"Vibe Coding is a Slot Machine" - Jeremy Howard - Machine Learning Street Talk (MLST) Recap

Podcast: Machine Learning Street Talk (MLST)

Published: 2026-03-03

Duration: 1 hr 27 min

Summary

Jeremy Howard critiques the current state of AI-based coding, arguing that while it creates an illusion of control, it ultimately lacks true understanding and reliability. He emphasizes the importance of human intuition and interaction in the coding process.

What Happened

In this episode, Jeremy Howard, a leading figure in deep learning and Kaggle grandmaster, shares his strong disdain for the way AI tools are currently being used in coding. He describes the experience as 'disgusting' and 'inhumane,' emphasizing his long-standing mission to prevent people from working in such a manner. Howard argues that despite the hype around AI, there's only a 'tiny uptick' in the quality of software being produced, and he compares AI-assisted coding to a slot machine, where users feel a sense of control while ultimately lacking understanding of the generated code.

Howard elaborates on the concept of 'vibe coding,' stating that while tools like large language models (LLMs) may appear to understand and generate code, they are merely simulating comprehension. He underscores the necessity of human interaction in coding, where developers must engage deeply with the problems at hand. Howard references the insights from great scientists like Feynman, who advocated for building mental models through hands-on interaction, contrasting this with the abstraction-based approach of machines that rely on statistical correlations in vast datasets. This distinction highlights the limitations of current AI tools in fostering genuine understanding and insight in software development.

Key Insights

• AI tools create an illusion of control in coding but lack true understanding.
• Human interaction and intuition in coding are irreplaceable for deep insights.
• The quality of software produced with AI assistance has not drastically improved.
• Building mental models through hands-on experience is crucial for true comprehension.

Key Questions Answered

What does Jeremy Howard think about AI coding tools?

Jeremy Howard expresses a strong aversion to the current use of AI coding tools, describing the experience as 'disgusting' and 'inhumane.' He believes that people should not be forced to rely on such methods, emphasizing that his mission has been to stop this trend for two decades. Howard argues that while these tools provide an illusion of control, they ultimately do not lead to a significant increase in the quality of the software produced.

How does Howard compare AI-based coding to a slot machine?

Howard likens AI-based coding to a slot machine, where users craft prompts and manipulate code with the hope of yielding beneficial results. However, he points out that, similar to a slot machine, the outcome is uncertain and reliant on chance. This analogy underscores the lack of true understanding and control developers have when using AI-generated code, making it risky to base significant projects on such outputs.

What key insight does Howard provide about human interaction in coding?

Howard stresses the importance of human interaction in coding, asserting that real insights emerge when developers engage deeply with the problems they are addressing. He references the practices of great scientists who build deeper intuitions through hands-on experimentation and mental modeling. This contrasts sharply with the approach taken by AI, which operates primarily through statistical analysis and abstraction.

What historical context does Howard provide regarding ULM Fit?

Howard discusses the historical significance of the ULM Fit paper, which he contributed to, as a groundbreaking approach to fine-tuning language models. He notes that it was one of the first instances where a general-purpose corpus was used for pre-training, which was a pivotal moment in NLP. His background in philosophy and cognitive science informed his understanding of this process, enabling him to create models that could effectively learn from vast datasets.

What are Howard's views on the current state of AI software engineering?

Howard is critical of the current state of AI in software engineering, stating that while there is excitement around the technology, it has not led to substantial improvements in software quality. He argues that the reliance on AI tools has resulted in a false sense of security, where developers do not fully grasp the workings of the code generated, which can lead to poor outcomes. He believes that the fundamentals of software engineering are still best served through traditional, hands-on methods.