Machine Learning Street Talk (MLST)

In this engaging discussion, Subbarao Kambhampati, a Professor at Arizona State University specializing in AI, tackles the limitations of large language models. He argues that these models primarily memorize rather than reason, raising questions about their reliability. Kambhampati explores the need for hybrid approaches that combine LLMs with external verification systems to ensure accuracy. He also delves into the distinctions between human reasoning and LLM capabilities, emphasizing the importance of critical skepticism in AI research.

Sayash Kapoor, a Ph.D. candidate at Princeton, dives deep into the complexities of assessing existential risks from AI. He argues that the reliability of probability estimates can mislead policymakers, drawing parallels to other fields of risk assessment. The discussion critiques utilitarian approaches in decision-making and the challenges with cognitive biases. Kapoor also highlights concerns around AI's rapid growth, pressures on education, and workplace dynamics, emphasizing the need for informed policies that balance technological advancement with societal impact.

Sara Hooker, VP of Research at Cohere and a leading voice in AI efficiency, shares insights on AI governance and the pitfalls of using compute thresholds, like FLOPs, as risk metrics. She critiques current US and EU policies for oversimplifying AI capabilities and emphasizes the need for a holistic view that includes data diversity. Hooker also discusses her research on 'The AI Language Gap,' revealing the complexities of creating inclusive AI that serves multilingual populations, highlighting ethical concerns and the societal implications of underrepresentation in AI development.

Murray Shanahan, a Professor of Cognitive Robotics at Imperial College London and a senior research scientist at DeepMind, dives deep into AI consciousness and the perils of anthropomorphizing machines. He discusses the limitations of current language in describing AI and stresses the need for nuanced vocabulary. Shanahan explores Reinforcement Learning and the 'Waluigi Effect,' as well as the complexities of agency in AI. He also touches on consciousness in relation to non-human entities, emphasizing how our perceptions shape understanding and the philosophical implications behind it.

David Chalmers, a leading philosopher and cognitive scientist from NYU, dives deep into the intersection of technology and reality. He asserts that virtual worlds could become indistinguishable from the real one, challenging the notion that they are mere escapism. Chalmers explores profound questions about consciousness, the nature of existence, and the implications of living in simulations. He also tackles the ethical dimensions of AI and the philosophical intricacies of metaverses, arguing that our interactions in virtual spaces can be genuinely meaningful.

Ryan Greenblatt, a researcher at Redwood Research known for his groundbreaking work on the ARC Challenge, discusses his innovative use of GPT-4 to achieve impressive accuracy. He explores the strengths and weaknesses of current AI models and the profound differences in learning and reasoning between humans and machines. The conversation touches on the risks of advancing AI autonomy, the effects of over-parameterization in deep learning, and the potential future advancements, including the promise of multimodal capabilities in forthcoming models.

Aidan Gomez, CEO of Cohere, dives into tackling AI hallucinations and reasoning improvements in this lively conversation. He discusses why Cohere avoids using GPT-4 for training and shares insights on the unique challenges enterprises face with AI, from onboarding to legislation. Aidan highlights their commitment to crafting robust, industry-specific solutions while addressing the societal implications of AI advancements and regulatory needs. Get a glimpse into Cohere’s ethos and their strategic vision for the future of AI technology.

In this engaging discussion, Jack Cole, a clinical psychologist and ARC Challenge winner, along with AI researcher Mohammed Osman and expert Michael Hodel, delve into the nuances of the ARC Challenge, which assesses AI reasoning. They present their winning approach of fine-tuning language models, emphasizing active inference and innovative data representation. The trio debates the philosophical implications of their methods on intelligence measurement while highlighting the addictive nature of ARC tasks and raising questions about the future of AI and generalization.

Nick Frosst, Co-founder of Cohere, previously at Google Brain alongside AI pioneer Geoff Hinton, shares insights on AI's future in business. He discusses Cohere's Command R models, which enhance language capabilities using retrieval augmented generation (RAG). Nick critiques the chase for AGI, emphasizing specialization in LLMs over generalization. He also touches on ethical data use in AI, the evolving role of software engineers in machine learning, and even gives a nod to his indie band, Good Kid, showcasing the intersection of creativity and technology.

Aman Bhargava, a PhD student at Caltech, and Cameron Witkowski, a graduate student at the University of Toronto, dive into their groundbreaking research on controlling language models using control theory. They discuss how language models operate as discrete systems and the surprising impact of prompt engineering on outputs. By examining the "reachable set" of outputs, they reveal that even minor tweaks in prompts can lead to significant changes in generated text. Their insights could pave the way for more reliable and capable AI systems.