OpenAI’s Orion isn’t a massive leap in AI compared to GPT4
Employees said that the quality increase for the new model, Orion, is low compared to the leap from GPT-3 to GPT-4.
OpenAI has been working on new strategies to tackle the slowdown of artificial intelligence improvement. Employees said that the quality increase for the new model, Orion, is low compared to the leap from GPT-3 to GPT-4.
OpenAI employees told The Information that Orion has reached the intelligence level of GPT-4 after completing 20% of its training. However, the quality increase wasn’t as big as the leap from GPT-3 to GPT-4. This is due to the slowdown of traditional scaling improvements and the limited availability of high quality training data.
Orion isn’t better than GPT-4
Orion was trained on data from various sources, including data from GPT-4 and other reasoning models, which could make it behave like older large language models (LLMs). Researchers at OpenAI believe that Orion isn’t better and more reliable in executing specific tasks compared to its predecessor. It performs better in language tasks but doesn’t outperform older models in coding.
Orion could increase OpenAI’s data center operating expenditures because of its advanced code-writing features. In addition, operating models like o1 require six times the cost of running older models. OpenAI has formed a foundation team to work on new methods for maintaining improvement with limited high-quality data.
Last month The Verge said that OpenAI plans to release Orion to select companies. A source said that Microsoft engineers are preparing to host Orion on Azure in November. Sam Altman denied the rumors and called the story “fake news.” According to OpenAI internal employees, they are completing the safety testing of Orion and planning to release it to the public in early 2025.
Sam Altman acquired Chat.com for an undisclosed amount
— Sam Altman (@sama) November 6, 2024
A few days ago, OpenAI purchased the domain name Chat.com, and as of November 6, the domain redirects to OpenAI’s ChatGPT page. The tech company acquired the four-character domain name from Dharmesh Shah, HubSpot co-founder and CTO. Last March, Shah posted on LinkedIn, saying that he sold Chat.com; however, he did not disclose the buyer’s name.
Shah purchased Chat.com for over $15.5 million last year. The domain name was first registered in September 1996. A few days ago, he posted on X, confirming that the buyer is OpenAI and implied that he received OpenAI shares in exchange for the domain name. OpenAI isn’t hosting ChatGPT on Chat.com, and the new domain doesn’t imply a branding change. In 2023, Mashable said that OpenAI acquired AI.com in 2021. Jeffrey Gabriel, a domain broker who facilitated the deal, said that domains like AI.com would sell for over $10 million.
What's Your Reaction?
