华为发布会翻车？官方回应大模型被指造假

文章来源: 于 - 新闻取自各大新闻媒体，新闻内容并不代表本网立场！
(被阅读 次)

5月16日，针对网传华为大模型文生图现场展示疑似人工操控的消息，华为昇腾社区回应称：并非调取预置图片，本次展示的均为真实代码，也将在昇腾社区上开放。

事情的起因是，6天前在鲲鹏昇腾开发者大会的一场面向开发者的技术讨论会上，华为演示了mxRAG SDK功能，即如何通过十几行代码即可完成RAG应用开发。

据网传视频及聊天截图，华为在演示文生图功能时，按下Crtl-C中断，显示对应代码为time.sleep（6），有网友解释代码的意思是：暂停6秒，然后读取本地的一张图片展示出来。

因此，华为被质疑大模型能力演示造假。





昇腾社区回应称，现场图片为实时生成，调用的是开源大模型。代码中有time.sleep（6）等表述，是命令等待读取外部开源大模型实时生成的图片，并非调取预置图片。本次展示的均为真实代码，也将在昇腾社区上开放，欢迎开发者使用并提出宝贵建议。

“具体什么情况很难说，但演示场景还是可以理解的。”一位AI相关开发者告诉《AI光年》，因为目前企业很难接受真实演示场景出故障，所以“大家做demo多少都有一些作假的成分”，他猜测，华为这次演示可能存在“赶工”抢发布热点的情况。

《AI光年》发现，华为此次展示的mxRAG SDK功能，目前还未在开发者资源下载中心上线，资源最新更新时间为4月22日。

据《AI光年》了解，mxRAG的功能为检索增强生成——检索（Retrieval）、增强（Augmentation）和生成（Generation）。这一能力是目前开发大语言模型（LLM）所需的重要能力之一。

据一篇香港理工大学、百度、新加坡国立大学等研究机构发表的一篇论文，LLM+RAG被应用于多个领域，包括但不限于：问答系统、聊天机器人、事实验证、金融领域的决策支持、科学领域的分子发现等。未来还可以提高模型的可信度，开发多语言和多模态的RA-LLMs等。



由于昇腾社区暂未公布源代码，开发者目前仅能通过网传代码截图进行经验判断和讨论，《AI光年》发现，大部分开发者认为文生图过程理论上“完全没必要sleep（6）”。但具体情况如何？华为mxRAG SDK功能如何？还需资源开放后再做验证。

pat4ng 发表评论于
狗粮们都是一帮不学无术的，sleep函数在多进程系统里是经常出现的，为的就是进程之间运行快的等待计算慢的进程而达到进程同步的作用，稍微懂点编程的都知道。狗粮们就别在这里丢人现眼了。。。
======================================
有不同意见可以好好说吗？其实文章没说清楚，因为当时现场演示产生这张图片用了6秒钟多一点，这其中包括sleep了6秒钟，也就是说整个生成过程只用了一秒钟也不到，创造了世界奇迹。这是大家质疑的根本原因
不允许的笔名 发表评论于
为什么是6秒？如果后台真在工作，6秒后台没完成不是大现眼？如果3秒就完成了不是白白遥遥领先了3秒？楼下五毛不学无术还要洗地。

正确的做法是wait event, 后台完成了就set event。
蟹粉小笼包 发表评论于
狗粮们都是一帮不学无术的，sleep函数在多进程系统里是经常出现的，为的就是进程之间运行快的等待计算慢的进程而达到进程同步的作用，稍微懂点编程的都知道。狗粮们就别在这里丢人现眼了。。。
roliepolieolie 发表评论于
华为当年手机照相能力造假，结果人家编辑对着裤衩照，手机上出现了月亮的环形山。这都是哪儿跟哪儿。最让人跌落眼睛的是，华为动用其势力掩盖丑闻，让编辑丢了职位。华为的水有多深，自己掂量掂量。
lzslsb 发表评论于
Here is the Microsoft Edge copilot explanation of the code (think it as chartGPT front end). 看看楼上谁真懂了这程序。
================
repost due to format issue
================
You
please explain the code in the image
(above code image copied to copilot and its interpretation below).

