Divya van Mahajan

The Era of AI Mass Intelligence Arrives

ai-daily-brief-podcast

The Era of AI Mass Intelligence Arrives

Overview

This episode of the AI Daily Brief (recorded around Labor Day weekend, late August/early September 2025) uses Professor Ethan Mollick’s essay “Mass Intelligence” (published on his One Useful Thing blog at oneusefulthing.org) as a framing device to reflect on the defining AI trends of summer 2025. The host argues that the most important story of this era is not incremental model capability improvements, but the dramatic democratization of access to powerful AI driven by collapsing costs and improving interfaces. No external speaker is featured; the host reads and annotates Mollick’s essay before offering his own analysis.

Source video: URL not provided (AI Daily Brief, published ~2025-08-31)

Prerequisites

  • Familiarity with major AI model families: GPT-4, GPT-4.5, GPT-5, O1/O3 (OpenAI); Gemini 2.5 (Google); DeepSeek R1
  • Basic understanding of the distinction between standard LLMs and reasoning/chain-of-thought models
  • Awareness of AI image generation tools and their general capabilities
  • General knowledge of enterprise AI adoption patterns
  • Familiarity with token-based pricing in LLM APIs
  • Understanding of benchmarks such as AIME (math olympiad) and preference-based evaluations (e.g., LMArena)

Main Points

The Arrival of Mass Intelligence (Mollick’s Core Thesis)

  • More than one billion people now use AI chatbots regularly; ChatGPT alone has over 700 million weekly users
  • Until recently, free users were limited to older, weaker models; the best reasoning models required $20–$200/month subscriptions
  • Two historical barriers to powerful AI access: confusion (users did not know which model to select) and cost (best models were paywalled)
  • Less than 7% of paying ChatGPT customers regularly selected the O3 reasoning model prior to GPT-5’s launch

GPT-5 as a Democratization Mechanism

  • GPT-5 is both a model family (from GPT-5 Nano to GPT-5 Pro) and a router that automatically allocates the appropriate model tier to a given task
  • Within days of GPT-5’s launch, the share of paying users who had used a reasoner rose from 7% to 24%; free users accessing top models rose from near zero to 7%
  • The rollout was messy because the routing logic was not well explained and initially performed inconsistently, producing uneven output quality across users

The Economics of Cost Collapse

  • GPT-4 at launch cost ~$50 per million tokens; GPT-5 Nano now costs ~$0.14 per million tokens — approximately 1/300th of the original cost for a more capable model
  • Google reported a 33x improvement in energy efficiency per prompt in the last year alone
  • Current energy cost of a standard LLM prompt: ~0.0003 kWh, equivalent to 8–10 seconds of Netflix streaming or a 2008 Google search
  • Collapsing marginal costs make new business models (e.g., ad-supported free AI) economically viable and allow free users to run prompts that would have cost dollars two years ago

NanoBanana and the Ease-of-Use Shift

  • Google’s Gemini 2.5 Flash Image Generator (codename “NanoBanana”) demonstrates that instruction-following in plain language has improved dramatically for image editing
  • Mollick’s demonstration: uploading an Apollo 11 photo and a tuxedo image, then prompting “Dress Neil Armstrong on the left in this tuxedo” produced a realistic composite
  • A second, more complex prompt (Armstrong playing trumpet, Aldrin holding a hamburger, an otter using a laptop in the middle seat on an airplane) also produced a coherent output
  • The model is better at editing than pure generation, but accessible to free users and does not require specialized prompting knowledge

The Societal Implications of Mass Intelligence

  • Every institution — schools, hospitals, courts, companies, governments — was built for a world where intelligence was scarce and expensive
  • With a billion users, AI is simultaneously being used to combat loneliness, diagnose disease, write obituaries, cheat on homework, launch businesses, and create religious texts
  • Key open questions: How do institutions rebuild trust when fabrication is trivially easy? How is human expertise preserved while knowledge is democratized? How is the chaos of billion-user deployment managed?
  • Less restrictive image generators approaching NanoBanana quality are expected within months, with fewer guardrails

The Host’s Analytical Framework: Beyond Benchmark Comparisons

  • The host argues that obsessing over “how much better is GPT-5 than GPT-4.5” is a fundamentally limited lens for understanding AI progress
  • Two under-appreciated dimensions of progress:
    1. Cost reduction — enabling mass deployment and new economic models at scale
    2. Use-case unlocking — small capability improvements that cross thresholds enabling entirely new applications (e.g., NanoBanana’s editing capability replacing years of Photoshop skill acquisition)
  • Current enterprise data (Menlo Enterprise Update): performance beats cost as a model-switching driver 61% to 36% — but the host argues this reflects the current early-deployment phase, not long-term patterns
  • Google’s monthly token processing grew from $480 trillion to $980 trillion between May and July 2025 (104% in two months), attributed largely to autonomous background coding agents — a signal of shifting usage patterns

The “Unlock Score” Proposal

  • The host proposes a new benchmark concept called the Unlock Score
  • Rather than academic benchmarks (gameable, increasingly saturated) or subjective preference rankings, the Unlock Score would measure:
    • What new use cases a model unlocks
    • How economically and socially valuable those use cases are
    • How widespread adoption of those use cases becomes
  • Intended as a comparative metric against a model’s most proximate predecessors, not an absolute capability score

Key Concepts

  • Mass Intelligence: The era in which powerful AI is as broadly accessible as a web search, available to over a billion users regardless of technical sophistication or ability to pay
  • Reasoning Model / Reasoner: An AI model that performs chain-of-thought processing to solve hard problems with lower hallucination rates; examples include OpenAI’s O1/O3 and DeepSeek R1
  • Router (GPT-5 context): A model-layer system that automatically selects the appropriate sub-model or compute allocation based on the complexity of a user’s request
  • Token: The fundamental unit of LLM input/output, used as the basis for pricing and consumption measurement
  • NanoBanana: Codename for Google’s Gemini 2.5 Flash Image Generator, notable for plain-language instruction-following and free-tier availability
  • Unlock Score: A proposed benchmark measuring the new use cases, their value, and their adoption breadth that a model enables relative to its predecessors
  • Autonomous Background Agents: AI systems operating independently on tasks (especially coding) while human users are occupied elsewhere — a usage pattern distinct from interactive co-pilot usage
  • DeepSeek Moment: The January 2025 event when DeepSeek R1 topped app store charts, exposing a mass audience to reasoning models for the first time and prompting OpenAI to reconsider access restrictions

Summary

The central argument of this episode is that summer 2025 marks the beginning of a qualitatively new era in AI — the era of mass intelligence — defined not by another incremental leap in benchmark scores but by the convergence of dramatically lower costs, improved interfaces, and broader access to genuinely powerful models. Drawing on Ethan Mollick’s essay, the host illustrates how the removal of two longstanding barriers (confusion over model selection and cost-based paywalling) has brought over a billion people into contact with tools previously available only to informed paying subscribers. The host extends Mollick’s analysis to argue that the dominant industry habit of evaluating AI progress through numbered-version comparisons (GPT-4 vs. GPT-5) obscures two more consequential dimensions: the ~300x cost reduction that enables entirely new deployment economics, and the threshold-crossing capability improvements that unlock use cases — like accessible professional-grade image editing — that were previously impractical for most people. With token consumption doubling in two months and autonomous agents beginning to operate at scale in enterprise environments, the host contends that the defining AI story of 2025 is not which model won the benchmark race, but what happens when a billion people gain access to unprecedented cognitive tools simultaneously.