The Electricity of Our Era

Anthropic’s recent Fable 5 release — and the brief supply cutoff that followed — made headlines again. It’s a powerful model, and everyone wants to use it. But Anthropic and the U.S. government have been sending mixed signals: on the one hand, relentless scarcity marketing (it’s unsafe, compare it with GPT, etc.), and on the other, safety pretexts — routing restrictions that prevent users from using the model however they want, and an outright ban on access for non-U.S. nationals.

Anthropic’s “America/democracy first” posture is nothing new. Frankly, I find it deeply repulsive. Let’s not forget: their own company is filled with non-U.S. nationals, many on EB-1 or H-1B visas. These people contributed their intellect and effort, helped create a remarkable model on this soil — only to be walled off by narrow national boundaries. And the company has the audacity to claim this is for some noble ideal. It’s laughable.

But this isn’t meant to be yet another anti-Anthropic piece. The question I want to ask is bigger: what kind of technology should AI be? What should AI, as a whole, bring to humanity?

When AI was just beginning to take off, I found myself wondering: if AI eventually takes over all work, what does the world look like? One scenario: the vast majority are rendered unnecessary, and AI serves only a tiny few. Another: everyone shares in the dividend — we truly break free from drudgery, free our minds, and set out to explore the boundaries of the world and the universe. I don’t want the first future. That color — dark, nauseating — is not one I’m willing to accept.

AI is already widening the gap between people. For a curious child, nothing beats an interactive encyclopedia. In the developed world — the U.S., China, Europe — these resources are at your fingertips. But in other countries, children don’t even have the internet, let alone access to these extraordinarily intelligent “tutors.” How can they possibly compete on the same starting line? And you don’t even have to look that far: for the poor in America or China, AI is a luxury. For someone fighting to make ends meet, a $20 monthly subscription is a decision that takes real deliberation.

On the other side of the coin, LLM development is showing unmistakable signs of diminishing marginal returns. The industry has shifted focus to post-training and vertical domain adaptation. The big labs have essentially abandoned the AGI pursuit, pivoting toward monetizing current LLM capabilities. Small models that run on a single machine can already reach 70–80% of a large model’s performance — exponential compute investment yielding less than linear intelligence gains. Expensive tokens are no longer irreplaceable. Agent technology has found its moat: a well-designed harness can make a weaker model outperform a stronger one. OpenRouter’s fusion technique has demonstrated that combining multiple models for cross-verification can surpass the performance of any single strong model.

Based on these observations, I believe the industry should converge on two principles regarding token supply:

1. Token supply is the electricity of our era. Only tokens that are affordable and stably supplied are good tokens.
2. In token usage, quantity breeds quality.

From these two principles, three implications follow:

1. Agent and harness design must be provider-agnostic. You never know when a model provider might cut off your token supply. Building a harness around a single provider is putting a noose around your own neck.
2. Token cost management will become an increasingly critical discipline — just as electrification brought electricity into every corner of life, tokens will seep into more and more of what we do.
3. The ability to systematically decompose tasks — reducing complexity so that smaller, weaker models can succeed — will become the defining differentiator of agent harness capability.

I genuinely hope everyone gets to share in AI’s dividend. Prometheus brought fire to humanity. He didn’t write: “Greek citizens only.”