Meta's most powerful open source model Llama 3 has been ignored by cloud vendors, and businesses are also unwilling to foot the bill.
Recently, according to the foreign media The Information, $Meta Platforms (META.US)$ the open source large model Llama 3 has been struggling to gain attention on the world's largest cloud provider - $Amazon (AMZN.US)$ AWS.
AWS spares no effort to provide various large model services for its enterprise clients. Currently, Anthropic's closed-source large model Claude is the most popular model on the platform. Surprisingly, despite the high reputation in the tech community, enterprise clients do not seem to endorse Meta's Llama.
In the midst of its continued success, some investors are betting that the company's seemingly unstoppable rise is about to come to an end.$Microsoft (MSFT.US)$On the Azure cloud platform, Llama is also neglected. The Information quotes a Microsoft employee as saying that Microsoft's salespeople only promote Llama to customers with data expertise, such as companies with engineers and data scientists.
The latest and most powerful model, Llama 3.1, released by Meta has been met with a particularly cold market response. Llama 3.1 has been released for one month and has accumulated a total of 3.6 million downloads from Hugging Face, much lower than the first month's 5.8 million downloads of Llama 3.
Meta has invested heavily in developing the Llama series models, with billions of dollars spent just on buying GPUs - obviously, the research and development costs need to be offset by financial returns. However, if the usage is poor, even a powerful open-source large-scale model like Llama will lose its value.
01 Venture companies complain: open-source models are not cost-effective, but rather expensive!
Why don't enterprises pay for Llama, which is known for its strong performance and open-source nature?
The truth is, although Llama is free, many developers are still willing to spend money on using closed-source models because the cost of using Llama is sometimes higher than using closed-source models like OpenAI's GPT.
Free things are the most expensive. In April of this year, after the release of the 70B and 400B models of Llama 3, a U.S. AI entrepreneur in the field of intelligent agents, Arsenii, complained in an interview that Llama 3, which seems to be free, turned out to be unaffordable.
Arsenii found that the efficiency of running open-source large models in his company is much lower than that of using closed-source large models. After experiencing the pitfalls, he summarized two points. First, open-source large models cannot be used like closed-source models that are well-tuned and commercially adapted. Companies still need to optimize and fine-tune them, which requires higher technical requirements for the technical team. Second, open-source models need to be downloaded before they can be used. Large-scale models with billions or trillions of parameters require extremely high IT infrastructure requirements locally, and small and medium-sized companies generally lack IT infrastructure construction that matches them.
After the launch of Llama 3.1, many industry insiders also raised the same question. Although Llama 3.1 is smarter than previous generation products, the cost of deploying it in small and medium-sized enterprises is too high.
Dylan Patel, Chief Analyst at semiconductor research company SemiAnalysis, did some calculations and found that the operating cost of Llama 3.1 405B is extremely high, requiring two H100 servers to run. Renting two H100 servers for a year costs over 0.3 million USD, which is a difficult expense for small companies to bear.$NVIDIA (NVDA.US)$In the United States, Meta Platforms founder Mark Zuckerberg talked about the question of whether open source or closed source is more suitable for the enterprise service market demand. At the WAIC event last month, he put forward a viewpoint that resonated with many people - open-source models have certain value in academic research and teaching fields, which can make the academic community more familiar with the working mechanism of large models and form theories. However, in most scale applications, open-source models are not suitable, especially in fierce commercial competition. Only closed-source models can make the business efficiency higher and the cost lower for enterprises.
In the United States, Meta Platforms founder Mark Zuckerberg talked about the question of whether open source or closed source is more suitable for the enterprise service market demand. At the WAIC event last month, he put forward a viewpoint that resonated with many people - open-source models have certain value in academic research and teaching fields, which can make the academic community more familiar with the working mechanism of large models and form theories. However, in most scale applications, open-source models are not suitable, especially in fierce commercial competition. Only closed-source models can make the business efficiency higher and the cost lower for enterprises.$Baidu (BIDU.US)$In the United States, Meta Platforms founder Mark Zuckerberg talked about the question of whether open source or closed source is more suitable for the enterprise service market demand. At the WAIC event last month, he put forward a viewpoint that resonated with many people - open-source models have certain value in academic research and teaching fields, which can make the academic community more familiar with the working mechanism of large models and form theories. However, in most scale applications, open-source models are not suitable, especially in fierce commercial competition. Only closed-source models can make the business efficiency higher and the cost lower for enterprises.
02 "Open source whitewashing" chaos occurs frequently, and the pace of continuous innovation is slow.
In addition to cost, another major concern for enterprises using large models is performance. So, in terms of performance, who can perform better between open source and closed source models?
Recently, the latest list of the most influential large-scale model evaluation benchmark - Stanford University's MMLU evaluation (Massive Multitask Language Understanding) showed that among the top ten ranked models, only Llama 3.1 is open source, while the other 9 models on the list are closed source.
It can be seen that currently, closed source models generally perform better than open source models.
In fact, although the emergence of powerful open source models like Llama 3 has stimulated the enthusiasm of the open source community, the speed of innovation in open source models is worrisome.
One reason is that the so-called open source models nowadays are not truly open source. Not long ago, Elizabeth Gibney, an editor of Nature magazine, pointed out the "open source whitewashing phenomenon" that currently exists in the AI open source community. Many AI models that claim to be open source are actually not transparent in terms of data and training methods. For example, Meta, which claims to be firmly open source, only opens the weights of Llama, while the actual code is still a 'black box'.
This so-called open source, which contradicts the core of open source philosophy centered around "open source code", is difficult to achieve the success of Linux, where everyone contributes and the flame burns high. This will seriously hinder open source innovation. In addition to catching up with closed source large models in terms of parameters, open source models also need to constantly refine their performance in application. However, the high hidden cost behind open source undoubtedly hinders enterprises.
Over time, the gap between open source models and closed source models will only widen, and open source models will become more and more lagging behind.
Editor/Somer