“The likelihood of getting to ASI is getting closer every month,” said Logan Kilpatrick, head of product at Google AI Studio. He cited the way former OpenAI chief scientist Ilya Sutskever jumped right into ASI development without an intermediate model, and pointed to “test-time compute” as a key element that made it possible.
|
▲[Korean Today] Photo = Google © Reporter Byun A-rong |
“Iteratively scaling test-time compute without going through intermediate models is more effective,” Kilpatrick said, citing Sootskever’s use of this approach to develop ASI as a good example. This means that it is more effective to rapidly improve performance after a single pre-training run than to release models incrementally.
Google also sympathizes with this approach and seems to have reflected it in the development of the Gemini AI model. Since November, Google has been releasing Gemini test versions on Chatbot Arena at a rate of one per week, showing off its performance improvements. This is consistent with Kilpatrick’s statement that the possibility of ASI development is increasing every month.
Kilpatrick also predicted that artificial general intelligence (AGI) will soon appear in product form. He said, “As companies compete to release new models, it is possible that some of them have already achieved AGI.” This is in line with the remarks of Sam Altman, CEO of OpenAI.
Kilpatrick has previously worked on the development of Qstar at OpenAI, and it is likely that he applied this experience to the development of Gemini at Google. He is currently in charge of Google’s Gemini API service and AI Studio products, and is considered a key figure in AI development.
As competition among major AI companies, including Google and OpenAI, intensifies, attention is focused on whether the development of ASI and AGI can open new horizons in AI technology.
© The Korean Today – All Rights Reserved