This is Atlantic Intelligence, a newsletter in which our writers help you wrap your mind around artificial intelligence and a new machine age. Did someone forward you this newsletter? Sign up here. |
|
Senior editor, technology |
|
|
OpenAI’s much-awaited new model, GPT-5, launched yesterday, and it’s already exhibited some embarrassing problems. In our tests, it has failed to find the correct answer to a basic arithmetic problem and to create a labeled map of Italy. Others on social media have pointed out that GPT-5 fails to correctly count the number of times the letter b appears in the word blueberry—a frequent problem for large language models, due in part to the way they process language as a series of tokens rather than individual letters or words. But as my colleague Matteo Wong argues in his latest story for The Atlantic, the overall “intelligence” of GPT-5 (or lack thereof) is somewhat beside the point. The glitches are not great (although the model performs admirably on a number of industry benchmarking tests), but OpenAI is now focused on developing the user experience for ChatGPT, prioritizing the product’s design and features such as personalization above all else. “As OpenAI has ascended to the scale of a typical tech giant—it is reportedly in talks for a $500 billion valuation—the firm has also started to act like its corporate rivals,” Matteo writes. That means building products that lure people into an ecosystem, like the iOS or Android platforms you’re likely reading this on, and keeping them there for the long haul. |
|
| | (Illustration by The Atlantic. Source: Getty.) | | | |
|
|
| GPT-5 achieves state-of-the-art performance on a number of AI benchmarks, according to OpenAI’s internal tests, but it is far from a clean sweep: On a few tests, competing products such as Google Gemini, Anthropic’s Claude, and xAI’s Grok outperform, or are just barely below the level of, OpenAI’s new top model. The GPT-5 announcement video and launch page also contained a number of errors—incorrect labels, numbers and colors that made no sense, and missing entries on charts—that made the program’s precise abilities, and the trustworthiness of OpenAI’s reporting, hard to discern (and led some observers to joke that perhaps GPT-5 itself had made, or hallucinated, the graphics). Yet that may not matter. OpenAI’s animating theme for GPT-5 is user experience, not “intelligence”: Its new model is intuitive, fast, and efficient; adapts to human preferences and intentions; and is easy to personalize. Before it is more intelligent, GPT-5 is more usable—and more likely to attract and retain users. “The important point is this,” Altman said, pinching a thumb and index finger together for emphasis: “We think you will love using GPT-5 much more than any previous AI.” | |
|
|
|
|
- Generative AI’s iPhone moment: “Whether the generative-AI race prompts genuine societal transformation or simply provides a new profit model, its winners will be whichever tech companies best execute Silicon Valley’s decades-old playbook,” Matteo wrote back in 2023.
- “The AI birthday letter that blew me away”: “Google is ushering in an era of custom chatbots,” Lila Shroff writes.
|
|
Most Popular on The Atlantic | |
|
|
Join Arthur C. Brooks and Jeffrey Goldberg for a subscriber-only event | | | Want to learn more about leading a life that feels full and meaningful? Join Arthur C. Brooks and The Atlantic’s editor in chief, Jeffrey Goldberg, on Monday, August 11, at 2:30 p.m. ET as they discuss Brooks’s new book, The Happiness Files: Insights on Work and Life. Add the event to your calendar today. | |
|
|
|
