CTB11的个人博客分享 http://blog.sciencenet.cn/u/CTB11

博文

GPT4之壁

已有 1470 次阅读 2023-4-20 08:23 |系统分类:观点评述

近期OpenAI的CEO Sam Altman 表示OpenAI 没有近期训练GPT5的计划,这跟之前的流言有很大的出入。之前看到的流言是GPT4.5会在今年下半年开放,然后GPT5会在今年12月训练完成,看来也只是谣言了。

我尝试问GPT4为什么OpenAI不再训练GPT5,下面是ta的回答:

While I cannot speak directly for OpenAI or its CEO, I can provide some possible reasons as to why they might not have plans to train GPT-5 in the near future:

  1. Technical limitations: Training larger models like GPT-5 would require significantly more computational resources and technical advancements. OpenAI might focus on addressing these limitations first before committing to a new model.

  2. Environmental concerns: The energy consumption required to train large models is substantial. OpenAI might be taking into account the environmental impact and looking for ways to reduce energy usage before developing a new iteration.

  3. Ethical concerns: OpenAI might be focusing on addressing the ethical implications of AI systems, such as biases, misinformation, and unintended consequences, before developing a more powerful model like GPT-5.

  4. Improving existing models: Instead of creating a new model, OpenAI might be focusing on improving the current GPT-4 model by refining its training data, architecture, and other aspects to enhance its capabilities.

  5. Commercialization and deployment: OpenAI may be prioritizing the commercialization and deployment of their existing technologies, ensuring that GPT-4 and other AI systems are effectively integrated into various industries and applications.

  6. Regulation and policy: The rapid advancement of AI technologies has led to increased calls for regulation and policy development. OpenAI might be engaging with governments and stakeholders to help shape the legal and regulatory landscape surrounding AI technologies.

  7. Collaboration: OpenAI might be looking to collaborate with other organizations, research institutions, and AI developers to pool resources, knowledge, and expertise to develop more advanced and responsible AI systems.

  8. Alternative AI approaches: OpenAI may also be exploring alternative AI technologies and approaches, such as reinforcement learning, unsupervised learning, or hybrid models, which could potentially surpass the capabilities of the GPT series.

说的已经很不错了。我作为人类一个个体的简单看法是:(1)OpenAI的算力应该已经很紧张了,没有办法在部署GPT4的同时训练和部署规模更大的模型。GPT4.5能否如期面世没准也要打个问号。(2)训练比GPT4更大的模型可能是人类的未知领域,已有的benchmark应该没有太大意义了。(3)没有多余的语料给GPT5了,我怀疑GPT4已经用光了internet中优质(甚至劣质)的语料。


接下来两年是应该国内追赶LLM技术的很好窗口。



https://wap.sciencenet.cn/blog-927304-1384923.html

上一篇:机器人会梦见电子羊么--浅谈生命科学中的学科交叉
下一篇:如何在没有经费的情况下做科研
收藏 IP: 50.234.189.*| 热度|

0

该博文允许注册用户评论 请点击登录 评论 (0 个评论)

数据加载中...

Archiver|手机版|科学网 ( 京ICP备07017567号-12 )

GMT+8, 2024-7-23 22:28

Powered by ScienceNet.cn

Copyright © 2007- 中国科学报社

返回顶部