Divya van Mahajan

The Power to Shape AI

ai-daily-brief-podcast

Overview

This episode of The AI Daily Brief (dated 2026-03-15) is a “big think” weekend episode hosted by the show’s regular presenter (name not explicitly stated). The central thesis is that despite widespread anxiety and a growing sense of helplessness around AI’s disruptive trajectory, individuals, organisations, and societies retain meaningful agency to shape how AI develops and is deployed. The episode is structured around a close reading of Professor Ethan Mollick’s essay The Shape of the Thing, paired with his earlier 2023 essay The Shape of the Shadow of the Thing, plus the host’s own commentary on the broader discourse around AI disruption.

Source video: No URL provided.

Prerequisites

  • Familiarity with major large language models (GPT-3.5, GPT-4, Claude, Gemini) and their approximate release timelines
  • Basic understanding of AI agents and the distinction between prompt-based (“co-intelligence”) and agentic AI workflows
  • Awareness of key AI companies: OpenAI, Anthropic, Google DeepMind
  • General knowledge of the AI capability debate (benchmarks, the “jagged frontier” concept)
  • Some familiarity with Ethan Mollick’s work and the concept of recursive self-improvement (RSI)

Main Points

Ethan Mollick’s 2023 Essay: Seeing the Shadow

  • Written in October 2023, near the end of what Mollick called the “first phase” of the AI era (beginning with ChatGPT’s launch in November 2022).
  • Mollick predicted that AI would soon function as a personal assistant, intern, and companion—talking, seeing, knowing context, and doing research—because all the component capabilities already existed.
  • He noted that the implications of AI for jobs, education, and society were genuinely unknowable, even to the AI labs themselves.
  • Key framing: AI’s impact is not deterministic—it “is not going to be imposed on us by machines.” Human agency and decisions would determine whether AI empowers or removes power.
  • His prediction that Gemini would outperform GPT-4 did not materialise exactly, but he correctly anticipated a prolonged period of models clustering around GPT-4-level capability throughout 2024.

Mollick’s 2025/2026 Essay: Seeing the Thing Itself

  • Mollick argues we have entered a new phase: from “co-intelligence” (human-AI prompting back and forth) to AI agency (delegating hours of human work to AI systems and managing the outputs).
  • Key enabling developments: agentic systems like Claude Code, OpenAI’s Codex, and OpenClaw.
  • Progress is illustrated by his long-running “otter on a plane” image generation test, which shows exponential improvement from 2022 to 2025, now extending to near-perfect video generation.
  • Benchmark data (including the “meter-long tasks graph” measuring autonomous AI work completion) shows the same exponential improvement curve across multiple evaluation frameworks.

Radical Changes to Work: The Software Factory Example

  • A three-person team at StrongDM built a software factory in which AI agents write, test, and ship production code with no human touching or reviewing the underlying code.
  • Rules: no human writes code; no human reviews code. Each engineer spends the equivalent of their salary (~$1,000/day) on AI compute tokens.
  • Agents take product roadmaps written by humans, build software, simulate customer environments, and loop feedback until results satisfy the AI, then ship.
  • Mollick’s point: the specific details matter less than the fact that such radical organisational experimentation is now possible and likely necessary.

Rolling Disruption: A Volatile Near-Term Environment

  • As AI capability crosses thresholds, it unlocks new use cases overnight, causing sudden market and organisational reactions.
  • A single week in late February illustrated this: a fictional-but-alarming Citrini Research financial crisis scenario; Block announcing large-scale layoffs (later clarified as not purely AI-driven); and a public clash between the Pentagon and Anthropic over Claude’s military use.
  • Mollick’s framing: even if each individual event was misread, collectively they preview what the near future will feel like—rapid capability revelations triggering market reactions, real (if debated) job impacts, and entanglement between AI and geopolitics.

