The impact of open source large models on closed source large models has become very violent.
In March of this year, Meta released Llama (alpaca), which quickly became the most powerful open source large model in the AI community and the base model for many models. Some people joked that the current large-scale model cluster is just a bunch of "alpacas" of various colors.
And just a few days ago, Meta launched a free commercial version of "Alpaca 2" - Llama2, which is said to be comparable in performance to GPT-3.5.
This is very explosive in the entire large model circle.
We know that various Internet and technology companies are competing to train and launch their own large-scale models, investing a lot of computing resources and costs. If they cannot be effectively commercialized, it will be difficult to recover the cost of these large-scale models. Subsequent iterations, updates, and upgrades will become problems. Not only will R&D companies lose money, but users who "waste all previous efforts" will probably be more distressed.
But now that there are free, open and powerful open source models, who is willing to give money to closed source models?
There really are.
Open source is the general trend, but the closed source big model still has its existence significance and commercial value. According to the current experience in the AI industry, to make good use of large models, you still have to rely on closed sources.
Today we are going to talk about this issue. Who needs a closed-source large model?
to the industry, to the industry
The end point of commercialization of large models is the industry, and it must be a consensus that does not require too much explanation.
Not long ago, I participated in an internal communication meeting of a domestic large-scale model, and the high-level executives of the other party clearly stated that they all use closed-source code and insist on the closed-source route, because they consider training large-scale models and cooperating with industry partners, and many of the private data are inconvenient to open source.
You can get a glimpse of the whole picture, at least in the short term, large-scale models will go to the industry, and the implementation still depends on closed sources.
In terms of models, the quality of closed-source large models is higher.
Take Llama 2, which is currently the most capable, as an example. Meta compared the results of Llama 2 70B with the closed-source model. The results are close to GPT-3.5 on MMLU and GSM8K, but there is still a significant gap in the coding benchmark, and many data are lacking in diversity and quality.
Of course, the optimization iteration speed of open source large models is very fast. But the essence of open source is very similar to "sexual reproduction", that is, through mass reproduction and mutation, just like the "alpaca cluster" at the beginning, in the face of an uncertain future, with the help of evolution's "survival of the fittest", the best quality offspring will continue to emerge. Therefore, there are many branches of open source software. For users, the cost of this choice is very high. In addition to the large number of developers, version control is a problem.
In terms of security, closed-source large models are more reliable.
Open source large models must abide by the open source agreement, and commercial use needs to be authorized. Overseas open source large models must also be subject to territorial jurisdiction. GitHub once banned Russian developer accounts. Using overseas open source large models to develop products and supply chain risks exist objectively.
So, what about using domestic open source large models? Safety is guaranteed, but from a commercial point of view, many customers, such as large government enterprises, also attach great importance to the reliability of large models in business, and often require the brand endorsement of large companies when purchasing. On the one hand, the investment in R&D is greater and the word-of-mouth is higher; on the other hand, in case the large model is generated improperly, resulting in commercial loss or goodwill problems, the use of the closed-source large model can hold the service provider accountable, and the use of the open-source large model can’t settle accounts with global developers, right?
For example, Huging Face, a large-scale model startup company, provides AI consulting for customers and is a pillar of the open source community. It said that a large number of customers want to use their private data/professional data for training models, and do not want to give these data to OpenAl.
In terms of industrialization, the long-term service capability of closed-source large models is stronger and more usable.
Landing a large model does not end with accessing APIs, inserting data, and tuning parameters. As an emerging technology, there are still many challenges in the integration of large models and business scenarios. For example, large models need to be compressed by distillation to reduce the size of the model before they can be deployed on the device side. Many companies simply do not have such professionals.
For another example, the combination of large models and business requires the participation of multiple roles such as product, operation, and test engineers. These service capabilities are difficult to provide for open source teams that are mainly coders. In addition, the long-term application of large models, computing power, storage, network and other supporting facilities must keep up, and the open source community cannot help users "one-stop" solve these detailed problems.
There are also data privacy concerns. Large models cannot be directly used by the industry. They must be optimized through proprietary scene data, and the models trained on these data will be open sourced and released, which makes enterprises worry a lot.
We once interviewed a smart medical research and development team. The other party said that a large amount of medical data is distributed in major hospitals and research institutions, and it also involves patient privacy. Everyone has concerns about using the data to jointly train an industry model. On the one hand, security cannot be guaranteed, and on the other hand, the quality of their own data is high, but they cannot get proper returns from it. Like other organizations with low-quality data, it is difficult to coordinate. In the co-construction of open source large models, there are still many difficulties in how to obtain data, grasp the formula, and determine the contributions of all parties.
Open-source large models need to balance the conflict between technological innovation freedom and copyright benefits, while closed-source large models do not have this trouble. The ownership and use rights of data and models are very clear, and they are firmly in the hands of the enterprise itself.
It can be said that the current open source large model cannot meet the actual business needs. However, open source large model users and ISV integrators need to obtain commercial returns. If the open source large model is not commercially available, the effect is not good, and it is difficult to make money, even if it is free, the enterprise will carefully consider whether to invest in people to develop it.
Therefore, for some time to come, closed source will still be a popular choice for the large-scale model landing industry.
to the masses, to the masses
Some people may not understand, open source is free for commercial use, and everyone can use a large model at the price of cabbage. It is so friendly to developers and enterprise users, why do you still say that closed source is better? Is it the platform of a big factory focused on making money?
No.
Anyone who understands open source will support open source. Anyone who supports open source will pay attention to the commercialization of open source.
Academician Mei Hong of the Chinese Academy of Sciences once said that open source originates from idealism and is vigorously fueled by commercialization. It is a model of open innovation. Without commercialization, there can be no open source.
Therefore, whether it is open source or closed source, whoever can be "commercial" earlier will have a better future. In this regard, closed-source large-scale models may have an advantage. After all, manufacturers with the confidence to close the source still have two brushes and R&D background.
So, what are the advantages of open source large models? If the closed-source large-scale model is going to the industry, then the open-source large-scale model must go to the masses, focusing on the strength of one person.
(LeCun believes that Llama-v2 will change the market structure of LLM)
The open source big model is different from traditional open source software, where the source code is put on it, and then developers from all over the world contribute the code and that's it. The collaboration and co-construction of large models is more reflected in the prosperity of the community. Everyone works together to optimize the model, enrich the data, improve the tools, and make the application comprehensive...
At this time, the open source model can bring several benefits:
1. Technological innovation. The open source community can bring together a large number of technology companies, research institutions, and developers to optimize, improve, and accelerate iterations of the model, making the model technology and supporting data sets, application tools, etc. rich and high-quality, so as to stay ahead.
2. Talent competition. As an emerging technology, large models are in short supply of talents. The gap can be widened by attracting outstanding talents from all over the world to contribute through open source communities and accelerating the upgrading of large models. There is pressure when there is competition, so after the release of LLama 2, it was soon reported that OpenAI also began to consider open-sourcing GPT-3.5 within half a year. The developers are blessed.
3. Ecological closure. At present, IT solutions and digital transformation in all walks of life use a large number of open source technologies and applications to build a large-scale open source ecosystem, allowing IT talents and enterprises to use related technologies, which is very helpful for later commercialization. For example, Microsoft, the partner/investor of OpenAI, also chose to become the primary partner of Llama 2 this time, supporting individual developers and small and medium-sized companies to call Llama 2 at the lowest cost, which is undoubtedly a great benefit for azure.
Not all large open source models can succeed, and ecology is the key moat.
Sandwich biscuits, where are you going?
Just like iOS and Android, the mobile operating system, the competition between open source and closed source is not a "life and death" fight in a certain field, but each takes a differentiated path and ushers in its own world. The same goes for large models.
Closed-source large-scale models are open to welcome customers, open-source large-scale models are booming, and everyone has a bright future.
That being the case, why do some experts believe that the open source of Llama 2 is a huge leap for open source, but a huge blow to closed source large model companies?
Who did it hit?
The answer should be that it is a basic large-scale model manufacturer who is not willing to only be an application layer, but also unable to overwhelm a large manufacturer.
Google researchers once wrote that because of the open source community, we (Google and OpenAI) have no moat. However, OpenAI also has closed-source large models such as GPT-4 as its killer feature. Only when it is forced to open source, it considers open-sourcing GPT-3.5. There is a technical gap in it. Moreover, the open source of GPT-3.5 only revealed the word of mouth, and the specific progress is still unknown.
Therefore, such leading technology manufacturers and cloud giants, such as overseas Google, OpenAI, and domestic BATH, have advantages in cards, money, talents, data, market awareness, and customer base. Taking the closed-source route to complete the commercialization and industrialization of large models has certain first-mover advantages and barriers.
This is a pain for those second- and third-tier manufacturers who want to train the basic general-purpose large model.
Previously, large and small technology companies and various scientific research institutions around the world flocked to train basic large models, such as some machine vision AI unicorns, which accidentally became "sandwich biscuits" between the basic layer and the application layer.
It can't beat GPT in terms of strength, and can't beat Llama in terms of cost. The basic general-purpose large model trained is outdated before it is officially opened for commercial use, and it is destined to be a thing of the past. The market cannot compete with giants, and the degree of openness is not as good as that of the open source community. It is almost impossible to recover the high development costs.
It may be a wise choice to give up the big model as soon as possible.
For example, the large-scale model of a domestic AI company was previously privatized at a price of 300,000 yuan a year, and then it was announced that it was completely open to academic research, and it was authorized for free commercial use. There is also the possibility of commercialization (such as Linux/Android/Red Hat) in the large-scale model open source community, and at the same time, it can avoid "head-to-head" with the general large-scale model of the head.
(Screenshot of a well-known investor's discussion on Llama2 open source/from the Internet)
For application layer developers and ISV integrators, making good use of closed-source large models with high industry acceptance can allow customers to accept them faster, be more suitable for the business needs of privatized customized deployment, and complete commercial landing and revenue growth faster.
For AI start-ups, open source can be used directly and avoid repeated wheel creation. It may be a more ideal and low-cost trial-and-error commercialization method. "Reporting to the group for warmth" contributes to large-scale open-source projects, promotes the development of large-scale open-source communities, and will also receive community feedback and business feedback.
The development of China's large-scale model to a high level requires not only the world's leading closed-source large-scale model to take the lead, but also an open-source large-scale model community with world influence.
The road is obstructed and long, but the journey is approaching. May wish to use a constructive attitude to look at the open source and closed source disputes, give some confidence to the domestic closed source large model, and also give some encouragement and support to the domestic open source community.
