The Personal Side of Jevons Paradox in the Age of AI
Something interesting happened back in January/February of 2025. If you were watching Google Trends, you might have seen this sudden, massive spike for the term “Jevons Paradox.” Everyone, all at once, seemed to start talking about it.
So, what is Jevons Paradox? It’s the phenomenon where making a resource more efficient to use doesn’t actually lead to us using less of it. Paradoxically, it often leads to increased consumption of that very resource. Because it’s more efficient and often cheaper or more accessible, more people start using it, or they find more purposes for it, and the total consumption shoots up.
Why the sudden chatter around this specific time, early 2025? A big piece of news hit around January 20, 2025: DeepSeek R1 was released. China being denied access to the most powerful NVIDIA GPUs, they had to fall back on less powerful hardware. The efficiency with which they trained that model took most AI watchers by surprise. NVIDIA’s stock took a nosedive in the couple of days following the release.
But apart from the plate tectonics of AI, there’s a very personal side to this.
Venkatesh Rao put it perfectly in a note:
Good sign that AI Jevons paradox has kicked in properly. I’m working much harder than before AI. It’s now clear that I was lazy and low productivity before because I couldn’t afford serfs/graduate student assistants.
Those “graduate student assistants” he’s referring to are the latest AI models, of course, especially the reasoning variety like O3 and Gemini 2.5 Pro.
Tina He wrote a lovely, insightful piece on her Substack, and she nails this personal pressure.
When your productivity increases, several mechanisms kick in simultaneously: Leisure’s opportunity cost skyrockets. When an hour of work generates what once took days, rest becomes luxury taxed by your own conscience. Every pause carries an invisible price tag that flickers in your peripheral vision… The game theory is unforgiving: when everyone can produce 10x more, the baseline resets, leaving us all running faster just to stay in place.
It’s this feeling that if AI can do so much for you, every moment you’re not directing it, you’re falling behind. The output potential is so high that the cost of leisure feels enormous.
Somewhat ironically, the frontier model labs themselves, perhaps unwittingly, are encouraging this. Take a look at Anthropic’s “Claude Code: Best practices for agentic coding”. They suggest a workflow:
Create 3-4 git checkouts in separate folders. Open each folder in separate terminal tabs. Start Claude in each folder with different tasks. Cycle through to check progress and approve/deny permission requests.
This is literally parallelizing your work with AI agents, requiring you to manage multiple streams. It reminded me of a short-order cook frantically running between pans on the fire.
Question is — why do you need human SWEs to carry out these steps and babysit the AI? Or until when?
The trend towards AI working even when we’re not actively directing it is growing. OpenAI introduced scheduled tasks in ChatGPT a few months ago. The AI can be working tirelessly, even without our direct, real-time input.
The folks from Letta AI (aka MemGPT) took this even further with their paper on “Sleep-time Compute: Beyond Inference Scaling at Test-time:
…allows models to “think” offline about contexts before queries are presented; by anticipating what queries users might ask and pre-computing useful quantities, we can significantly reduce the compute requirements at test-time.
So, the AI is pre-processing, anticipating, constantly churning in the background. If you start doing this, the appetite for hardware, for GPUs, is basically going to be insatiable because there’s no real upper bound on how much you could anticipate and pre-compute.
This all reminds me of a recent podcast with Tyler Cowen where he was asked about the bottlenecks to AI progress. His answer was blunt. The interviewer, Dwarkesh Patel, asked, “What are the specific bottlenecks?”
Humans. Here they are. Bottleneck, bottleneck. Hi. Good to see you. And some of you are terrified. You’re going to be even bigger bottlenecks.
So, where does that leave us? In this world where Jevons Paradox at the personal level just seems to squeeze our time more and more, what’s the play? I love how Tina He framed it in the conclusion of her essay:
This may be the good news for those that didn’t dare to fully lean into what they love and want to do. What if the most game-optimal play in the new system is actually to become relentlessly, unapologetically you?