This is a little experiment in documenting what I’m reading. The following is just a list of some of the articles, papers, and the occasional podcast I came across recently. Sharing them here is my way of trying to get into a better groove of reading, but I hope that by sharing I’ll also get to interact more with the world. Perhaps something here will be useful to you. Maybe you’ll send along something you think I should read next. Either way, this is an experiment and the process and format will likely evolve.
Here are some of the curiosities I read this past week:
Yamada et al. introduce The AI Scientist-v2, an autonomous system that formulates hypotheses, conducts experiments, and writes scientific papers, achieving a milestone by having an AI-generated manuscript pass peer review, showcasing AI's expanding role in scientific discovery.
Sholto Douglas and Trenton Bricken on the Dwarkesh Podcast explore the transformative shifts in AI research over the past year, delving into the scaling potential of reinforcement learning (RL), the intricacies of mechanistic interpretability, and strategies for nations, workers, and students to brace for the advent of artificial general intelligence (AGI).
The release of ether0, a 24B open-weights reasoning model from FutureHouse trained with reinforcement learning, marks a significant leap in AI-driven scientific discovery, excelling in designing drug-like molecules by reasoning in natural language and offering a glimpse into the future of general-purpose scientific reasoning models.
Rane et al. propose that principles from animal cognition research, such as designing adversarial control conditions and analyzing failure types, can significantly enhance the evaluation of LLMs by providing a more nuanced understanding of their cognitive capabilities and limitations.
In a leap for computational chemistry, a team from Microsoft Research have harnessed deep learning to dramatically enhance the accuracy of density functional theory (DFT), enabling reliable predictions of experimental outcomes
The Sergey Levine explores the intriguing notion that while LLMs have adeptly emulated human cognitive abilities by reverse-engineering the mental processes reflected in internet text, they fall short in acquiring new skills and representations from real-world experiences, highlighting the need for AI systems to evolve beyond mere replication of human intelligence to truly embody flexible and adaptive learning.
Huang et al. introduce Biomni, a general-purpose biomedical AI agent that autonomously executes diverse research tasks by integrating LLM reasoning with dynamic planning and code-based execution.
Peter Putnam’s Theory of Mind: Peter Putnam, once associated with luminaries like Einstein, developed a groundbreaking, yet largely incomprehensible theory of the mind. Amanda Gefter tells his amazing story.
Kosmyna et al. claim that using AI assistants like ChatGPT for essay writing can lead to a measurable decline in cognitive engagement and learning skills, as evidenced by weaker neural connectivity and reduced perceived ownership of work compared to traditional methods. More thoughts on this eyebrow-raising paper later…
Liu et al. show that cutting-edge LLMs, such as GPT-4o and Llama-3, inaccurately model human decision-making by assuming people are more rational than they truly are, aligning more closely with expected value theory rather than actual human behavior, which interestingly mirrors the common human expectation of rationality in others.
Anthropic's exploration into multi-agent research systems reveals that these architectures excel in handling breadth-first queries by dynamically distributing tasks among parallel agents, with prompt engineering emerging as a crucial factor in optimizing performance and coordination.