|
Listen to this story
|
Amazon Web Services recently announced that it would begin offering access to generative algorithms Bloom and Stable Diffusion in Sagemaker Jumpstart – their service for open-source, pre-trained, deployment-ready algorithms. These models have become fairly well-known through the generative AI space.
The creators of Stable Diffusion, a text-to-image generative algorithm, claim it is a “collective effort to create a single file that compresses the visual information of humanity into a few gigabytes”. Stable Diffusion can also be applied to other common generative AI use cases such as inpainting, outpainting, and image-to-image translations guided by a text prompt. It uses the diffusion method to generate images, wherein the algorithm removes noise from the image progressively until the final image resembles the prompt given by the user.
Bloom is a multilingual model which contains over 176 billion parameters. It is the largest collaboration of AI researchers ever involved in a project, and was trained on a behemoth 386 GPUs for 3.5 months. It can solve a variety of logical thinking problems in mathematics, coding statistics, and more. It can also generate text in 46 natural languages and 13 programming languages, making it one of the most robust large language models available today.
Why haven’t Microsoft and Google caught up?
Even though tech giants have spent billions of dollars researching and developing generative AI algorithms, they are still reluctant to open it up to the public. Arguably, Microsoft Azure was one of the first platforms to make these services available to enterprise customers. In 2020, it obtained an exclusive license for the GPT-3 language model, which they then made part of the Azure OpenAI platform. The service provides REST API access to many OpenAI language models, including GPT-3.
Even though Azure offers API access to OpenAI’s powerful algorithms, it is not open to all. Those who wish to get access to the service must first go through an application process and then a use case review to ensure that it is a low-risk scenario. As we can see, this is a continuing trend among cloud service providers as they realize the societal impact of releasing highly accurate generative AI algorithms to the public.
Google, on the other hand, has completely closed off the results of its AI research to the general public. They created a text-to-image generator called Imagen, which was recently updated to be able to generate video as well. However, the model has been kept a “trade secret” by Google, with the researchers quoting the possible societal impacts of releasing such an algorithm to the public. Citing “potential risks of misuse”, they decided against making the code public, and are yet to offer it as a service on their cloud platform.
The dark side of cloud-powered generative AI
Microsoft states that while the generative models do have considerable potential benefits, they also have a huge potential to be misused to create large volumes of harmful content. Moreover, they acknowledge that the datasets used to train these algorithms also have inherent biases due to their uncurated nature. Microsoft has also taken a strong position against the irresponsible use of certain AI algorithms, seeing the probable negative consequences of unchecked AI bundled with easy-to-use cloud computing power.
Google has also blocked access to Imagen with the same reasoning. In its blog post regarding the use of the algorithm, it stated that while it is working on a framework to balance the value of external auditing and the risks of unrestricted open access, it also found that Imagen encodes several social biases and stereotypes. In its testing, this resulted in a bias of generating images of people with a lighter skin tone, a tendency to portray women in professions reinforced by gender stereotypes, and many other issues, which led them to hold off on publishing the model.
When looking at the extreme attitude other cloud service providers have approached generative AI with, it is puzzling to see AWS open both Bloom and Stable Diffusion up to the public with no restrictions. Moreover, they are also ready to deploy in any enterprise setting, with the power of AWS nigh-infinite infrastructure behind it. They have also pre-trained the model, allowing engineers to run it as is for inferencing or further fine-tune it to adjust the type of content they want to create.
For instance, a user with malicious content can use the already-capable Bloom — now hosted on the cloud with the ability to scale — to create voluminous amounts of content targeted towards a certain group of people, which can then be used to further a maligned goal. On the other hand, Stable Diffusion can be used to create offensive or hateful content images, further compounding the inherent biases that such models already have. Offering such content on the open market seems like a recipe for disaster for AWS. It remains to be seen what impact this move will have on the world as a whole.
