Welcome to this week's exploration of the cutting-edge intersections between AI, ML, bio-ML, and neuroscience. We came across quite a few papers on LLMs this week (and every week, we predict, till we die). And yet, not everything is about LLMs: This week’s selection is diverse, with articles, Substacks, and other essays that range from blood-DNA cancer detection to the maturities of scientific fields, and hybrid human-AI thinking systems. We hope you enjoy each of these.
Andy Clark argues that embracing our inherent nature as hybrid thinking systems, which seamlessly integrate non-biological resources like generative AI, can transform our understanding of human-AI collaborations from a threat to an opportunity for enhancing creativity and cultural evolution.
Sampaio et al. explore the complex social dynamics of octopus-fish hunting groups, showing the composition-dependent properties of leadership across multiple species. They find that social influence is hierarchically distributed over various scales of role specializations—that is—some fish drive environmental exploration while octopuses seem to have the final call for when the group moves to a new location. Find out more about what makes a de facto leader and what “partner control mechanisms” like punching do for your social influence!
Krakauer et al. explore the concept of emergence in LLMs, arguing that while these models exhibit emergent capabilities—such as sudden improvements in tasks with increased scale—they fall short of demonstrating emergent intelligence, which requires the internal use of coarse-grained representations to solve a broad range of problems, akin to human reasoning.
In a compelling exploration of the role of human-generated scientific writing amidst the rise of LLMs, the editors of Nature Reviews Bioengineering argue that while LLMs can enhance readability and assist in brainstorming, the act of writing remains a crucial cognitive process that helps researchers distill years of data into coherent narratives, thereby fostering deeper insights and innovation. Don’t let your LLM think for you.
In an essay for Asimov Press, Slime Mold Time Mold (the author’s real, legal name, we’re sure) argues that a mature science emerges when it describes its domain through entities, properties, and rules, moving beyond mere generalization to mechanistic models, as illustrated by the evolution from alchemy to chemistry and the complex systems in games like Dwarf Fortress.
In another incredible math video by 3Blue1Brown, Grant Sanderson goes into the attention mechanism of the transformer architecture.
Liquid biopsies, which analyze cell-free DNA in the bloodstream, offer a promising new approach to early cancer detection by potentially identifying malignancies without invasive procedures. Siddhartha Mukherjee chronicles this fascinating quest in The New Yorker.
The Platonic Representation Hypothesis posits that as AI models grow larger and more sophisticated, they converge towards a shared representation space. Jack Morris explores the idea that this convergence could enable universal embedding inversion, allowing for the translation between different models' embedding spaces.
Hwang et al. introduce GEOFACT-X, a geography-based multilingual factual reasoning benchmark, and BRIDGE, a novel training method that enhances multilingual reasoning fidelity by aligning LLM reasoning processes with the input language, demonstrating significant improvements in reasoning accuracy and language consistency across low-resource languages.