← Originals

Jeremy Howard

5 interviews

What they argue

Jeremy Howard is an AI researcher and engineer who focuses on the practical application of machine learning, specifically regarding efficient model training and software engineering. He works in the territory of 'Practical AI R&D,' advocating for strategies that prioritize high-impact products and deep technical understanding over vanity metrics.

  1. The GPU-Poor Strategy

    Developers should focus on fine-tuning, optimized inference, and application-layer innovations rather than pre-training large models from scratch. This strategy allows for building high-impact products with limited compute resources.

    From AI Magic: Shipping 1000s of successful products with no managers and a team of 12 — Jeremy Howard of Answer.ai (episode)

  2. The Fallacy of Fine-Tuning

    Fine-tuning is not a separate mode of learning but is simply pre-training on a different distribution. The standard three-step approach is flawed because it often lacks the diversity of the original pre-training, leading to catastrophic forgetting.

    From The End of Finetuning — with Jeremy Howard of Fast.ai (episode)

  3. Continued Pre-training Framework

    Instead of a standard three-step process, models should undergo continuous training that maintains a diverse data mix. This prevents the model from losing general intelligence while it acquires new capabilities.

    From The End of Finetuning — with Jeremy Howard of Fast.ai (episode)

  4. The 'Slot Machine' Illusion of Control

    AI-based coding currently functions like a slot machine where users pull a lever of prompts to see what code the model spits out. This creates a danger where humans ship code they cannot explain, degrading their ability to engineer complex systems.

    From The Dangerous Illusion of AI Coding? - Jeremy Howard (episode)

    • A related idea Nicholas Carr — The Slot Machine AnalogyCarr compares the internet's notification system to a slot machine. This illustrates how digital environments are designed to capture and fragment our attention.The Big Switch

  5. Interactive Loop for Learning

    To avoid 'vibe coding,' developers should use a loop that begins by identifying the core constraint of a problem. This ensures the human is actually learning and understanding the underlying mechanics rather than just generating functional software.

    From The Dangerous Illusion of AI Coding? - Jeremy Howard (episode)

    • Pushes the other way David Yaffe-Bellany — The 'Vibe Coding' WorkflowThis is an iterative process for moving from an idea to a functional tool using AI. It begins by identifying a repetitive, tedious 'itch' in one's professional life to…Is Google Search Cooked? + We’re Getting a U.S. Crypto Reserve? + What You’re Vibecoding

  6. Practical AI R&D

    There is a massive, underserved opportunity in the art of building products efficiently with small teams. This involves a stack that prioritizes speed and simplicity, such as using QDoRA and FSDP to train large models on limited hardware.

    From AI Magic: Shipping 1000s of successful products with no managers and a team of 12 — Jeremy Howard of Answer.ai (episode)

    • Pushes the other way Dario Amodei — The Big Blob of Compute HypothesisThe primary drivers of AI progress are not 'clever' architectural tricks, but rather raw compute, data quantity, data quality/distribution, and training techniques.…Dario Amodei — “We are near the end of the exponential”
    • A related idea Dylan Patel — The RouterThe next frontier of AI is the development of smarter 'routers' that dynamically allocate compute resources to solve the economic problem of scaling intelligence. This…Dylan Patel on GPT-5’s Router Moment, GPUs vs TPUs, Monetization

Every idea names the summary it came from. Written by a local model from this mind’s own work, never from memory.

Interviews 3

Open Jeremy Howard in Originals