Monthly Archives: March 2023

Currents 088: Melanie Mitchell on AI Measurement and Understanding

Jim talks with Melanie Mitchell about her critique of applying standardized exams to LLMs and the debate over understanding in AI. They discuss ChatGPT and GPT-4’s performance on standardized exams, questioning the underlying assumptions, OpenAI’s lack of transparency, soon-to-be-released open-source LLMs, prompt engineering, making GPT its own skyhook to reduce hallucinations, the number of parameters in GPT-4, why LLMs should be probed differently than humans, how LLMs lie differently than humans, Stanford’s holistic assessment for LLMs, a College Board for LLMs, why the term “understanding” is overstressed today, consciousness vs intelligence, the human drive for compression, working memory limitations as the secret to human intellectual abilities, episodic memory, embodied emotions, the idea that AIs don’t care, calling for a new science of intelligence, the effects of differing evolutionary pressures, whether a model of physics could emerge from language learning, how little we understand these systems, and much more.

Melanie Mitchell is Professor of Computer Science at Portland State University, and External Professor and Co-Chair of the Science Board at the Santa Fe Institute. Mitchell has also held faculty or professional positions at the University of Michigan, Los Alamos National Laboratory, and the OGI School of Science and Engineering. She is the author or editor of seven books and numerous scholarly papers in the fields of artificial intelligence, cognitive science, and complex systems, including her latest, Artificial Intelligence: A Guide for Thinking Humans.

EP 182 Brad DeLong on An Economic History of the 20th Century

Jim talks with Brad DeLong about his book Slouching Toward Utopia: An Economic History of the Twentieth Century. They discuss how everything changed around 1870, the idea of a polycrisis, Friedrich von Hayek’s affirmation of the market system, the calculation problem, Karl Polanyi’s response, a quantitative index of technological knowledge, the pace of growth, the necessity of a grand narrative, Malthusianism, the lead-up to the Industrial Revolution, the invention of the industrial research lab, the Edison-Tesla fight, science as an institution, the transition away from force & fraud dominance, theories about the rise of global empires, communities of engineering practice, causes of World War I, Max Weber’s German chauvinism, 30 glorious years of social democracy, the Macintosh launch commercial & the neoliberal turn, the evaporation of cultural conservatism, the liminal age, and much more.
Brad DeLong is a professor of economics at the University of California at Berkeley. He was a deputy assistant secretary for economic policy at the U.S. Treasury during the Clinton Administration. He is a New York Times instant bestselling author, for Slouching Towards Utopia: An Economic History of the Twentieth Century, which was called: “magisterial” by Paul Krugman, “required reading” by Larry Summers, “immense scope and depth” by Diane Coyle, and “impressive… written with wit and style and a formidable command of detail” by Ryan Avent. He has been too online since 1995, now in the form of a SubStack, formerly at TypePad.

Currents 087: Shivanshu Purohit on Open-Source Generative AI

Jim talks with Shivanshu Purohit about the world of open-source AI models and a significant open-source LLM coming soon from Stability AI and EleutherAI. They discuss the reasons for creating open-source models, the release of Facebook’s LLaMA model, the black box nature of current models, the scientific mystery of how they really work, an opportunity for liberal arts majors, OpenAI’s new plugin architecture, the analogy of the PC business around 1981, creating GPT-Neo & GPT-NeoX, the balance between data & architecture, the number of parameters in GPT-4, order of training’s non-effect on memorization, phase changes due to scaling, Stability AI and EleutherAI’s new collaboration & its specs, tradeoffs in price & size, the question of guardrails, reinforcement learning from human feedback, the missing economic model of generative AI, necessary hardware for the new suite, OpenAI’s decreasing openness, Jim’s commitment to help fund an open-source reinforcement learning dataset, the status of GPT-5 & other coming developments, and much more.

Shivanshu Purohit is head of engineering at Eleuther AI and a research engineer at Stability AI, the creators of Stable Diffusion.

EP 181 Forrest Landry Part 1: AI Risk

Forrest Landry
Jim talks with recurring guest Forrest Landry about his arguments that continued AI development poses certain catastrophic risk to humanity. They discuss AI versus advanced planning systems (APS), the release of GPT-4, emergent intelligence from modest components, whether deep learning alone will produce AGI, Rice’s theorem & the impossibility of predicting alignment, the likelihood that humans try to generalize AI, why the upside of AGI is an illusion, agency vs intelligence, instrumental convergence, implicit agency, deterministic chaos, theories of physics as theories of measurement, the relationship between human desire and AI tools, an analogy with human-animal relations, recognizing & avoiding multipolar traps, an environment increasingly hostile to humans, technology & toxicity, short-term vs long-term risks, why there’s so much disagreement about AI risk, the substrate needs hypothesis, an inexorable long-term convergence process, why the only solution is avoiding the cycle, a boiling frog scenario, the displacement of humans, the necessity of understanding evolution, economic decoupling, non-transactional choices, the Forward Great Filter answer to the Fermi paradox, and much more.

