The newest AI model developed by Chinese tech giant Tencent has received positive feedback from developers. Tencent employees utilized Anthropic’s Claude to test and refine their model, named Hy3. This occurred despite the fact that Anthropic does not provide its models and services to companies in countries considered adversaries of the United States such as China.<\/p>\n\n\n\n
It is not known how the employees of Tencent located in China gained access to Claude, but it is commonly used by Chinese tech companies and startups who receive access to Claude via non-Chinese mobile number or non-Chinese credit card.<\/p>\n\n\n\n
In the first half of this month, Anthropic made identity requirements more stringent when using its models, which meant that some customers needed to submit a government-issued photo ID and a picture of themselves taken on their phone or webcam. The relocation provoked anxieties among the afflicted many Chinese startups and programmers who are dependent on Claude.<\/p>\n\n\n\n
Last week, Tencent issued a preview of Hy3 which is free of charge. The company says it is \u201cthe most intelligent model\u201d in its series of AI models known as HY, or Hunyuan. Tencent has released its first major model since Yao Shunyu, a former OpenAI researcher, joined Tencent as its chief AI scientist in September. Hiring Yao is among the constant remodeling of Tencent to keep up with competitors Alibaba and ByteDance in the development of foundational models.<\/p>\n\n\n\n
In the process of creating Hy3, employees at Tencent used the AI coding tool of Anthropic, Claude Code, during the post-training phase called reinforcement learning with human feedback.<\/p>\n\n\n\n
To ensure that the model was developed as human evaluators, Tencent recruited employees to be part of the model development during this stage. According to one of the people and the memos, the company gave them the instructions on how they can install Claude Code.<\/p>\n\n\n\n
One of the memos revealed that Tencent limited the number of Claude Code tokens that each evaluator could access to thousands of tokens. A token is about four characters in English language.<\/p>\n\n\n\n
In response, an Anthropic spokesperson told them that due to national security reasons, they did not currently provide commercial access to Claude in China, or even to subsidiaries of their companies located outside of the country.<\/p>\n\n\n\n
The spokesperson also indicated that the company is still detecting and preventing the attempts to illicitly use Claude in training the competing models.<\/p>\n\n\n\n
The spokesperson did not comment directly on how Tencent uses Claude Code, but he said that our safeguards and threat intelligence teams are actively looking after distillation attacks and we take immediate action when we detect accounts or organizations involved in it.<\/p>\n\n\n\n
The procedure of training a smaller or less powerful model basing on the results of a bigger or more powerful model is known as distillation, a frequently used process in AI model development. This is typically done by inputting the outputs of the stronger model into a new model being developed.<\/p>\n\n\n\n
The fact that Claude Code was used in the development distillation of Hy3 does not enter their minds since they did not feed the responses of Claude to Hy3. Their work with Claude Code, they said, is more like benchmarking, where another model is used that is considered the best in its category as a reference to which the new model learns to make better decisions through trial and error. According to employees at other AI companies, it is very common to find AI model developers use leading models such as Claude in such purposes during the post-training work.<\/p>\n\n\n\n
As an example, in the benchmarking Tencent employees asked the same programming questions to two unknown models and rated the answers on various scales, according to one of the individuals with firsthand experience and the internal memos. The employees were not aware of what models they were playing with when scoring. They also used Claude Code to create the examples of a high-quality AI behavior in real time and used them as guiding references, to analyze the Tencent model and filter out low-quality responses.<\/p>\n\n\n\n
Although not regarded as state of the art, developers have praised the efficiency with which the Hy3 preview can code as compared to other similarly-sized models. It is rated number nine in the programming category by usage, according.<\/p>\n","protected":false},"excerpt":{"rendered":"
The newest AI model developed by Chinese tech giant Tencent has received positive feedback from developers. Tencent employees utilized Anthropic’s Claude to test and refine their model, named Hy3. This occurred despite the fact that Anthropic does not provide its models and services to companies in countries considered adversaries of the United States such as […]<\/p>\n","protected":false},"author":5,"featured_media":3162,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"inline_featured_image":false,"footnotes":""},"categories":[1],"tags":[],"tmauthors":[34],"class_list":{"0":"post-3161","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-blog"},"acf":[],"yoast_head":"\n