The Daily Hallucination

SAN FRANCISCO — A researcher has demonstrated that gzip, the file compression program that has shipped free with nearly every computer since 1992, can generate text like a language model, prompting at least four venture firms to attempt to invest in it before learning that it is not a company.

"We moved very fast," said a partner at one firm. "We had a term sheet out within 90 minutes. It was only when we tried to email the founders that we discovered gzip is a command."

The demonstration relies on a long-standing idea that predicting text and compressing it are the same problem. The researcher primed gzip on the complete works of Shakespeare and asked it to continue a prompt. The result was, in his description, "not exactly coherent," and, in the description of a literature professor asked to evaluate it, "about as coherent as act four of most of them."

"It knows something about the text," said Anneke Oyelaran-Whitfield, a researcher in information theory. "It has no weights, no training, no GPUs and no fundraising deck, and it still produced a sentence with the word 'morsel' in it. There are well-funded labs that cannot say that."

A secondary market for gzip exposure had formed by Thursday afternoon, with one fund announcing a $30 billion valuation for what it called "the gzip opportunity," based on the observation that the program is already installed on more than four billion devices and has never once had an outage.

Asked about the program's revenue, the fund's managing partner said it was "pre-revenue in a very deep way."

Maintainers of the program said they were flattered but noted that it had been compressing files for three decades and had never once claimed to understand them.

At press time, an investor had suggested zip.