Machine Learning Street Talk (MLST)

This discussion features Andrew Dudzik, a mathematician specializing in category theory; Taco Cohen, a researcher in geometric deep learning; and Petar Veličković, an expert in graph neural networks. They delve into why LLMs struggle with basic math by highlighting their pattern recognition flaws. The conversation proposes category theory as a framework to transition AI from trial-and-error towards a scientific approach. They explore concepts like equivariance, compositional structures, and the potential for unifying diverse machine learning perspectives.

Join Andrew Gordon, a behavioral science researcher at Prolific, and AI expert Nora Petrova as they delve into the flaws of current AI benchmarking. They challenge the notion that high scores mean better models, using a Formula 1 car as an analogy. The discussion touches on critical issues like AI safety, especially in sensitive contexts like mental health, and critiques the biases in popular ranking systems. Discover how Prolific's innovative HUMAINE framework and TrueSkill methodology aim to create a more human-centered evaluation of AI.

In a captivating discussion, Professor Yi Ma, a pioneer in deep learning and computer vision, challenges our perceptions of AI. He explains how language models primarily memorize rather than understand, and he distinguishes between 3D reconstruction and true comprehension. Yi introduces the principles of parsimony and self-consistency as crucial to intelligence. The conversation touches on the evolution of knowledge, the limitations of current AI models in achieving abstraction, and the potential of coding rate reduction to enhance learning mechanisms.

Pedro Domingos, a leading computer science professor at the University of Washington and author of The Master Algorithm, unveils his groundbreaking concept, TensorLogic. He discusses how this innovative programming language could unify the fragmented worlds of Deep Learning and Symbolic AI. Pedro reveals TensorLogic's capabilities in logical reasoning and learning from data, emphasizing its potential to prevent AI hallucinations. He also shares insights on how TensorLogic can express complex systems and improve AI education, paving the way for a more integrated future in artificial intelligence.

In this engaging discussion, Llion Jones, co-founder of Sakana AI and co-author of the Transformer architecture, shares insights on the need for innovation beyond Transformers in AI research. Joined by Luke Darlow, a specialist in biologically inspired models, they explore the limitations of current AI paradigms and introduce the Continuous Thought Machine (CTM). This novel model emphasizes internal reasoning and adaptive computation, aiming to enhance how AI processes information. Expect fascinating analogies and thought-provoking concepts that challenge the status quo!

Phelim Bradley, Co-founder and CEO of Prolific, discusses the vital role humans play in AI development. He explains how human intelligence and data train models, emphasizing quality over quantity. Phelim reveals the challenges of matching the right experts to specific tasks and highlights how Prolific fosters long-term relationships to improve data quality. He also addresses the geopolitical concerns surrounding AI centralization and the evolving future of work, where expert judgment will be increasingly in demand as AI continues to grow.

In this engaging discussion, Prof. Chris Kempes, a quantitative biophysicist at the Santa Fe Institute, explores the search for a universal theory of life that transcends Earth-bound definitions. He introduces a three-level hierarchy: Materials, Constraints, and Principles, highlighting how different life forms could emerge from diverse substrates. Chris delves into the convergence of evolution, using the eye as a compelling example, and raises thought-provoking questions about whether concepts like culture and AI can also be considered forms of life.

Blaise Agüera y Arcas, a pioneering scientist and author of "What Is Intelligence?", shares revolutionary ideas on the relationship between life and intelligence. He argues that DNA functions as a computer program, proposing that evolution's complexity comes from merging systems rather than just mutations. Blaise also discusses his BFF experiment, showing how self-replicating programs can emerge from randomness. He explores how both AI and human intelligence are part of a larger collective, reshaping our understanding of purpose and consciousness.

Sara Saab, VP of Product at Prolific with a background in cognitive science, and Enzo Blindow, VP of Data and AI at Prolific and an expert in economics, discuss the pivotal role of human feedback in AI. They stress that non-deterministic AI systems require human oversight more than ever, as optimizing for benchmarks can mislead usability. Exploring the ecological context of intelligence, they advocate for a participatory approach to evaluation that captures social norms and emphasizes the importance of cultural alignment.

Dr. Ilia Shumailov is a former DeepMind AI security researcher now focused on building security tools for AI agents. He delves into the unique challenges posed by AI agents operating 24/7, generating hacking tools at unprecedented speeds. Ilia emphasizes that traditional security measures fall short and discusses new adversarial threats, including prompt injection attacks. He also explores the risks of model collapse and the importance of fine-grained policies for AI behavior, warning that as AI evolves, its unpredictability could lead to significant security vulnerabilities.