Recursive Self-Improvement (RSI): The Accelerant

  • RSI is the feedback loop in which AI systems are used to build better AI systems, potentially steepening the already-exponential capability curves.
  • Evidence it is no longer theoretical:
    • Dario Amodei (Anthropic) noted at Davos that Anthropic engineers “barely write code themselves anymore.”
    • OpenAI stated its Codex model was “instrumental in creating itself.”
    • Demis Hassabis (Google DeepMind) confirmed closing the self-improvement loop is an explicit goal across all major labs, while acknowledging missing capabilities and real risks.
  • Mollick concludes RSI is now “an explicit item on the roadmap of every major AI company,” not science fiction, though bottlenecks in compute, data, or research difficulty remain possible.

The Host’s Argument: Against Learned Helplessness

  • The host identifies a pattern of feigned or implicit helplessness in mainstream AI discourse—an attitude that because the forces are large, human agency is irrelevant.
  • Example critique: the organisation Alliance for Secure AI and its site jobloss.ai track AI-driven layoffs but offer no policy proposals, no remediation, and no actionable direction—functioning only to amplify fear.
  • The host distinguishes between awareness of disruption (necessary and productive) and learned helplessness (counterproductive).
  • Analogises to Archimedes: large forces can be moved with the right lever; the question is finding the fulcrum, not denying agency.

Individual and Societal Agency in Practice

  • On an individual level: the host cites initiatives like “Claw Camp” (a self-directed programme for building AI agents) with nearly 7,000 participants as evidence that people are choosing to engage actively rather than passively.
  • On a societal level: public controversies (Pentagon vs. Anthropic; Bernie Sanders’ proposed moratorium on AI data centres; Andrew Yang’s proposal to tax AI instead of workers) are expanding the Overton window and creating space for a broader policy conversation.
  • The host is explicitly critical of Sanders’ data centre moratorium as counterproductive, but welcomes the elevation of the conversation itself.
  • Core message: markets and societies are mechanisms for getting people what they need and want—a reminder that agency exists at every level, even if expressed in small ways.

Key Concepts

  • Co-intelligence: The earlier phase of human-AI interaction characterised by iterative prompting, where humans and AI collaborate task by task.
  • AI agency / agentic AI: A newer paradigm in which AI systems autonomously complete extended, multi-step tasks (potentially hours of human work) with humans managing outputs rather than steering each step.
  • Jagged frontier: The uneven landscape of AI capability, where AI is superhuman in some tasks and surprisingly weak in adjacent ones.
  • Recursive self-improvement (RSI): A feedback loop in which AI systems are used to design and train the next generation of AI systems, potentially accelerating capability growth.
  • Software factory: An organisational model in which AI agents autonomously write, test, and ship production code, with humans providing only high-level roadmaps and final approval.
  • Rolling disruption: Mollick’s term for the unpredictable, threshold-driven pattern of AI advances causing sudden and outsized reactions in markets, organisations, and public discourse.
  • Meter-long tasks graph: A benchmark measuring how much autonomous human-equivalent work an AI can reliably complete; used as a key indicator of agentic AI progress.
  • Learned helplessness: A psychological state (applied here socially) in which individuals or groups stop believing their actions can influence outcomes, even when they can.
  • Overton window: The range of ideas considered politically acceptable in mainstream discourse at a given time.
  • SaaSocalypse: The host’s term for the cascading market disruption in software-as-a-service valuations triggered by successive AI capability announcements.

Summary

Drawing on Ethan Mollick’s paired essays from 2023 and 2025/2026, this episode traces the arc from barely being able to see AI’s “shadow” to now being able to discern the “thing itself”—a world of workable agentic AI systems that can autonomously complete hours of human work, combined with exponentially improving capabilities and the early-stage emergence of recursive self-improvement. Mollick and the host both argue that this moment is genuinely historic and genuinely destabilising, pointing to radical organisational experiments like the StrongDM software factory and a single chaotic week in February 2026 as previews of the rolling disruption ahead. However, the episode’s central argument is a rebuttal of the implicit fatalism that pervades much public AI discourse: the host contends that awareness of disruption is necessary and productive, but that organisations like Alliance for Secure AI which catalogue harm without proposing remedies actively cultivate learned helplessness. Both Mollick and the host insist that because AI’s shape is not yet fully fixed—because norms, policies, and organisational models are still being written—the choices made by individuals, companies, and governments right now carry outsized weight, and the window to exercise that agency, while open, should not be squandered.