Google’s artificial intelligence (AI) product manager also said that the opportunity of achieving ‘super artificial intelligence (ASI)’ is increasing. He also pointed to ‘test-time compute’, which expands AI reasoning capabilities, as the important thing.
On the thirty first (local time), Logan Kilpatrick, head of Google AI Studio product, posted on Ilya refers back to the former top scientist at Sützkever, who developed ‘Strawberry’, or ‘QStar’, the premise of the inference model in OpenAI.
He then cited for instance the case where former top scientist Sutskever founded SSI and immediately began developing ASI with none intermediate processes. “As a substitute of coming up with an intermediate model, Ilia went straight to constructing ASI,” he identified.
Regarding this, he said, “Many individuals, including myself, initially had doubts about Ilya’s approach to gaining momentum by constructing an intermediate model and directly difficult ASI construction without going through the means of widening the gap with other corporations.”
Nevertheless, he explained, “Ilya can have realized this early on, nevertheless it appears to be example that simply repeating the ‘test-time compute’ extension is simpler.”
In other words, it is mostly faster to repeat test-time computation after pre-training once, somewhat than regularly releasing models with improved performance through pre-training over several years. The reason is that the founding father of Sutskever SSI also adopted this method to construct ASI in a single go.
This could be interpreted to mean that Director Kilpatrick agrees with this method, and that Google is already experiencing this as well.
Artificial general intelligence (AGI) is predicted to look in product form within the near future. Specifically, as models with cutting-edge performance appear one after one other, it is predicted that almost all people is not going to notice this. This doesn’t mean that “that is AGI” will appear sooner or later, but that there’s a high probability that models which have already achieved AGI will probably be partially included within the means of competitively launching models at Open AI, Google, or other corporations.
This is comparable to what OpenAI CEO Sam Altman or OpenAI employees recently revealed. It’s claimed that the event of AGI is just across the corner, or that it is feasible that AGI has already been achieved. Some even said that it’s just that most of the people doesn’t really notice it.
Specifically, Director Kilpatrick is a key figure whom Google considers to be the ‘secret weapon in AI development.’ He left his position as developer relations manager at OpenAI last March and joined Google, and is in command of ‘Google AI Studio’ products and ‘Gemini’ API service.
He also participated in the event of Q-Star at OpenAI. And it is feasible that this was applied by Google.
Since last October, Google has also reported that the performance of Gemini 2.0 just isn’t meeting expectations and that it’s specializing in inference to resolve this problem. Nevertheless, since last November, Gemini test versions have been released once every week on Chatbot Arena, showing off gradual performance improvements.
In other words, the improved performance of those test versions could mean that, as Kilpatrick says, “the likelihood of going straight to ASI is increasing every month.”
Nevertheless, he didn’t reply to X users’ requests for further explanation.
Reporter Lim Da-jun ydj@aitimes.com
