在 OpenAI 擅长的领域打败 OpenAI 微软称自研模型仅凭对方 1/4 算力就超越同类产品Beat OpenAI in areas where OpenAI excels Microsoft said the self-developed model surpassed similar products by only 1/4 of the computing power of the other party


AI free token/AI免费积分
在 OpenAI 擅长的领域打败 OpenAI 微软称自研模型仅凭对方 1/4 算力就超越同类产品

IT之家 10 月 16 日消息,据《商业内幕》报道,当地时间 15 日,微软 AI 业务 CEO 穆斯塔法 · 苏莱曼在员工大会上称,微软自研的一款 AI 模型表现超过 OpenAI 的同类机型,而且训练成本有望更低。“笔者的预训练基础模型表现超过了当下驱动 GPT-5.6Sol 的模型,训练时使用的算力实际上只有 OpenAI 的四分之一。”

预训练是 AI 模型开发的第一步,模型需要借助海量文本、图像等数据学习其中的规律。由于预训练需要投入大量算力,成本高昂,如果微软可以提高训练效率,就有望大幅降低成本。

当日早些时候,微软发布了这款自研模型,名称为 MAI-Base-2。

苏莱曼透露,微软曾私下将 MAI-Base-2 与多款 OpenAI 模型开展比较,其中涵盖代号为 Spud、支撑 GPT-5.6 运行的模型。在测试 AI 理解和预测信息规律的能力时,微软的模型取得了更好的成绩。

IT之家从报道中获悉,比较依据是微软掌握的 OpenAI 模型工艺资料。苏莱曼表示,微软维度“私下以保密方式”将其与 OpenAI 的模型作了比较,暗示微软掌握着 OpenAI 模型表现维度的特殊信息。

另外,苏莱曼还称,微软的新模型在计算效率维度同样具有长处,“以 DeepSeek 为例,笔者所需的计算量只有它的十二分之一。”不过,他没有透露比较的对象是何种 DeepSeek 模型。

本文来源:https://www.ithome.com/1/013/761.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT House October 16th news, according to the "Business Insider" report, local time on the 15th, Microsoft AI business CEO Mustafa Suleiman said at the staff meeting, Microsoft's self-developed AI model performance than OpenAI's similar models, and training costs are expected to be lower. "My pre-trained base model outperforms the current model that drives GPT-5.6Sol, and the hashrate used during training is actually only a quarter of that of OpenAI."

Pre-training is the first step in AI model development. Models need to learn the rules with the help of massive text, images and other data. Because pre-training requires a lot of computing power and is expensive, if Microsoft can improve the efficiency of training, it is expected to greatly reduce costs.

Earlier in the day, Microsoft released a self-developed model called MAI-Base-2.

Suleiman revealed that Microsoft has privately compared MAI-Base-2 with a number of OpenAI models, including the model code-named Spud that supports GPT-5.6 operation. Microsoft's model achieved better results when testing AI's ability to understand and predict information patterns.

Beat OpenAI in areas where OpenAI excels Microsoft said the self-developed model surpassed similar products by only 1/4 of the computing power of the other party

IT House learned from the report that the comparison is based on the OpenAI model process data held by Microsoft. Suleiman said that Microsoft dimension compared it with OpenAI's model "privately and confidentially", suggesting that Microsoft has special information on the performance dimension of the OpenAI model.

In addition, Suleiman also said that Microsoft's new model also has advantages in the computational efficiency dimension. "In the case of DeepSeek, the amount of computation required by the author is only one-twelfth of that. However, he did not disclose what kind of DeepSeek model was compared.

This article is from: https://www.ithome.com/1/013/761.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry