news

ai has multiple capabilities such as dialogue, grading and lecturing. tal cto tian mi shared the development prospects of the education big model at the 2024 yunqi conference

2024-09-21

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

jimu news reporter hu hao
correspondent hao weilai
from september 19th to september 21st, the 2024 yunqi conference with the theme of "cloud-starting intelligence, industrial transformation" was held in yunqi town, hangzhou. tian mi, cto of tal education group, was invited to attend the yunqi conference and ai native architecture summit, and delivered a keynote speech entitled "the big education model makes the dream of "large-scale teaching in accordance with students' aptitude" come true."
in his speech, tian mi first shared the "impossible triangle" theory that exists in the education industry, that is, personalization, high quality and large scale are difficult to achieve at the same time. "to break the impossible triangle, we can only rely on the development of ai technology. the emergence of large-model ai technology has given us hope."
tian mi said that the education industry is recognized as one of the preferred fields for the application of large models. in 2023, xueersi officially launched the jiuzhang large model (mathgpt). the xueersi team conducted "post-training" including fine-tuning and reinforcement based on the industry's excellent open source base model, continuously improving its own capabilities, so that jiuzhang achieved lower costs and better results than general large models in relevant scenarios.
recently, the new model o1 released by openai has attracted much attention and popularity in the industry. o1 has been trained through reinforcement learning, so it has better performance in reasoning tasks such as physics and mathematics. tian mi also mentioned o1 when sharing the training details of the jiuzhang model, and said that "the jiuzhang model has done a lot of reinforcement work more than half a year ago. there are many innovations at this stage, and the improvement in results is also very obvious."
at present, the jiuzhang model has been applied in many educational scenarios. according to tian mi, the jiuzhang model has the ability to solve problems, dialogue, correct, explain and recommend problems in multiple disciplines, and has been put into practical application in various ways.
for example, chinese and english composition grading and whole-paper math test grading have been implemented in the xueersi learning machine, bringing practical help to parents; english dialogues are used in both the learning machine and the "abc english corner"; the "jiuzhang anytime question" has been launched on the app and mini-program, which can conduct multiple rounds of dialogues and heuristic explanations like a one-on-one math ai teacher; the adaptive recommendation capability can recommend targeted exercises and personalized q&a tutoring to users through step-by-step decomposition and labeling of questions.
tian mi said that xueersi’s success in developing big models is inseparable from the support of the entire industry ecosystem and strong partners such as alibaba cloud. at the same time, xueersi has officially opened up the big model’s problem-solving, lecture, and grading capabilities to the entire industry. the jiuzhang big model will soon be available on the alibaba cloud bailian platform to support and serve more industry users.
it is reported that this year's yunqi conference has invited the world's top technology experts, academic elites, industry leaders and innovation pioneers to gather together, and will hold 400 forums and special discussions, focusing on the integrated development of cloud computing and ai, cutting-edge achievements in large models, ai infrastructure, large model service tool chains, and in-depth industry practices and other industry hot issues.
report/feedback