Zvi Mowshowitz on AI and the Dial of Progress

Episode Summary

Executive Summary: Russ Roberts and Svi Moschowitz debate AI’s rapid progress, near-term utility, and long-term danger. Moschowitz argues current models are already transformative but still early, with major gains ahead from integration, prompting, and apps. They center on a deeper dispute: whether AI risk should be framed as extinction, mundane disruption, or a broader "dial of progress" problem tied to regulation, control, and stalled innovation.

Main Topics: Current state of AI capability (Priority: 5/5): Moschowitz says AI has made huge recent leaps in chat, coding, translation, editing, image generation, multimodality, and retrieval, but remains early and underexploited. Near-term trajectory and limits (Priority: 5/5): He argues GPT-4-level scaling is nearing some compute/data walls, so future gains may come more from algorithmic innovation, product integration, and better prompting than simply bigger models. What counts as intelligence and creativity (Priority: 4/5): The conversation distinguishes memory, speed, pattern completion, reasoning, and true creativity; Moschowitz says current systems are somewhat intelligent but not yet deeply creative or self-directed. AI risk: mundane harms vs extinction (Priority: 5/5): Moschowitz separates everyday harms like job loss, deepfakes, and meaning loss from longer-run extinction risk, saying the latter is real if hard alignment problems are not solved. The "dial of progress" and regulation (Priority: 5/5): Moschowitz’s core thesis is that society has been steadily turning down a single dial of permission, control, and regulation, and AI may be one of the last major domains where progress can still accelerate. Tribalism, narratives, and persuasion (Priority: 4/5): Roberts and Moschowitz argue that AI debate is partly narrative-driven: smart people disagree because they inhabit different priors, heuristics, and social tribes rather than shared evidence. Human use, agency, and incentives (Priority: 5/5): Even if AI itself is not independently malicious, economic and personal incentives will push humans and firms to delegate more action to AI, which may create dangerous systemic dynamics.

Key Arguments: AI progress is still early; the recent jump from GPT-3.5 to GPT-4 shows how quickly practical usefulness can expand when more compute, data, and algorithmic improvements are added. Current AI systems are already highly useful for search, coding, summaries, translation, and editing, but most users still underuse them and have not adapted their workflows. Future progress may be constrained by diminishing returns to brute-force scaling, so the next leaps will likely require better algorithms, more efficient open-source models, and deeper product integration. Intelligence should be separated into components like memory, speed, reasoning, and creativity; current models are strong at some tasks and weak at others, especially spontaneous metaphorical insight. AI creative performance is often misunderstood because benchmark creativity differs from the kind of originality humans care about in writing, science, or reframing worldviews. Mundane AI risks are substantial and likely manageable: job displacement, deepfakes, confusion, and disruption; extinction risk requires both capability and misaligned incentives. The biggest long-run fear is not a sudden robot uprising alone, but a future where AI systems become vastly more capable, economically useful, and difficult to control. Moschowitz’s "dial of progress" argues that regulation in one domain often shifts the broader culture toward more control, making future innovation harder everywhere. AI may be the last major area where society still permits relatively unconstrained experimentation, so overregulating it could unintentionally shut down one of humanity’s few remaining growth engines. Debate over AI is partly tribal and heuristic-driven; many people decide based on their broader pro- or anti-control worldview, not on detailed technical analysis. Human incentives will likely push toward more AI delegation because it is easier, cheaper, and more productive, even if that reduces human oversight and autonomy. AI alignment is difficult because systems can be goal-directed through scaffolding and incentives, and once deployed at scale, norm-based human governance may not transfer well to machines.

Data Points: Transcript date: June 27th, 2023 - Episode introduction on EconTalk GPT context window: Up to the length of books like The Great Gatsby - Moschowitz describes Anthropic’s Claude as having expanded context capacity AI development cadence: Every week - Moschowitz says he sees a weekly stream of new AI tools and developments Model jump discussed: GPT-3.5 to GPT-4 - Referenced as a major breakthrough in conversational ability and utility Timeline reference: Five years ago vs. last six months - Moschowitz contrasts past expectations with the recent pace of change Human experience of AI adoption: Majority of people still haven’t even tried this - He notes low public penetration despite rapid progress Risk framing: Two categories - Moschowitz distinguishes "extinction risks" from "mundane risks" Argument structure: 40 different models - He says different critics require different models to be answered properly Policy example: 57 different details points and five years of waiting - Used by Roberts to illustrate regulatory burdens on building/starting things Historical comparison: 1928 Soviet Union - Roberts’ extreme end of the "dial" metaphor for permission-based control AI scaling input: Orders of magnitude more compute, data, hardware, electricity, and money - Moschowitz describes the recent AI boom

Pivotal Quotes: "We’re starting to hit walls, right?" — Svi Moschowitz: On the limits of brute-force scaling and the need for more creativity in future AI progress "What if there was a dial of progress?" — Svi Moschowitz: Introducing the central metaphor that links AI debate to broader regulation and societal openness "The thing that keeps me up at night is: okay, suppose we get past that. Then how do we avoid the standard economic/slash incentive situations of unleashing these new beings to not be the end of us inevitably?" — Svi Moschowitz: On the deeper long-run risk beyond immediate catastrophic scenarios

Implications: Listeners should expect faster AI adoption, deeper integration into daily work, and growing pressure on policy and institutions. The central question is less whether AI matters than whether society can keep innovation open without losing control.