Forrest Landry is a philosopher, writer, researcher, scientist, engineer, craftsman, and teacher focused on metaphysics, the manner in which software applications, tools, and techniques influence the design and management of very large scale complex systems, and the thriving of all forms of life on this planet. Forrest is also the founder and CEO of Magic Flight, a third-generation master woodworker who found that he had a unique set of skills in large-scale software systems design. Which led him to work in the production of several federal classified and unclassified systems, including various FBI investigative projects, TSC, IDW, DARPA, the Library of Congress Congressional Records System, and many others.

Currents 086: Monica Anderson on Bubble City

Jim talks with Monica Anderson about her paper “Bubble City Design Proposal: A Twitter Alternative Which Is Not a Social Medium.” They discuss the origins of the Bubble City idea, its architecture, quenching the flood of social media information, only seeing the messages you want, research bots, the difference between a bubble and a Slack channel, fine-tuning bubbles, law enforcement, filtering, the place of curators, federating feeds into the system, how the system supports itself financially, how identity is handled, viscosity, the Pacer speed control, the clickbait problem, trusted streams, Google Wave, how LLMs are changing programming, version changes to Bubble City, Understanding Machine One, a call for fundraising, and much more.

Monica Anderson is an independent AI researcher and ex-Googler operating from Silicon Valley. Her company Syntience, Inc. has researched computer-based Natural Language Understanding since Jan 1, 2001.

Currents 085: Jonny Miller on Self-Unfoldment

Jim has a wide-ranging conversation with Jonny Miller about self-development and emotional resilience. They discuss being a natural human, self-help as deconditioning, self-unfoldment, ecologies of practices, giving power back to the individual, Jamie Wheal’s hedonic engineering, pushing outside the window of tolerance, emotional anti-fragility, facilitated breath repatterning, affirming anger, principles of decision-making, decision paralysis, self-destructive patterns in relationships, common barriers to communication, surrendering to grief, conditions of play, preserving unscheduled time, critiquing “mental health,” the importance & decline of friendship, sparring in schools, the resistance to unproductive activity, video games & disembodiment, the Nervous System Mastery course, and much more.

Jonny Miller is a Nervous System Specialist and host of the Curious Humans podcast. He’s spent cumulatively thousands of hours researching, training & mentoring high-performers and professionals — from the CEO of a rocket ship company to startup founders recovering from burnout as well as busy parents, early-stage solopreneurs & school-teachers.

Currents 084: Mirta Galesic on Global Collective Behavior

Jim talks with Mirta Galesic about the ideas in her co-authored paper “Stewardship of Global Collective Behavior.” They discuss the meaning of collective behavior, a crisis in network structures, the analogy of the printing press, consequences of person-to-person communication, the capacity for collective forgetting, unpredictable developments in chatbots, bottom-up vs top-down influence, advertising-driven information ecosystems, emergent knobs in social media design, ChatGPT’s political bias, the widespread trust in algorithms, suggestions for reforming Twitter, information decay, viscosity, opportunities & dangers of mass surveillance data, the Twitter Files, free speech & cultural evolution, and much more.

Mirta Galesic is a Professor at the Santa Fe Institute and External Faculty at the Complexity Science Hub in Vienna, Austria, as well as the Vermont Complex Systems Center, UVM. She is also an Associate Researcher at the Harding Center for Risk Literacy and a non-resident system thinking expert at the United States Institute of Peace. She studies how simple cognitive mechanisms interact with social and physical environments to produce seemingly complex social phenomena. Her projects focus on developing empirically grounded computational models of social judgments, social learning, collective problem solving, and opinion dynamics. She is also interested in how people understand and cope with the uncertainty and complexity inherent in many everyday decisions.

EP 180 Lynne Kiesling on the Electrical Grid

Jim talks with Lynne Kiesling about the electrical grid and what could and should change in its architecture in the years to come. They discuss electricity as a product, the move away from centralized control rooms, energy storage as the holy grail, base load vs peak load, distributed & intermittent energy resources, moving power to & from the grid, temporal patterns of supply & usage, varying demand to meet supply, programming thermostats, digitization of the electric grid, how rooftop solar systems coordinate with the grid, distributed energy resource management systems, advancements in storage, cyberattacks & solar flares, the Transactive Energy Service System (TESS), machine learning in energy bidding, the challenge of testing complex systems, the Olympic Peninsula Testbed Project, responding to events like the Great Texas Freeze of 2021, institutional design in a new technological landscape, wholesale power generation, power law distributions, and much more.

Lynne Kiesling is an economist focusing on regulation, market design, and the economics of digitization and smart grid technologies in the electricity industry. She is a Research Professor in the School of Engineering, Design and Computing at the University of Colorado-Denver, and Co-Director of the Institute for Regulatory Law & Economics. Lynne also provides advisory and analytical services as the President of Knowledge Problem LLC, and is an Adjunct Professor in the Masters of Science in Energy and Sustainability program at Northwestern University. In addition to her academic research, she is currently a member of the U.S. Department of Energy’s Electricity Advisory Committee, has served as a member of the National Institute of Standards and Technology’s Smart Grid Advisory Committee, and is an emerita member of the GridWise Architecture Council. Her academic background includes a B.S. in Economics from Miami University (Ohio) and a Ph.D. in Economics from Northwestern University.