April 25thMilletHead of the Large Model TeamRooseveltIN AN INTERVIEW WITH "LANGUAGE IS THE WORLD", ABOUT THE CURRENT BIG MODEL COMPETITION PATTERNS, THE AI SMART BODY PARADIGM SHIFT, AND AGI The process gives a series of specific judgments。
I don't knowAGI WILL BE ACHIEVED IN TWO YEARSI DON'T KNOW. SHE ESTIMATED THAT THE CURRENT PROGRESS WAS ABOUT 201 TP3T AND THAT THIS YEAR IT IS EXPECTED TO MOVE TO 601 TP3T TO 701 TP3T。
According to Roosevelt, many companies, including Kimi and MiMo, now have base models with more than 1T parameters, and the gap between China and the United States at the pre-training stage is “basically non-existent”。
In her view, there is a certain advantage in the pre-training structure of the national team if the response to the Agent paradigm is fast enoughIn fact, it's only two or three months.
"The next two or three months will be wonderful," she said, and the window period will be a key test of the overall level of research, technical agility and ability to embrace the new paradigm。
At the same time, Roosevelt pointed out that the big model competition had moved from a pre-trained Chat era to a post-trained Agent era。
The 1T parameter size is the current "entry ticket" near the top Agent level, with a reversal of the logic of calculus distribution - the ratio of pre- and post-training calculus input of the top team has been 5:1 in the past and narrow to 1:1 in the present。