Little Black's Picky Eating

In California, in a conference room in San Francisco's financial district.

pnA's Adam Kelly hung his coat over the back of a chair.

David Long from l pMn and Simon Hart from Anthropic had been waiting for him for some time.

These three companies had been fighting for years. They had poached each other's employees, stolen each other's clients, and taken shots at one another during product launches.

Yet today, the people in charge of artificial intelligence at all three companies were sitting in the same conference room.

There was only one possible explanation for this.

They had run into a problem none of them wanted to face alone.

It all began after the release of the Unending General-Purpose Large Model.

Before Unending was released,

the market had already priced in far too much of generative AI's commercial value.

After all, most corporate projects were still stuck in the proof-of-concept stage.

Every token they generated was burning money.

And once they entered real business operations, they still had to pay for permissions, audits, and human review.

The employee hours they saved then had to be spent checking whether the models were making things up.

When Unending appeared, the excessively inflated AI bubble finally burst completely.

Alphabt, l's parent company, was the first to see its stock price fall. pnA and Anthropic's enterprise renewal rates, cancellation margins, and secondary-market valuation quotes also began declining.

Because users had learned that large models could, in fact, avoid hallucinations.

From then on, every mistake PT, mn, or Claude made was like advertising Unending.

The three companies had tried expanding their training datasets, using synthetic data, private sampling, and external validators.

They had even cut prices. When they truly could not catch up with Unending, they quietly opened write-back permissions for model self-iteration one after