The Dangerous Illusion of AI Coding? - Jeremy Howard

Item: The Dangerous Illusion of AI Coding? - Jeremy Howard
Rating: 4
Author: PostSingularity Institute

Humanities, Social Sciences & Thought AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence

🎙 Jeremy Howard 👥 218K 📅 March 3, 2026 ⏱ 86 min 👁 163K 📄 expert opinion 🧭 2026-08-15

Available in: English (current) Français

Keywords

AI-assisted codingfine-tuningULMFiTsoftware engineeringcognitive science

Summary

In this interview, Jeremy Howard, co-founder of fast.ai and creator of ULMFiT, discusses the origins of fine-tuning in deep learning and critically examines the current hype around AI-assisted coding. He traces the development of ULMFiT, highlighting the importance of pre-training on a general corpus and the techniques of discriminative learning rates and gradual unfreezing. Howard argues that while LLMs can perform impressive combinatorial creativity, they fundamentally interpolate within their training data and cannot truly extrapolate or innovate. He draws parallels between AI coding and a slot machine, where developers have an illusion of control but ultimately rely on outputs they don’t fully understand. He emphasizes the value of interactive programming and building deep intuition through hands-on exploration, warning that over-reliance on AI erodes technical skills and understanding. The conversation also touches on cognitive science, the importance of desirable difficulty in learning, and the broader societal risks of power centralization and enfeeblement.

151 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information is high, offering a nuanced perspective from a pioneer in the field. Howard’s arguments are well-structured, combining personal experience with references to research and historical context. He effectively challenges the prevailing narrative of AI coding productivity, using analogies like the slot machine and the interpolation illusion to illustrate his points. His reasoning is solid, though some claims are anecdotal and would benefit from more empirical backing.

Scientific Rigor, Source Quality, Title Accuracy

The discussion demonstrates scientific rigor through references to key papers and studies, such as the ULMFiT paper, the DINO paper, and the METR study on AI OS development. Howard also cites philosophical works by Dennett and Searle, and historical references like Ebbinghaus. The title accurately reflects the content, focusing on the potential dangers of AI coding. The sources are credible and relevant, though some are from commercial blogs, which may introduce bias. Overall, the title-content alignment is strong, and the sources support the arguments presented.

171 words

Title / Content Match

The title accurately reflects the central theme of the interview, which critically examines the promises and pitfalls of AI-assisted coding.

Quality & Reliability

8/10

The discussion is grounded in the speaker's extensive experience and references several credible sources, but it is primarily opinion-based and lacks systematic empirical validation.

Chapters

Cited Sources

Self-Supervised Learning and NLP — Howard references his blog post on self-supervised learning, which underpins the discussion on pre-training and fine-tuning.
Gemini Deep Think — Mentioned in the context of AI's ability to solve mathematical problems through interpolation.
The Claude C Compiler: What It Reveals About the Future of Software — Discussed as an example of AI-generated code that is actually interpolation from existing code.
Building C Compiler — Referenced in the discussion about Claude's ability to write a C compiler, illustrating the interpolation illusion.
Scaling Agents — Mentioned in the context of AI coding tools and their limitations.
NB Dev Merged Driver — Howard discusses his work on Jupyter notebooks and interactive programming.
Is Avoiding Extinction from AI Really the Most Important Thing? — Howard's response to the AI risk letter, discussed in the context of existential risk.
The Brain Abstracted — Referenced in the discussion on cognitive science and abstraction.
Consciousness Explained — Mentioned in the debate about AI understanding and consciousness.
Infinite Alphabet / Laws of Knowledge — Referenced in the discussion on organizational knowledge and the slope.
Why Creativity Cannot Be Interpolated — MLST archive article referenced in the discussion on creativity.
Early 2025 AI Experienced OS Dev Study — Referenced to support the claim that AI coding has only a tiny uptick in productivity.
No Silver Bullet — Fred Brooks' classic paper referenced in the discussion on software engineering complexity.
Minds, Brains, and Programs — Searle's Chinese Room argument referenced in the debate on AI understanding.
Emerging Properties in Self-Supervised Vision Transformers (DINO) — Referenced in the discussion on self-supervised learning and pre-training.
Sculptor Identification Paper — Referenced as an example of transfer learning in computer vision.
AI Skill Formation — Referenced in the discussion on automation and skill formation.
Ebbinghaus: Memory / Spaced Repetition — Historical reference on memory and learning.
Slope vs Intercept — Referenced in the discussion on organizational knowledge and the slope.
Inventing on Principle — Bret Victor's talk referenced in the discussion on interactive programming.
I Don't Like Notebooks — Joel Grus's talk referenced in the notebook debate.

Concurring Sources

METR Study: AI OS Development — Supports the claim that AI coding has minimal productivity gains.
No Silver Bullet — Aligns with the argument that software engineering complexity remains a challenge.

Dissenting Sources

Cursor Blog: Scaling Agents

External References

Contribution & Novelties

The interview provides a unique perspective from a pioneer in deep learning, offering insights into the origins of fine-tuning and a critical analysis of AI coding. Howard’s emphasis on the importance of interactive programming and intuition building is a valuable counterpoint to the current trend of relying on AI-generated code. The discussion on the ‘slot machine’ nature of AI coding and the interpolation illusion adds depth to the debate.

Pour aller plus loin :

ULMFiT paper — The original paper on universal language model fine-tuning.
The Bitter Lesson — Rich Sutton’s essay on the importance of computation and learning.
Desirable Difficulties — Concept from cognitive psychology relevant to the discussion on learning and friction.

114 words

Radar Profile

The radar profile shows high scores in quantity and quality of information, with a moderate technical level and reliability. This indicates a content-rich discussion that is accessible yet grounded in expertise, though some claims are opinion-based.

Reliability 7/10

💬 Très positif. Sur les 30 commentaires analysés, la grande majorité exprime une forte appréciation de la profondeur et de la pertinence de la discussion, saluant l'intelligence et l'honnêteté de Jeremy Howard.