AI startups like OpenAI and Anthropic and established tech giants like Google alike are struggling to create more advanced AI, sources with knowledge of the matter have told Bloomberg.
ChatGPT creator OpenAI’s upcoming flagship AI model codenamed ‘Orion’ has failed to live up to hopes it would outpace earlier models, sources told the publication.
Despite completing an initial round of training in September, Orion has not met expectations. Sources revealed that Orion failed to perform well in tasks such as answering coding questions it hadn’t been specifically trained on, and it has not represented as significant a leap forward as OpenAI's GPT-4 was over GPT-3.5.
This setback means OpenAI is unlikely to roll out Orion to its users until early next year, one source said.
Meanwhile, Google’s next Gemini iteration, meant to be a major upgrade, has also underperformed relative to internal expectations, according to three sources with knowledge of the matter.
Similarly, the release of Anthropic’s long-awaited Claude 3.5 Opus model has been delayed.
Challenges facing AI developers include rising costs for the computing power and time required to train new models and a lack of high-quality, human-made training data.
Two sources told Bloomberg Orion’s poor coding performance was due to a lack of sufficient coding data to train on.