“The bubble is bursting a little”: Bloomberg learned about OpenAI and Google’s problems with new AI models

According to them, in September OpenAI completed the initial training of a new neural network codenamed Orion. The company expected the model to outperform previous versions, but was unable to achieve the desired performance; For example, Orion was unable to answer programming questions, agency interlocutors say.

ChatGPT developer “is not alone in his problems”: Google has difficulties with AI, as the next version of its Gemini model “did not live up to expectations”, as does Anthropic, which has no time to launch Claude. 3.5 Opus neural network in time, writes Bloomberg.

One of the obstacles companies have faced is that they are finding it increasingly difficult to find new, untapped sources of high-quality data to use to train AI, agency sources said. Furthermore, they say, even “modest improvements” may not justify the “enormous costs” of developing new models.

The failures of OpenAI, Google and Anthropic could affect the development of artificial general intelligence, which is expected to equal or surpass humans in many intellectual tasks, Bloomberg writes. “The AGI bubble is bursting a little bit,” Margaret Mitchell, chief ethics officer at artificial intelligence startup Hugging Face, told the agency.

Previously, sources at The Information reported on a slowdown in innovation at OpenAI. They noted that Orion’s new neural network is barely superior to its GPT-4 predecessor, as well as being less easy to program and may be more expensive to use in data centers.

In late October, The Verge’s interlocutors said that the creator of ChatGPT was preparing to launch a new AI model in December, which the company’s director, Sam Altman, called fake. According to Bloomberg, OpenAI is unlikely to launch Orion until early 2025 as it is “not yet at the level the company would like.”