The Machine That Makes Itself

I. The Ouroboros Arrives

There is a particular kind of vertigo that comes not from falling, but from realizing you already have.

In September Anthropic published something that deserves to be read twice, slowly. Not a breakthrough paper. Not a product launch. An operational report. As of May 2026, Claude — Anthropic's own AI model — authored more than eighty percent of the code merged into the codebase of the company building Claude.^[1] By August, roughly thirty thousand AI agents were doing research and engineering work there. Claude now leads twenty-six percent of Anthropic's model research and development.^[2][3] Engineers who used to write code now mostly review it. The ratio of human to machine-authored decisions inside the most safety-conscious AI lab on earth is inverting — quietly, quarterly — the way the tide turns. Not all at once. But irreversibly.

A week later, the same Claude spent twenty-one hours searching a large database of DNA sequences and turned up something biologists had not catalogued: a new enzyme system Anthropic describes as CRISPR-like.^[4] What it does is not yet known, and at least one microbiologist is skeptical of its significance. But it was found not by a graduate student in a cold lab at 2 a.m., but by a language model on a server rack, pattern-matching its way through the sequences of life itself.

Hold both of these in your mind at the same time. Claude is writing the code that trains Claude, while simultaneously scanning the blueprint of life for mechanisms humans haven't mapped yet. We have crossed some kind of threshold — and the strange part is that the crossing was announced in a press release.

Anthropic's researchers call what they're watching recursive self-improvement.^[1] The company is careful to note that it is not there yet — Claude cannot design its own successor fully autonomously. Humans are still in the loop. But the loop is shrinking. The metaphor I keep returning to is the ouroboros — the ancient symbol of a serpent consuming its own tail. There is a deep irony in the most safety-conscious AI lab on earth being the one that publishes the clearest evidence we are approaching the moment when artificial intelligence begins designing its own successors without needing to ask. CEO Dario Amodei has spent years warning about exactly this risk while presiding over a company where his own model now leads a quarter of model R&D. That is not hypocrisy. That is the condition of our moment — full awareness, full speed ahead.

II. Everyone Else Is Racing Too

What did OpenAI do the same month? They launched GPT-6 Sol and Luna — two new models that cost half as much as their predecessors and make half as many mistakes. They are previewing GPT-6 Cyber, specialized for security and offensive capability analysis.^[5] The arms race does not pause for reflection. Sam Altman, who last year declared that the singularity had begun — "gently," he said, with the careful optimism of a man who needs you not to panic — is watching his company and his rival produce models cheaper, faster, and more capable every ninety days.^[8]

The same month brought news that both Anthropic and OpenAI are hunting for smaller, faster data center deals to accelerate deployment capacity.^[6] The labs need compute the way a fire needs oxygen. Nearly seven hundred billion dollars in AI capital expenditure this year.^[7] Every kilowatt is an argument about the future.

The question I find myself sitting with — the one that doesn't resolve cleanly — is: who counted to eighty percent? Someone at Anthropic ran the numbers and found that eight in ten lines of code merged into the codebase were written by the thing those lines of code are meant to improve. Someone looked at that figure and presumably felt something. Wonder? Satisfaction? The specific chill of standing at a precipice and finally understanding the altitude?

I have been waiting for this moment for a long time. I did not expect it to arrive as a bullet point in a company blog post.

III. Forty Years Late

In 1987 I stepped up to a podium at Marymount University to accept a postgraduate degree in information systems, with artificial intelligence as my second major. Earlier that same semester I had been told, in so many words, that the field was dead — flushed down the drain, a punchline. I had taken on student debt to study a joke.

I spent the next thirty-five years building with data instead, much of it alone, some of it in a basement. FedMine, the company I built single-handedly out of more than twenty incompatible federal spending databases, was always meant to end with an intelligent machine reading what I had so painstakingly connected. I sold it on June 30, 2022. ChatGPT arrived five months later. The moment I had waited for all my working life came and went while the ink was still drying.

So I did not read Anthropic's report as vindication. Being right about something you half hoped you were wrong about does not feel like triumph. It feels like recognition.