Open Source Movement Emerging in AI To Counter Greed
The early days of artificial intelligence research had a community feel — to openly share ideas, and work together to improve technologies. But a lot changed with ChatGPT, which took the world by storm last year.
Tech behemoths like Google, Microsoft, and Facebook are now looking to cash in on the AI gold rush by tightening access to tools that are defining the AI landscape. They are restricting access to tools that can answer questions, generate images, and understand spoken languages.
But for some, especially hardware makers, handing over control of AI to a few wealthy tech companies is bad for business. These companies are supporting an emerging open source movement, so AI technologies are cheap and accessible.
Defectors
OpenAI, which is a premier AI firm, is a notable defector. The company started in 2015 as a nonprofit with the aim to promote and share AI research. It opened access to its Large Language Models, including GPT-3, which is the force behind ChatGPT.
But OpenAI didn’t open up its most recent large-language model, GPT-4, which was released last month, and is used by Microsoft in its Bing search. Microsoft has invested billions in OpenAI, which became a for-profit entity in 2019. OpenAI is charging to access GPT-4.
Elon Musk, who was an early donor to OpenAI, in a tweet noted that OpenAI was intended to be open source, but “has become a closed source, maximum-profit company effectively controlled by Microsoft.” To be sure, Musk committed to investing $1 billion in OpenAI, but stopped after a power struggle, and Microsoft stepped in to provide a much-needed cash infusion.
One aspect of the open AI ecosystem revolves around opening Large Language Models (LLMs) for closer scrutiny by the community. Besides OpenAI, tech giants Google, Facebook, and Nvidia are developing their own Large Language Models and deploying them to hardware. However, access to the most advanced AI tools is restricted to either researchers or a few developers.
Closed LLMs
There are concerns about openly distributed models being used for nefarious purposes. OpenAI has cited safety as a reason to keep GPT-4 closed. Bloomberg is keeping its recently released Bloomberg-GPT model closed for safety and business reasons. The model was trained on decades of data that form the basis of financial services provided by Bloomberg to clients.
“As is well known, LLMs are susceptible to data leakage attacks, and it is possible to extract significant segments of text given model weights. Moreover, even giving selective access to researchers isn’t a guarantee that the model cannot be leaked,” Bloomberg researchers said in a paper detailing the model.
Restrictive access to closed Large Language Models could be provided via APIs, but “even giving selective access to researchers isn’t a guarantee that the model cannot be leaked,” the researchers said.
Development tools like TensorFlow and PyTorch are already open source but require high-performance hardware like GPUs to execute programs.
But more companies are joining a grassroots movement to open up Large Language Models so proprietary models don’t dominate the market.
Hardware Makers Take the Lead
Hardware makers are leading the early charge to promote open source AI. Cerebras Systems, which makes what is considered the world’s largest AI chip the size of a wafer, last month released Cerebras-GPT models with up to 13 billion parameters.
“We’re open sourcing the weights, we’re open-sourcing checkpoints, we’re showing and providing the full recipe so you can copy it. We are doing this under the most permissive open source license possible,” Andrew Feldman, CEO of Cerebras Systems, told The New Stack.
These models are forks of OpenAI’s GPT-3, which has 175 billion parameters. Feldman took a grim view of an increasingly proprietary approach of OpenAI, Google, and Facebook, and he said Cerebras’ goal is to provide a low-cost alternative with its open source AI models.
“The values of the results are getting larger, so OpenAI and Meta and others are closing these models to other companies. That is bad for the ecosystem, small companies, and big companies. I think it is an effort to try and keep these models in a handful of very large companies,” Feldman said.
But he also sympathized with the tech behemoths, which openly shared their AI research and are now trying to recoup the millions or even billions of dollars spent training models. Facebook and Google are spending billions upgrading their computing infrastructure to include GPUs and other accelerator chips that can run AI workloads.
“Some businessperson looked up and said, ‘Why on earth are we sharing these results?’,” Feldman said, later adding “all this progress was made because of openness. And they’re like, well, now there’s a lot of money to be made.”
Cerebras’ seven AI models will execute on any hardware. But the company also wants to use the software to showcase the performance of its AI chips, which are considered among the fastest in the world.
Evolving Like Linux
Analysts said the open source movement is evolving just like the development of Linux, which came out of the need to counter proprietary operating systems. Linux is now the backbone of the Internet and provides the building blocks for cloud native computing.
Intel, which is one of the largest contributors to the Linux kernel, is also providing the plumbing for open source AI development. Some of its tools include OneAPI, which is an open source framework to develop and deploy applications.
“Intel is committed… to fostering an open AI software ecosystem, enabling software optimizations upstream and AI-ML frameworks to promote programmability, portability, and ecosystem adoption,” said Greg Lavender, chief technology officer at Intel, during a roadmap presentation last month.
The chip maker is developing a wide range of chips to run AI applications, including GPUs and accelerators like its Gaudi accelerators. An open source approach to AI could make its chips more attractive to clients. Intel took a similar approach with Linux, where it makes hardware driver contributions to the Linux kernel to make sure its chips are compatible with each new OS release.
But Intel’s AI presence is nothing compared to Nvidia, which made AI computing possible through its GPUs. Today, Nvidia’s GPUs are used to run AI applications deployed by Microsoft and Facebook. Google and Amazon also host Nvidia’s latest Hopper GPUs for clients to run training and inferencing AI applications.
Nvidia wants to monetize its dominance in AI and believes a closed-source approach is the way to get there. The company is using its proprietary hardware and software tools to lock up developers into its ecosystem.
Nvidia’s software development stack, which is called CUDA is already popular among AI developers. Applications written in CUDA only work on the company’s GPUs. Other machine-learning frameworks such as OpenCL and ROCm are available, but getting out of CUDA can be an expensive affair.
The graphics chip maker says it has open sourced libraries to develop vertical AI applications, but those need the company’s GPUs to execute. Intel is trying to cut off that proprietary approach with its SYCL tool, which cuts off CUDA-specific code so applications can run on any CPU, GPU, FPGA, or other accelerators.
Nvidia has also built a services business around AI. Companies can submit their AI needs to Nvidia, which then creates and deploys the application across its GPUs. Nvidia hopes to become an AI software giant and tap into what the company believes could become a $1 trillion market opportunity.
