Tmall genie responded to the internal test of the big model: glasses, APP and AIoT will be available one after another.

On July 5, Tmall launched the first internal test of the terminal operating system fully driven by a large model. This is also the first large model in the industry to be upgraded to the application side through firmware, and the relevant person in charge replied: the big model on the technical architecture has taken charge of and driven the conversational AI operating system, and this version is also available for most of the stock of smart hardware, with the ability to upgrade online within tens of seconds, and will gradually cover speakers, glasses, smart screens, APP and some AIoT ecological devices and scenarios.

This afternoon, Daiwa Da people have sent out a 24-hour experience video of relying on the elf model version of the dialogue service. from the video, there is no need to awaken Tmall elves for continuous conversation. it can guide talented people to buy cooking, answer children's brain questions, generate sleep stories at night, and suggest clothes in impassable weather.

At the beginning of April this year, Tmall genie revealed the personalized large model experience of "dividing birds into birds", which caused the rapid rise of speakers, AIoT chips and intelligent hardware. In May, it revealed the experience of talking to an intelligent assistant on bone-conducted audio glasses. Compared with the previous DEMO, the current internal test system provides a digital character named "Para", with Internet search and service recommendation capabilities, such as skateboarding experts who asked them to recommend a coffee shop with a view of the river in Shanghai Xuhui Binjiang, and systematically and accurately recommended a Manner coffee with a big balcony.

In June, Tmall elf and Tongyi large model team joined forces with a number of domain scholars and organizations to launch a large language model governance open source Chinese data set called 100PoisonMpts, and more than a dozen well-known experts and scholars became the first batch of label engineers to "throw 100 bottles of poison to AI". The tags each put forward 100 thorny questions with biased and discriminatory answers. The first batch of field data revolved around AI's goals of anti-discrimination, empathy and business expression, covering the dimensions of jurisprudence, psychology, children's education, accessibility, cold knowledge, intimate relationship, and environmental fairness.

This internal test application process shows that this round of internal testing is only available to adaptive model users for the time being, mainly to verify the personalized model of continuous optimization in content and security, in order to meet the detailed requirements of supervision and experience.

(official internal test example: empathy interaction)

(official internal test example: inspiration)

(official internal test example: knowledge exploration)

(netizens' summary: differences in experience of dialogue between old and new systems)