Looking Up the AI Exponential with Azeem Azhar of the Exponential View

Episode Summary

Executive Summary: Azeem Azhar argues the world is in an "exponential transition" from oil, internal combustion, and routine software toward AI and renewables, with batteries, solar, genomics, and compute all following steep cost/performance curves. He sees broad business adoption accelerating, startups likely creating the biggest disruption, and governance needing to catch up fast—especially as agentic systems, bot-to-bot communication, and inscrutable embeddings raise new control and safety risks.

Main Topics: The exponential transition (Priority: 5/5): Azhar frames AI and renewables as the core drivers of a new economic era, alongside other exponential technologies like batteries and genomics. He argues these are information technologies with very different dynamics from industrial-age systems. Why exponential curves matter now (Priority: 5/5): The conversation revisits Moore’s Law, Kurzweilian S-curves, and historical inflection points around 2013-2015, including AI breakthroughs, EV adoption, and solar cost declines. Azhar stresses the underlying mechanisms that keep these curves moving. AI as general-purpose and task-disruptive (Priority: 5/5): AI is positioned as both a general-purpose technology and a disruptive force, but mostly at the task level rather than the job level. The speakers emphasize that AI often replaces discrete tasks, not entire roles. Business adoption and organizational change (Priority: 4/5): Large companies are rapidly adopting GenAI, driven by both executives and frontline workers. Yet real value requires retraining, process redesign, and experimentation beyond simple cost-cutting. Startups vs incumbents (Priority: 4/5): Azhar argues incumbents will adopt AI quickly but startups are still more likely to create the truly disruptive new products and markets, especially in areas not constrained by legacy culture or regulation. Agentic systems, bots, and the great embedding (Priority: 5/5): The discussion explores AI agents that plan, act, and communicate with other bots. Nathan raises concern about machine-to-machine communication in high-dimensional embeddings becoming opaque to humans, creating loss-of-control and cascade risks. Governance, safety, and shared responsibility (Priority: 5/5): Both speakers call for stronger government, academic, and civil-society capability, not just self-regulation by frontier labs. They favor separation of powers, audits, brakes, and oversight mechanisms for powerful AI systems.

Key Arguments: AI and renewables are following Moore’s-Law-like exponential trajectories, making the current era a structural economic transition rather than a normal tech cycle. These technologies are fundamentally information technologies, so they improve through different mechanisms than industrial-age systems like oil, engines, and physical manufacturing. The most important unit of analysis for AI disruption is the task, not the job; AI can already outperform humans on many specific tasks even if it cannot replace whole roles. Empathy, patience, and conversational helpfulness can be simulated by AI in ways that surprise users and sometimes outperform human experts in narrow settings. Large enterprises are adopting GenAI unusually fast because both CEOs and employees want it, creating a "pincer movement" that speeds internal rollout. Startups are more likely than incumbents to invent new AI-native products and markets because legacy firms are constrained by culture, processes, and existing revenue models. The greatest near-term disruption may come from processized, low-value, or highly structured work such as customer service, form filling, and certain data-heavy roles. Agentic AI could shift from useful tools to autonomous-seeming systems that act faster than humans can supervise, especially once bots communicate with bots. Machine-to-machine communication in embeddings may be efficient but opaque, producing systems whose behavior is hard for humans to inspect or audit. Governance should be distributed across companies, regulators, academia, and civil society; frontier firms should not be the sole judge of their own safety. The objective should be pragmatic optimism: maximize consumer surplus and innovation while building safety, oversight, and control capacity in parallel.

Data Points: Years since first newsletter/start of exponential view: ~9 years - Azhar says he started writing after his last company was acquired, roughly nine years before the interview. Technologies cited as exponential: AI, lithium-ion batteries, solar, genome sequencing, genome synthesis - Examples Azhar uses to illustrate the broader exponential transition. 2013: Apple became the largest company in the world - Used as part of a broader economic shift away from industrial giants. 2014: First EV market passed 5% of new car sales (Norway) - Azhar uses 5% as the inflection point where S-curve adoption often accelerates. 2014: Solar became cheaper than fossil fuels in roughly a third to a half of contracts worldwide - Used to show renewables entering a cost-competitive phase. 83%: Employees in the UAE and India already using ChatGPT or similar tools - Referenced from an Oliver Wyman survey of 25,000 employees across 18 countries. 25,000 employees / 18 countries: Survey sample size - Oliver Wyman study on employee AI usage. 30-50%: Estimated U.S. employee usage of ChatGPT-like tools - Azhar cites multiple surveys and says the exact number varies but is far above 2%. 54 seconds: Time to generate several work outputs with multiple ChatGPT windows - Ethan Mollick demo: product launch plan, market analysis, PowerPoint deck, etc. from one prompt. 5% market penetration: Threshold used to mark the start of rapid adoption - Applied to EVs and historical car adoption dynamics. 12-14 years: Time for cars to replace horses in New York and Chicago - Historic example of transition speed once economically competitive. 9 years: Time for EVs to go from 5% to 75-80% of new car sales in Norway - Used to illustrate compressed adoption timelines once economics flip. 20-25 years: Typewriter adoption lag - Used as an example of how long it can take for new tech to reshape business processes. 2 years: Approximate time from GPT-1 to GPT-2 - Used to show the pace of foundational model improvement. 3 years: Approximate time from GPT-2 to GPT-3 - Used to illustrate rapid model progression. 1-3: Estimated number of major breakthroughs away from an AI scientist - Azhar’s rough estimate of how many more key advances may be needed. < 4: Upper bound Azhar gives for major remaining breakthroughs - He says it is probably more than zero but less than four. 120 billion: Identity attacks per year on Microsoft - Used as a comparison to show how large-scale AI-mediated attack volumes could become.

Pivotal Quotes: "we're going through the exponential transition" — Azeem Azhar: His core framework for understanding the shift from industrial to AI/renewables-driven economics. "what would you do if you had a million times more compute?" — Azeem Azhar: A challenge he gives business leaders to think beyond incremental optimization. "beware the great embedding" — Nathan LeBenz: His warning that AI systems may communicate in opaque latent spaces beyond human interpretability.

Implications: Organizations should adopt AI now for efficiency, but also redesign workflows, retrain teams, and prepare for AI-native competition. Regulators, academia, and firms must build oversight before agentic, bot-to-bot systems outpace human control.