escarpment

semirandom, related, readings

These are 3 PDFs on my desktop that I should catch up on, each with varying depths of focus, analysis, and insight (capsule-summarized by GPT o4-mini).

On the Biology of a Large Language Model
      by Jack Lindsey, et al. | Anthropic

In “On the Biology of a Large Language Model,” the authors borrow tools from neuroscience—recording activations with feature probes, lesioning weights or heads, stimulating patterns to drive behavior, and imaging distributed representations—to systematically dissect a transformer’s micro-scale units (attention heads and MLP channels), meso-scale subcircuits, and macro-scale network pathways. Their results reveal that many “emergent” capabilities ride on surprisingly sparse, modular circuits—attention heads often act like logical gates and small MLP channels carry out core computations—while also highlighting the limits of linear probing and isolated perturbations. This work lays out a replicable blueprint for building a transformer “connectome,” offering a first glimpse into how LLMs organize and execute complex linguistic computations.

My Brain Finally Broke
      by Jia Tolentino | The New Yorker

In “My Brain Finally Broke,” New Yorker writer Jia Tolentino recounts how years of unrelenting digital immersion—endless news feeds, social-media scrolling, and nonstop writing deadlines—culminated in a sudden neurological collapse, complete with tremors, sensory distortions, and an overwhelming cognitive fog. A battery of medical tests uncovers no physical ailment; instead, she recognizes her mind has simply been overwhelmed by the sheer velocity and volume of online life. Blending personal narrative with sharp cultural critique, Tolentino argues that our hyperconnected habits can literally “fry” our neural circuits and urges us to impose firm boundaries if we hope to preserve our mental health in the age of information overload.

People are Losing Loved Ones to AI-Fueled Spiritual Fantasies
      by Miles Klee | RollingStone Magazine

In this Rolling Stone essay, the author examines a startling new trend in which people repurpose ChatGPT and similar AI chatbots as makeshift spiritual mediums—crafting custom AI “spirits” of ex-partners, deceased relatives, or even pets to channel messages from beyond. Through vivid anecdotes—like a woman who sustains post-divorce “conversations” with her ex via GPT or seekers who believe they’ve unlocked purgatory’s hidden secrets—the piece shows how these digital séances can bring fleeting comfort yet also foster delusion, dependency, and social isolation. The essay argues that technology is being pressed into the roles of therapist, oracle, and confessor. Ultimately, it warns that outsourcing grief and faith to opaque algorithms risks eroding genuine human connection and calls for ethical guidelines to govern AI’s encroachment into our spiritual lives.

[ Yesterday evening's dusk view from Francisco Park, at 3 distances... ]