Copilot
Analyzing the image: Privacy blur hides faces from Copilot

Sure, I’d be happy to explain the code in the image. Here’s a breakdown:

Python

if __name__ == '__main__':
prompt = input('>>> ')
save_path = '/home/demo/output.jpeg'

th = Thread(target=ModelInference, args=(prompt, save_path))
th.start()

try:
time.sleep(6)
except KeyboardInterrupt:
running = False
traceback.print_exc()

th.join()
AI-generated code. Review and use carefully. More info on FAQ.
if __name__ == '__main__': This line checks if the script is being run as the main program. If it is, the code within this block will execute.
prompt = input('>>> ') This line takes user input from the console and assigns it to a variable named prompt.
save_path = '/home/demo/output.jpeg' This line defines a file path where the output will be saved.
th = Thread(target=ModelInference, args=(prompt, save_path)) This line creates a new thread to run a function called ModelInference with prompt and save_path as arguments.
th.start() This line starts the execution of the thread.
The try-except block is used to handle exceptions. In this case, it’s used to catch a KeyboardInterrupt exception, which is typically raised when the user hits Ctrl+C while the script is running.
time.sleep(6) This line makes the script pause for 6 seconds.
running = False If there’s a keyboard interruption, it sets a running variable to False.
traceback.print_exc() This line prints the traceback of the exception.
th.join() This line makes the main program wait for the thread to complete execution before proceeding.
This script appears to be part of a larger program for running some kind of model inference. The model inference is run in a separate thread, and the main program waits for a short period of time (6 seconds) for potential keyboard interruption. If there’s a keyboard interruption, it stops the execution and prints the traceback of the exception. Finally, it waits for the thread to complete execution before proceeding. The results of the model inference are presumably saved to the specified path. Please note that the actual behavior might depend on the implementation of the ModelInference function, which is not shown in the image.
lzslsb 发表评论于
Here is the Microsoft Edge copilot explanation of the code (think it as chartGPT front end). 看看楼上谁真懂了这程序。

>> You
please explain the code in the image
>> ')
save_path = '/home/demo/output.jpeg'

th = Thread(target=ModelInference, args=(prompt, save_path))
th.start()

try:
time.sleep(6)
except KeyboardInterrupt:
running = False
traceback.print_exc()

th.join()
AI-generated code. Review and use carefully. More info on FAQ.
if __name__ == '__main__': This line checks if the script is being run as the main program. If it is, the code within this block will execute.
prompt = input('>>> ') This line takes user input from the console and assigns it to a variable named prompt.
save_path = '/home/demo/output.jpeg' This line defines a file path where the output will be saved.
th = Thread(target=ModelInference, args=(prompt, save_path)) This line creates a new thread to run a function called ModelInference with prompt and save_path as arguments.
th.start() This line starts the execution of the thread.
The try-except block is used to handle exceptions. In this case, it’s used to catch a KeyboardInterrupt exception, which is typically raised when the user hits Ctrl+C while the script is running.
time.sleep(6) This line makes the script pause for 6 seconds.
running = False If there’s a keyboard interruption, it sets a running variable to False.
traceback.print_exc() This line prints the traceback of the exception.
th.join() This line makes the main program wait for the thread to complete execution before proceeding.
This script appears to be part of a larger program for running some kind of model inference. The model inference is run in a separate thread, and the main program waits for a short period of time (6 seconds) for potential keyboard interruption. If there’s a keyboard interruption, it stops the execution and prints the traceback of the exception. Finally, it waits for the thread to complete execution before proceeding. The results of the model inference are presumably saved to the specified path. Please note that the actual behavior might depend on the implementation of the ModelInference function, which is not shown in the image.
man008 发表评论于
人家有机器人，我们有人机器。
莫言无语 发表评论于
五毛需要sleep6静候，主子统一指挥。
sandanke 发表评论于
这怎么不见五毛出来洗地？他们主子也不好意思了？
worley 发表评论于
华为应该事先告诉演讲者不要按下Crtl-C中断，就不会露馅了。演示效果是六秒后出现高质量照片。
wisfan 发表评论于
在淘寶上看華為手機用戶的評論，大多數是“還好吧”，也有不少抱怨華為手機的問題。大多數購買華為手機的理由是幫助國產貨。而小米手機用戶的評論幾乎都是讚揚，沒人提購買的動機是支持國貨。
Timberwolf 发表评论于
从亩产万斤到手磨芯片到sleep6秒，你不得不服墙国各界在歪门邪道上的想象力创造力之强大以及对人们的理性思维能力之轻蔑。
barryv 发表评论于
通过sleep来等待后台进程结束，也太低级了。
chinesegod3 发表评论于
文生图这种开源的很多，不是个问题，华为真正的挑战是GPU的生态，一时半会赶不上的。
shakuras2000 发表评论于
华为造假太多了
- 用单反拍照，然后冒充手机拍的
- 有月亮的时候，找张月亮的照片贴图，宣传自己摄像头清晰
我要真普選 发表评论于
造假本領，遙遙領先！
tudoutudou99 发表评论于
为了不让美帝发现华为早已遥遥领先，故意加了6秒的延迟，居然被指造假。掩盖遥遥领先，确实是弄虚作假，华为认了。