4.5 billion words a day: 300 apps now run on GPT‑3

Nine months after opening the API, OpenAI says over 300 applications are now built on top of GPT‑3, pulling in everything from customer feedback summaries to entire virtual beings. The scale is staggering. Developers using the platform are collectively generating an average of 4.5 billion words daily, and the company shows no signs of the growth slowing down.

Some of the use cases are genuinely clever. Viable, for instance, hoovers up survey responses, help desk tickets, and live chat logs, then uses GPT‑3 to surface themes and sentiment and spit out plain-English summaries. Ask it what’s annoying your customers about checkout, and you might get back: “Customers are frustrated with the checkout flow because it takes too long to load.” That’s the kind of specific, actionable insight that normally requires a data analyst with too much coffee and not enough time.

Then there’s Fable Studio, which is doing something much weirder. They’ve hooked GPT‑3 into Lucy, the hero of Neil Gaiman’s Wolves in the Walls, turning her into a conversational virtual being who pitched her own movie at Sundance. Fable’s team frames it as blending an artist’s vision with emotional intelligence and AI — a phrase that normally makes me roll my eyes, but the Sundance appearance suggests it’s more than a gimmick.

On the search side, Algolia has integrated GPT‑3 into its Answers product, and the results are hard to ignore. After running the model across 2.1 million news articles, the company says it hit 91% precision or better and answered complex natural-language questions four times more often than BERT. Peter Buffington, a product manager at ABC Australia, noted the system surfaced “very relevant, evergreen content” for queries like “Why does a volcano erupt?” — exactly the kind of long-tail question that makes traditional keyword search fall flat. Alongside the customer stories, OpenAI also rolled out new endpoints for answers, classifications, and enhanced search, plus reiterated its safety review process, presumably to head off critics who still worry about what happens when you let anyone prompt a model this powerful.