LessWrong (Curated & Popular)

Yoshua Bengio, a Turing Award winner for deep learning research, discusses the importance of deep models and understanding in ML. Topics include Unitary Evolution Recurrent Neural Networks and gradient explosion/death in recurrent nets. They also explore imposter syndrome and progress in fields through feedback loops and admitting lack of knowledge.

The podcast compares Anthropic's dictionary learning technique with a sparse autoencoder approach in analyzing language models. It discusses the similarities, differences, and success of the dictionary learning approach. It also compares the language models and sparse autoencoder architecture used by the two teams. The podcast explores the differences in dictionary learning approaches and training methods, including architectural variations, training set sizes, dead neuron handling, and feature interpretability. The effects of editing model activations in an AI language model and a form of automatic interpretability are also discussed.

Gretta Duleba, CEO of MIRI, introduces the new leadership team and discusses the shift towards broad public communication. The podcast explores the transition of leadership at MIRI, introduces the new CEO and team, and discusses their strategic plans to address AI's existential risk.

The podcast discusses cohabitive games that combine cooperation and competition. It explores the importance of negotiation in board games and introduces character concepts. The podcast also focuses on enhancing cooperative bargaining games and exploring the distinction between maximizing scores and real-world elements. It discusses the factors contributing to group success in cohabitive games and the potential of virtual reality in gameplay.

Discover the new content feature on LessWrong called Dialogues, allowing for in-depth explanations of world-models. Learn how to invite partners and find interesting conversation topics. Explore the features and etiquette of writing dialogues while finding someone to try out this exciting feature.

The podcast discusses the debate around aligning AI with human values, including GPT-4's potential as a human value function. It explores the challenges of instructing AI and the value identification problem in AI, emphasizing the need for accurately specifying objectives.

The podcast discusses the challenges and solutions for understanding the behaviors of neural networks. It explores the use of features instead of individual neurons and the decomposition of language models into interpretable parts. The concept of interpretability in language models is also explored, highlighting the importance of features and the influence of activating specific features. The potential for decomposing models into interpretable features is discussed, along with the universality of learned features and anthropics investment in mechanistic interpretability.

Guest Quintin Pope argues that the sudden growth in human capabilities was due to cultural transmission, not evolution, and this does not apply to AI. The podcast explores the parallels between human evolution and AI development, the challenges of alignment techniques, and the importance of discernment in autonomous learning.

Thomas Kwa shares his research experience working with MIRI. They discuss concrete problems, information hazard prevention, agent reflection, deconfusion, empirical feedback, and understanding cognition.

Explore the newly renovated Light Haven Campus available for bookings. Learn about the range of services, buildings, and goals of Lighthaven. Discover the wide variety of events that can be hosted with preferential pricing for projects beneficial to the world.