Jeremy Howard
Deep learning education
Co-founder of fast.ai, former Kaggle president. Pioneer in practical deep learning education.
Recent activity
-
High-stakes standardized tests and the emphasis on quantifying learning metrics can lead to a dehumanizing education system that prioritizes test scores over creative expression and meaningful relationships, ultimately harming students' love for learning and overall well-being. AI summary
Read more → -
Vibe coding, a practice of generating large amounts of complex AI-generated code without human review, can be likened to gambling due to its addictive qualities and the "Loss Disguised as a Win" phenomenon, where users feel a sense of accomplishment despite not achieving actual progress or success. This can lead to "dark flow," a variation of true flow characterized by an insidious, superficial experience that becomes addictive rather than growth-producing. As a result, many developers experience frustration and disappointment when they realize the true value of their creations is minimal. AI summary
Read more → -
Large Language Models (LLMs) can aid in the ancient art of close reading by providing a tool for identifying outward connections and clarifying questions, allowing readers to pause and ask 'What does this term mean?' or 'How does this connect to what came before?' This can lead to a more engaging and personalized reading experience, as seen in experiments with the fast.ai course 'How to Solve It With Code'. AI summary
Read more → -
Adults often glorify boredom in children, despite its negative effects on their well-being, disengagement, and stress. Boredom is a state of lack of interest, dullness, or tedious repetition, not a natural part of childhood, and can lead to a lack of curiosity, creativity, and resilience. Unstructured play, on the other hand, is essential for children's development and autonomy. AI summary
Read more → -
Solveit's platform features support an interactive, incremental approach to development, allowing users to build solutions piece by piece, with the AI suggesting code based on the user's full working context. The platform includes features such as real-time collaboration, Markdown support, and a symbol browser, with a focus on providing working code that users can understand and modify. Solveit also offers tools like VS Code integration, code block extraction, and thinking mode for quick AI continuation. AI summary
Read more → -
Chris Lattner, creator of LLVM, Clang, and Swift, advocates for building software from first principles, focusing on fundamental design and architecture that can last decades, rather than prioritizing speed and efficiency. He emphasizes the importance of craftsmanship and engineering culture in producing technically excellent systems. Lattner's approach has led to the development of widely-used programming languages and tools. AI summary
Read more → -
To implement a GPT tokenizer, one can use the Byte Pair Encoding (BPE) algorithm, which is a technique to reduce the number of unique tokens in a vocabulary, or use a regex-based pre-tokenization approach. Alternatively, libraries like tiktoken provide pre-built tokenizers, and developers can build their own using the minbpe exercise, which involves implementing a basic BPE implementation, adding regex preprocessing, loading GPT-4 merges, handling special tokens, and exploring SentencePiece. AI summary
Read more → -
fast.ai's "How to Solve it With Code" course is now available, offering a new approach to harnessing AI for creative activities like software development and writing, focusing on human agency and growth. The course is designed for experienced coders, AI practitioners, and data scientists, and includes a software platform, community, and ongoing R&D. The "solveit method" aims to provide a sustainable way to learn and grow using AI, with a course, platform, and community available at solve.it.com. AI summary
Read more → -
The fasttransform library in Python provides a simple way to make data transformations reversible and extensible through the power of multiple dispatch, allowing users to easily invert transformations and visualize their data. This is particularly useful for machine learning model debugging, where understanding the data transformations applied to the input can be crucial for interpreting model predictions. AI summary
Read more → -
Foundation models for computational pathology, such as Prov-GigaPath and UNI, can analyze images from multiple types of tissue and multiple tasks, including cancer cell grading, identifying early signs of rejection, and detecting genetic mutations, by leveraging patterns learned from diverse datasets. These models have achieved state-of-the-art performance on dozens of pathology tasks. The success of these models is attributed to the concept of foundation models, which are trained on large, diverse datasets to generalize across different tasks and datasets. AI summary
Read more →