The Last Decade: Our Road to AGI
My current views on the next 10 years of AI
I’ve been reflecting on AI progress and how the next decade might unfold. This is not an opinion piece on what should or will happen, but what I believe is a reasonable scenario for what might happen.
Now: The Augmentation Era
Synchronous AI, Early Labor Shifts
Aggregate AI Demand Limited by # Humans: Most AI interactions remain synchronous (e.g., prompt → result → review). Single threaded human focus means most AI usage is mutually exclusive with other AI usage: either you are consuming Gemini tokens or ChatGPT tokens or Cursor tokens, not all four simultaneously.
Early Displacement: Low-complexity roles in copywriting, graphic design, legal research, and coding are increasingly automated.
Early Human Passthrough Effect: Workers start to defer critical thinking to AI, acting as intermediaries rather than decision-makers.
2025-2030: Autonomous Digital Workforce + Mobility
Substantial Knowledge Worker and Logistics Labor Shifts
Technology Developments:
24/7 Digital Labor: AI agents operate autonomously for longer and longer durations, handling many jobs at near-human salary performance equivalents, but a fraction of the price. Demand for datacenters, energy, and GPUs ramp quickly.
Autonomous Delivery + Driverless Trucking: Zipline drones and Waymo/Tesla/etc dominate logistics for people, meals, and stuff. 4% of the US workforce.
Economic Developments:
Large Company Restructuring Automation Feedback Loop: Automate → restructuring → stock price surges → automate more → repeat.
One Labor Market Scenario: Displaced knowledge workers reconcile with a loss of value in domains where they spent decades acquiring detailed expertise (eg legal, tax, compliance, coding, etc). Big ‘class’ adjustments potentially required where people in upper middle class knowledge worker jobs suddenly need to reevaluate career choices, potentially towards more physical or spatial work.
Alternatively / Hopefully: The overall economy is so supercharged that the value of the complement of these AI agents (humans with judgement in areas where training data is sparse) goes up massively, so new jobs are created even faster than old ones are automated.
Either way, I expect wealth concentration as AI empowers high productivity people, companies, and capital pools to have even more leverage.
⠀Physical Constraints:
Compute Demands: Compute is already at record high levels of demand, even though we’re only in the copilot era and have yet to meaningfully scale autonomous vehicles.
Fab and Energy Bottlenecks: Even if we could run human-level AGI on a single datacenter GPU (unlikely for many years), doubling the size of the human workforce would require billions of GPUs. The current chip industry produces ~5 million GPUs per year, and even if we repurposed the entire new Arizona TSMC facility for producing H100s it would only add ~14M per year (and we wouldn’t have any capacity for iPhone chips then). Fabs takes ~2 years to bring online when moving at max speed (eg in Japan, the Arizona timeline was more like 5 years) . So, it’s going to be many years of energy and fab buildouts to realize the benefits of the technology even after it exists.
2030 - 2040+: AGI and Humanoids
Scaling labor with electricity
Technology Developments:
AGI: By 2030, it is reasonably likely that we will have AI systems that are better than almost all humans at almost all digital tasks. This is not the end of the world (unless they are profoundly misaligned).
Humanoid robots entering mass production: The rapid progress now visible in humanoid robots will take another 5 years to become reliable, safe, fast, and general enough for widespread use. Much like self-driving cars, we will see early deployment success stories over the next 5 years, but it will take time to get to mass rollouts of humanoids doing tasks people do today.
⠀Societal Shifts:
Full Human Passthrough Effect: “What did ChatGPT/Claude/Gemini/etc say to do?” becomes the most common response in any situation where judgement is required, and humans more and more often choose to ‘skip the middleman’ and let the AIs work it out.
Policy by Algorithm, Enforcement by Algorithm, Compliance by Algorithm: Legislation increasingly drafted by AI, votes negotiated by AI, results measured by AI.
Eventually, Maybe: A Golden Age: We master our biology and expand to the stars. The economy shifts into post-abundance, managed by AI itself.
The error bars are very large, and there are many things that could shift this timeline:
Significant trade decoupling or disruption in production of key components (eg chips).
A significant natural disaster or large scale war.
AI-related negative externalities, like a bioterrorism or lab leak.
Unexpected fundamental limitations in model performance (eg failure to solve long horizon planning)
What do you think?