What Vinge Saw

 

Copyright: Sanjay Basu

Vernor Vinge died in March 2024, a year after GPT-4 was released and six months before OpenAI released the first widely deployed reasoning model. He was 79. He had spent most of his adult life either writing about the arrival of superhuman intelligence or teaching mathematics to people who would help build it. The timing is worth sitting with. He did not miss the singularity by decades. He missed it by weeks.

Most obituaries mentioned the 1993 essay. Some mentioned True Names. A few mentioned Rainbows End. Almost none tried to say what he had actually gotten right, because doing that requires reading four books written across twenty five years and holding them in the same head at once. I want to try.

True Names, 1981

True Names is a novella about hackers who meet in a shared virtual space and hide from a government that wants to find them. The word “cyberspace” would not exist for another year. The dominant reading, then and now, is that Vinge predicted virtual reality.

That reading misses what the novella is actually about. True Names is about the asymmetry between an identifier and a person. A user’s true name is the string that lets an adversary reach past the pseudonym and arrest the body. Everything else is a wrapper.

We now live inside that asymmetry. A phone number, a face embedding, a writing style, a browser fingerprint, a purchase pattern. Any one of these can be a true name if the adversary has enough compute. The industry that Vinge described in 1981 without knowing what to call it is the industry that companies like Clearview AI and Palantir sell into today. The novella predicted the industry more precisely than it predicted the interface.

The 1993 Essay

The Coming Technological Singularity is a NASA paper, not a novel. Vinge presented it at a symposium on the future of space exploration, which is one of the stranger venues for a document that ended up shaping every AI safety conversation of the next thirty years.

The essay lists four routes to superhuman intelligence. First, an AI system that becomes generally more capable than any human. Second, biological enhancement of the human brain. Third, a computer network that becomes conscious as a system. Fourth, cognitive enhancement through human machine interfaces. Vinge did not predict which one would arrive first. He predicted that the first one to arrive would make the others obsolete or trivial within a short interval.

The essay is often summarized as “AI will exceed humans and then things get weird.” That summary loses the interesting part. Vinge’s real claim was about the topology of the transition. Once any one of the four routes lands, the others become downstream effects. The arrival is not a slope. It is a phase change.

Reading the essay in 2026, the striking part is not that Vinge was right about the general direction. It is that he was right about which of the four paths would matter. Route one has arrived. Routes two and four have not. Route three, in the shape he imagined it, was never coherent. The systems that now write code and pass medical boards are not conscious networks. They are large parameterized functions trained on the network’s output. Vinge got the substrate wrong. He got the ordering right.

Rainbows End, 2006

If True Names and the 1993 essay predicted the shape of the transition, Rainbows End predicted the texture of the world during it.

The novel is set in San Diego in 2025. Robert Gu, a poet in his seventies, has been cured of Alzheimer’s by a treatment called You-Gene. He has to go back to high school to learn how to function in a world that has moved past everything he knew. Teenagers wear contact lenses that overlay information on the physical world. People message each other silently by subvocalization. Reading is done by pointing at things. Libraries are being physically shredded and rescanned in a process called the Librareome Project, which nearly everyone in the book has an opinion about. Belief circles let groups of people run consensus reality overlays with their friends and see the world through a shared fictional layer.

I do not need to labor the point. Meta Ray Bans, Apple Vision Pro, AirPods used for silent conversation, the ongoing digitization wars around Google Books and Anna’s Archive, algorithmic feeds that make each user’s phone a private belief circle, Ozempic as a soft You-Gene, adult reskilling anxiety at every professional level. Vinge described the daily granularity of 2026 with a precision that no other science fiction novel of the 2000s matched.

The one thing Rainbows End did not predict was the direction of the intergenerational anxiety. In the novel, teenagers are competent and adults must reskill. In 2026, the picture is muddier. The people best positioned to use frontier AI are not always the youngest. Many of them are the same middle aged engineers who built the previous generation of infrastructure. The reskilling is happening in every direction at once.

Zones of Thought

The point I have been circling. Vinge’s Zones of Thought novels, A Fire Upon the Deep and A Deepness in the Sky, propose that the galaxy is divided into regions where different kinds of cognition are physically possible. In the Slow Zone, only sublight travel and near human intelligence are possible. In the Beyond, faster than light travel and superhuman intelligence become tractable. In the Transcend, minds that are effectively gods can exist.

This is not a plot device. This is a claim about the topology of intelligence.

Intelligence in the Zones of Thought is not a scalar. It is not a number that goes up. It is a function of the substrate. Move the same brain from one region of the galaxy to another and its ceiling changes. There is no fixed thing called “how smart you are.” There is only “how smart you can be given where you are running.”

This is the frame that current AI conversation does not have and desperately needs. We talk about capability benchmarks as if they measure a property of the model. They measure a property of the model on a particular hardware substrate with a particular inference budget with a particular context window with a particular tool loop. Move any of those variables and the number moves. There is no fixed capability of a frontier model. There is only a capability surface across a topology of deployment conditions.

Vinge saw this in 1992. He wrote it as a galaxy. He meant it as a claim about minds.

What He Missed

He missed the political economy. He assumed the transition would be distributed across many actors, most of them scientific or amateur. He did not foresee that four or five companies would end up owning most of the frontier compute. His hackers and academics are still around, but they are not the ones training the models.

He missed the interpretability problem. His superhuman AIs are opaque, but they are opaque in the way a human genius is opaque. Our large models are opaque in a way that no organism ever was. We do not know why they produce the tokens they produce. Vinge assumed that once a system was smart enough to be dangerous, it would be smart enough to explain itself. That has not proven to be true.

He missed how personal the deployment would become. Rainbows End is close but not close enough. The AI in the novel is background infrastructure. The AI in 2026 is a chat window that people use to write eulogies and legal filings and love letters. Vinge did not predict how intimate the technology would be.

Closing

Vinge’s contribution was not the singularity as a slogan. It was a specific claim: that intelligence is a topological property, that the transition to superhuman cognition would be a phase change rather than a slope, and that the daily texture of the world during the transition would be composed of small legible pieces that most people would learn to use without understanding.

He was right on all three. He was writing science fiction. It should have been called something else.