logic error. Understanding that self-attention mechanisms eventually "collapse under their own weight" helps leaders set realistic expectations for reliability and monitoring.
Real-world experimentation
When organizations accept this stochastic nature and embrace experimentation, the results can be transformative. McLuckie’s team experienced this firsthand when they built an internal knowledge management server in just two weeks, a project he describes as "shockingly useful."
By ingesting documentation from Google Drive, GitHub, and Discord, the system allowed the team to query the organization's collective brain using natural language. This did more than just retrieve information; it flattened the organization. It removed the need to interrupt an engineering manager to find out what an engineer is working on; you could simply ask the knowledge server.
However, the experimental mindset also means accepting failure. The team attempted to build an AI-powered root cause analysis tool for Kubernetes clusters, but found that when presenting raw time-series data, the models would "collapse the context." Traditional Bayesian analysis proved superior. This failure was as valuable as the success, offering a clear boundary of where AI excels and where it struggles.
The era of disposable code and 'vibe coding'
Perhaps the most disruptive shift is the dramatic reduction in the cost of producing code. Because generating code is now "infinitely cheaper," we are entering the era of "disposable code," where prototypes can be spun up, tested, and discarded overnight.
This has given rise to vibe coding, or rapid, weekend prototyping that drives innovation. McLuckie shares a success story where an engineer had an idea on a Friday and showed up on Monday with a working proof-of-concept. This democratization of coding allows designers and marketers to produce functional prototypes that previously required engineering resources.
However, this speed creates new tensions. As McLuckie warns, "bad code takes work to become good code." Organizations must develop rigorous processes to evaluate whether a prototype should be engineered for production or simply discarded. Furthermore, the risk of shadow IT increases as non-technical teams build their own solutions using code as an "intermediate language."
Navigating the transition
For engineering leaders beginning this journey, the advice is to start simple: first use the tools to understand what is possible, and only then attempt to build and optimize them.
By starting with simple tool usage, teams develop the intuition needed to navigate the stochastic nature of AI. Embracing the AI Maximalist philosophy doesn't mean abandoning discipline; it means applying discipline to the process of experimentation. By maintaining appropriate guardrails while asking "Why can't we use AI?", engineering organizations can harness the transformative potential of these new technologies while navigating the entropy they introduce.
For the full conversation on the shift toward AI Maximalism, listen to Craig McLuckie discuss these ideas in depth on the Dev Interrupted podcast.
