李沐，一位曾在新华社、人民日报、中央电视台、华尔街日报、纽约时报等资深新闻媒体供职的专业新闻记者和编辑，在创业一年之际，对BosonAI的进展、纠结和反思进行了深刻的回顾。
李沐在亚马逊工作了七年半后，决定辞职创业。他回忆说，如果有什么事情是自己这辈子一定要尝试的，那就是创业。他意识到，一旦开始创业，就需要不断学习新事物，因此他感慨自己为何没有早点开始。
BosonAI的名字来源于他在创业前用Gluon命名的系列项目，灵感来自于量子物理中的玻色子，象征着该项目是亚马逊和微软的联合项目。虽然取名对程序员来说是一个挑战，但李沐和他的团队最终决定使用“Boson”作为新公司的名称，希望人们能理解“玻色子和费米子组成了世界”这个科学梗，并会心一笑。然而，这导致许多人误以为他的公司位于波士顿，而非位于湾区的实际位置。
在2022年底，李沐提出了两个使用大型语言模型（LLM）进行生产力工具的想法。在一次偶然的会面中，他与字节跳动的创始人张一鸣进行了深入交流，张一鸣建议他们为什么不自己开发LLM。李沐最初有所犹豫，因为他之前在亚马逊的团队已经从事了多年的LLM开发工作，但张一鸣鼓励他放眼长远，于是李沐决定继续推进LLM的开发。
在融资方面，李沐遇到了挑战。在他即将签署投资协议的前一天，领投方突然撤资，导致跟投方也随之退出。尽管如此，仍有投资方支持他完成了第二轮融资，并最终拿到了开发LLM所需的资金。
在购买GPU方面，李沐遇到了供货延迟的问题。他灵机一动，直接联系了NVIDIA的创始人黄仁勋，并通过超微的CEO迅速获得了GPU。虽然多付了些钱，但他的团队得以在20天后就拿到了机器，成为第一批使用H100的创业公司之一。
在商业运营方面，李沐表示，BosonAI在创业第一年实现了收支平衡。收入主要来自于为大客户定制模型，这得益于客户们的支持。尽管算力和人力成本巨大，但李沐感谢OpenAI和NVIDIA的支持。
展望未来，李沐预计会有更多公司开始尝试使用LLM，他相信公司会有更好的发展。
英语如下：
News Title: “Li Mu: From AI to Entrepreneurship, Challenges and Growth”
Keywords: Entrepreneurship, Reflection, Pandemic
News Content: Li Mu, a professional news journalist and editor who has served at prestigious media outlets such as Xinhua News Agency, People’s Daily, CCTV, The Wall Street Journal, and The New York Times, has embarked on a profound review of the progress, dilemmas, and reflections of BosonAI upon the first anniversary of his entrepreneurial venture.
After working at Amazon for seven and a half years, Li Mu decided to resign and venture into entrepreneurship. He recalls that if there is one thing he has always wanted to try in his life, it is entrepreneurship. He realized that once he started his own venture, he would need to continuously learn new things, and he felt regret for not starting earlier.
The name BosonAI is derived from a series of projects he named Gluon before starting his entrepreneurial journey, inspired by bosons in quantum physics, symbolizing the joint venture between Amazon and Microsoft. Although choosing a name was a challenge for programmers, Li Mu and his team ultimately decided to use “Boson” as the name of the new company, hoping that people would understand the scientific pun of “bosons and fermions make up the world” and smile in agreement. However, this led to many people mistakenly thinking that his company was located in Boston, rather than its actual location in the Bay Area.
At the end of 2022, Li Mu proposed two ideas for productivity tools using large language models (LLMs). During a chance meeting, he had an in-depth conversation with Zhang Yiming, the founder of ByteDance, who suggested that they might as well develop their own LLM. Initially hesitant, Li Mu, who had previously worked on LLM development in his Amazon team for years, was encouraged by Zhang Yiming to look to the future, and so he decided to continue with the LLM development.
In terms of financing, Li Mu faced challenges. On the day before he was set to sign an investment agreement, the lead investor suddenly pulled out, leading to the exit of follow-on investors as well. Despite this, there were still investors who supported him through the second round of financing, ultimately securing the funds needed for LLM development.
In terms of purchasing GPUs, Li Mu encountered supply delays. He came up with a clever solution by directly contacting NVIDIA founder Jensen Huang, and through the CEO of Advanced Micro Devices, he quickly acquired GPUs. Although he paid more for them, his team was able to receive the machines in just 20 days, becoming one of the first startups to use the H100.
In terms of business operations, Li Mu stated that BosonAI achieved a break-even point in its first year of entrepreneurship. Revenue primarily came from customizing models for large clients, which was made possible by the support of these clients. Despite the significant costs of computing power and human resources, Li Mu expressed gratitude for the support from OpenAI and NVIDIA.
Looking ahead, Li Mu predicts that more companies will begin to experiment with LLMs, and he is confident that the company will see better development.
【来源】https://www.jiqizhixin.com/articles/2024-08-15
Views: 2