Source: Quantum Bit

The core team OneFlow acquired by Wang Huiwen's Lightyear Beyond is about to start a new business.

The news comes from Oneflow founder Yuan Jinhui, and was forwarded on Jike by the team's R&D engineer Yao Chi.

Subsequently, Yuan Jinhui also responded on Jike, emphasizing that the new entrepreneurial project aims to address the cost issue of large model inference.

OneFlow is a well-known domestic open-source deep learning framework and development platform, backed by the company Yi Liu Technology.

Previously, Wang Huiwen's Lightyear Beyond acquired 46.52% equity of Yi Liu Technology, and the legal representative of the company is currently Wang Huiwen.

After acquiring Yi Liu Technology, Lightyear Beyond has obtained a core technical team.

With the progress of financing, a new round of financing was completed in early June, and within 100 days of starting a business, it successfully became a unicorn.

On June 26, just as people were looking forward to Lightyear Beyond having the funds and talent in place and preparing to make a big splash, founder Wang Huiwen's illness was reported.

Three days later, Meituan acquired 100% equity of Lightyear Beyond for 20.58 billion, and the OneFlow team of Yi Liu Technology, as its core asset, also came under the ownership of Meituan.

Fifty days later, Yuan Jinhui announced that he would lead the original OneFlow team to start a new business again, whether it will be independent of Meituan and whether a new company will be established is currently unknown.

In addition, enterprise check shows that the company address of Yi Liu Technology has also changed recently, moving from Wudaokou to the neighboring Zhongguancun.

OneFlow starts a new business, targeting AI inference

Yuan Jinhui, founder and CEO of Yi Liu Technology, is a well-known AI architecture expert (online name "Teacher Mu"), with a Ph.D. from the Department of Computer Science at Tsinghua University, under the guidance of the Chinese AI pioneer, Academician Zhang Bo, and won the Outstanding Ph.D. Dissertation Award at Tsinghua University.

Before founding Yi Liu Technology, he was a research director at Microsoft Research Asia, where he developed the fastest topic model training algorithm LightLDA at the time, and built a deep learning platform based on heterogeneous clusters, winning the Special Award from the Director of Microsoft Research Asia.

Yuan Jinhui was also responsible for the development of the Chinese Eagle Eye system, widely used in sports events and daily training of national teams.

Since its establishment in 2017, Yi Liu Technology has launched the heterogeneous distributed deep learning framework OneFlow, the MLOps full-lifecycle management platform OneBrain, and the large-scale model training open-source toolbox Libai (Li Bai), etc.

In short, it is more focused on the training phase of AI.

At the time of Oneflow's first entrepreneurship, there was no clear concept of "large models" in the industry. At that time, the visual model based on the convolutional neural network architecture was the main focus, and a parameter scale of several tens of millions was considered large. Efficient training of larger models was the direction Yuan Jinhui was optimistic about.

Now, as the team embarks on its second entrepreneurship, large models that have been trained and deployed, whether open-source or proprietary, are flourishing both domestically and internationally. Therefore, shifting focus to AI application inference and deployment is a natural choice.

Entrepreneurship in AI infrastructure is heating up

Whether it's AI training or inference, it can be attributed to a larger concept, AI infrastructure (AI Infra).

In today's large model entrepreneurship market, in addition to developing basic models and applications for large models, entrepreneurship in large model infrastructure is also experiencing a new wave of explosion.

Lepton AI, a startup company founded by former Alibaba VP Jia Yangqing, is also entering AI Infra.

Similar to Yuan Jinhui, Jia Yangqing was initially known as an AI framework expert. During his doctoral studies, he created the well-known deep learning framework Caffe, and while working at Google and Facebook, he was also a core member of mainstream frameworks such as TensorFlow and PyTorch.

Now, based on the brief introduction on the Lepton AI official website, "efficiently and on a large scale running artificial intelligence applications" is also targeting the inference deployment of large model applications.

The open-source LlaMA2 large model deployed by Lepton AI was praised by Vercel CEO as "the fastest I've seen".

(The image is not accelerated, the alpaca eats very fast)

Speaking of mainstream deep learning frameworks, a group of members from the PyTorch team recently left to start the AI Infra company fireworks.ai.

Among the three co-founders, CEO Lin Qiao was previously the Senior Engineering Director at Meta, Dmytro Dzhulgakov was the Technical Director, and Ning Li was the Deputy Director of Engineering.

However, unlike the new project of OneFlow and Lepton AI, from the information on the official website, fireworks.ai currently focuses more on fine-tuning of large models.

Reference links:
[1]https://m.okjike.com/reposts/64df593e58c9056cf4f020e2
[2]https://twitter.com/rauchg/status/1692638409230094644
[3]https://www.lepton.ai
[4]https://www.fireworks.ai/

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。