Peter Williams’ Post

Best practices and lessons for fine-tuning Anthropics Claude 3 Haiku on Amazon Bedrock In this post, we explore the best practices and lessons learned for fine-tuning Anthropic’s Claude 3 Haiku on Amazon Bedrock. We discuss the important components of fine-tuning, including use case definition, data preparation, model customization, a... https://lnkd.in/eBWh2_wg #AI #ML #Automation

Deep Dive into Anthropic’s Sparse Autoencoders by Hand ✍️ "Anthropic’s journey ‘towards extracting interpretable features’ Following the story of Zephyra, Anthropic AI delved into the expedition of extracting meaningful features in a model. The idea behind this investigation lies in understanding how different components in a neural network interact with one another and what role each component plays. According to the paper “Towards Monosemanticity: Decomposing Language Models With Dictionary Learning” a Sparse Autoencoder is able to successfully extract meaningful features from a model. In other words, Sparse Autoencoders help break down the problem of ‘polysemanticity’ — neural activations that correspond to several meanings/interpretations at once by focusing on sparsely activating features that hold a single interpretation — in other words, are more one-directional. To understand how all of it is done, we have these beautiful handiworks on Autoencoders and Sparse Autoencoders by Prof. Tom Yeh that explain the behind-the-scenes workings of these phenomenal mechanisms." https://lnkd.in/dnwVXtCf

Excellent article by Rich Mironov

"As Socrates taught, being able to say 'I don't know' is an essential step on the path to wisdom." In his new book, NEXUS, Yuval Noah Harari makes an impassioned case for all of us to recognize our limited understanding of the potential threat — and promise — of AI and to work together as a species to create humanistic guardrails that will help us navigate the future and rediscover our shared humanity. I can't recommend highly enough. https://www.ynharari.com/

I'm a bit disappointed with day 3 of the 5-day Gen AI with Google intensive course.  Tools and features are the most valuable benefits of AI. Only tools give AI the ability to act in the real world.  The weakest white paper last days I read. https://lnkd.in/dZnJd3XU

Huge thanks to @alexalbert for sharing a sneak peek of Claude 3! I'm particularly excited about Claude FM, which promises to: Reduce prompt engineering time – This is a game-changer for efficiency! Improve agentic workflow – Claude 3 seems to be taking big strides in automation. Extract features from images – Multimodal capabilities open exciting new avenues. All this while maintaining a high bar for responsible AI development. Additionally, the potential of using Haiku as a sub-agent (as explored in https://github.com/haiku) is intriguing. This is definitely a development to keep an eye on! #Claude3 #AI #MachineLearning #ResponsibleAI

That crazy genius that was Salvador Dali exhorted us to conquer the irrational. In 1935 he wrote: “We are swimming between two kinds of water, the cold water of art  and the warm water of science”. Today, with the impact of Artificial Intelligence, Dali’s words resonate. In my latest blog post I try to connect the ideas of the master of the irrational with the realities and challenges that AI pose today.  https://lnkd.in/eY_T-Pru

NEW: Watch the latest #KempnerInstitute Seminar Series talk: Yoav Goldberg of Bar-Ilan University and Ai2 describes his work on the limits of LLMs, and how new abilities become possible when LLMs are embedded in larger systems. https://lnkd.in/ejH3Z2gQ #ML #AI

Excited to share my perspective on the fascinating topic of `LLM-as-Judge`

Check out the latest research from Lin Zhang, Jaime Smith, and Chao Wang on expanding the use of AI to expedite access to clinically relevant data in China: