The chat-GPT company OpenAI and Google dominate the AI market. Can open source AI be the solution?
Mark Zuckerberg is not necessarily known for artificial intelligence. But his company is investing a lot in their open development.
Carlos Barria / Reuters
It’s an unusual alliance, but times are also unusual in the tech world: the Facebook group Meta, software companies like IBM and Oracle, the chip manufacturer Intel and prestigious research institutions like Harvard University, but also the Swiss ETH and that Cern. On Wednesday they announced that they wanted to work together on open artificial intelligence. AI Alliance is the name of the association.
The background: A year ago, the company OpenAI made powerful voice AI generally accessible with the chatbot Chat-GPT. Writing essays or solving programming tasks at the push of a button, all of this was suddenly possible. Private individuals are trying out how they can use AI for their own benefit, and companies are incorporating it into their products.
This set a machine in motion from which two players in particular are currently benefiting: OpenAI and Microsoft. The majority of users rely on OpenAI’s language models because they are the largest on the market and pay fees for them. Microsoft is an investor in OpenAI and provides the cloud infrastructure for voice AI. This means that the Microsoft Cloud Azure has a decisive advantage over competitors such as Amazon Web Services (AWS) or the IBM Cloud.
It’s difficult to compete with OpenAI’s dominance for four reasons: First, training this type of AI requires tons of data. In addition to data, this training also requires huge amounts of computing power. Third, you need enough of the rare and well-paid AI specialists to adapt the model. And once it is ready and many users are interested, fourthly, enormous amounts of computing power are required to generate the requested texts, images and programming codes.
There are only a few companies that currently have the resources and minds to play in the same league as OpenAI and Microsoft: above all Google, which has just launched Gemini, a new alternative to OpenAI’s models, and Meta.
Meta relies on open AI – almost
Meta pursues an open strategy in AI research: While OpenAI now presents its models in blog posts, Meta publishes research papers on them. With Llama2, Meta has also made one of the largest language models publicly usable. You can download it free of charge upon request, adapt it for your purposes and use it.
However, Meta’s AI doesn’t really meet open source standards. Meta reserves the right to exclude companies with more than 700 million users from using it. This affects its biggest competitor, Google, but also apps like Snapchat. And the data and source code used for training are secret.
Open source developers like Frank Karlitschek, the founder of Nextcloud, see this critically: “You have to differentiate between free software, is it ‘free’ as in free beer or as in freedom?” It is actually not responsible to build on a model for which one does not know what data was used for it.
Despite this criticism, it is good news that Meta now wants to work more closely with IBM and many other companies and research institutes to jointly offer and further develop open models. Because alternatives to OpenAI’s monopoly are necessary.
Anyone who only relies on OpenAI will become dependent
When companies experiment with AI, they almost never develop their own AI model. This is too complicated for the reasons mentioned above. They often buy usage rights from OpenAI instead.
The chatbot from Helvetia Insurance, for example, was optimized with chat GPT technology. When Helvetia customers ask him a question, Helvetia sends a request to Chat-GPT in the background. Voice AI is good at “understanding” what questioners are asking and formulating answers. These strengths are combined with the data from Helvetia, from which the chatbot gets the information for the answer.
The advantage: You can build a functioning application very quickly. The disadvantage: You are dependent on OpenAI charging acceptable prices, handling the data it receives well, and maintaining its services.
If you’re just running a chatbot on the side, that’s acceptable. But there are also many startups whose entire business model relies on the availability of OpenAI. They depend on OpenAI.
There are opinions and worldviews in AI – but whose?
It is also socially worrying if only a few Silicon Valley companies develop powerful voice AI and no one knows what exactly is in the models. Especially when more and more people no longer question Google, but rather their chatbot. Then it becomes very relevant which worldviews and opinions the AI represents in its answers.
Real open source models, in which training data and methods are transparently visible, benefit society.
The company Hugging Face produces such completely open models. Leandro von Werra is a researcher and machine learning engineer there. He emphasizes that AI models are based on the works of millions of people. They have the right to know when their content is being used – and need ways to protect it from being exploited by AI.
Hugging Face is also part of the AI Alliance. The hope is that the collaboration of many players in the open source sector will enable real competition to the big tech companies. The highlight of open source is that everyone benefits when you share basic software and everyone can detect errors and suggest improvements.
In the future there will be both types of AI
It’s not just about who has the most powerful model, but also about its applicability in everyday life.
A business solution like the one from OpenAI is the easiest choice for many companies at the moment. You buy AI there as a complete package. Using open, collaborative software also means that you have to invest more knowledge and time until the programs are ready for use. In addition, there are the costs for the cloud, which are already priced into OpenAI.
Open source AI companies have to close this gap. Traditionally, open source companies make money by helping corporate customers introduce and maintain free software. If the developers of open AI agree on standards and procedures, all providers in the area can benefit. This is the logic behind the AI Alliance.
When it comes to regular computer programs, most people use both private and open source software in everyday life. Companies buy some things from Microsoft, SAP and Co. – others they develop themselves and rely on open source solutions.
Leandro von Werra from Hugging Face also expects such coexistence with AI: “There will be a place for closed general-purpose models like Chat-GPT or Claude, which can assist with many questions. But I expect that many companies will build their own AI models based on open source.”
Worries about abuse
There are also critics of open source AI. Their main argument is: Artificial intelligence is a powerful tool, for example for producing disinformation, fakes or otherwise undesirable content. If anyone can download and customize an AI model, it will be impossible to limit abuse. Because filters can be deleted from the source code. So is open AI a danger?
The open source supporter Karlitschek says: “This is a fundamental question: Would we prefer if a few large companies have control and decide what is good and what is bad? Or do we rely on transparent and decentralized AI?” Of course he would prefer the latter.
The aim of the AI Alliance is now to follow up the declaration of intent with action. There are no clear agreements on this, and it is clear from ETH that no money will flow. Alessandro Curioni, director of the Zurich IBM Research Center, puts it this way: There are no fixed obligations for the members of the alliance to contribute something. “But it is in our own interest to get involved, whether with ideas, data, reputation or expertise.”