Amazon wants to host companies’ custom generative AI models
AWS, Amazon‘s leading cloud computing service, aims to establish itself as the premier platform for companies seeking to host and optimize their custom generative AI models. In line with this goal, AWS has unveiled Custom Model Import, a new feature introduced in preview as part of Bedrock, AWS’ suite of enterprise-focused generative AI services. This feature enables organizations to seamlessly import and access their proprietary generative AI models as fully managed APIs within the AWS infrastructure.
Once imported, companies’ proprietary models receive the same level of support and infrastructure as other generative AI models available in Bedrock’s library, such as Meta’s Llama 3 or Anthropic’s Claude 3. Moreover, organizations gain access to a suite of tools designed to expand their models’ capabilities, fine-tune their performance, and implement safeguards to address potential biases.
Vasi Philomin, Vice President of generative AI at AWS, emphasized the significance of this capability in an interview with TechCrunch, stating, “This Custom Model Import capability allows them to bring their own proprietary models to Bedrock and see them right next to all of the other models that are already on Bedrock — and use them with all of the workflows that are also already on Bedrock, as well.” This integration ensures seamless integration and compatibility with existing workflows, enhancing efficiency and accessibility for AWS customers leveraging generative AI technology.
Importing custom models
According to a recent poll conducted by Cnvrg, Intel’s AI-focused subsidiary, the majority of enterprises are adopting generative AI by developing their own models and customizing them to suit their specific applications. However, these enterprises cite infrastructure, particularly cloud compute infrastructure, as their primary obstacle to deployment.
In response to this need, AWS has introduced Custom Model Import, a feature aimed at addressing infrastructure challenges while keeping pace with competitors in the cloud computing space. Amazon CEO Andy Jassy hinted at this initiative in his recent annual letter to shareholders.
Google’s Vertex AI, a counterpart to AWS Bedrock, has long enabled customers to upload, customize, and serve generative AI models through APIs. Similarly, Databricks has provided toolsets for hosting and fine-tuning custom models, including their recently launched DBRX platform.
When questioned about what sets Custom Model Import apart, Vasi Philomin emphasized its comprehensive model customization capabilities, positioning Bedrock as offering a broader range and depth of options compared to competitors. He highlighted Bedrock’s versatility in serving models and its extensive workflows, noting that tens of thousands of customers currently leverage the platform. Philomin underscored the ability for users to experiment across multiple models using the same workflows and seamlessly transition them to production, all from a unified platform.
In essence, Custom Model Import and Bedrock aim to streamline the deployment and customization of generative AI models while providing a robust infrastructure and workflow environment, setting AWS apart from its competitors in the cloud computing landscape.
So what are the alluded-to model customization options?
Philomin emphasizes two key features within Bedrock that aim to enhance model safety and evaluation: Guardrails and Model Evaluation. Guardrails enable Bedrock users to set thresholds to filter model outputs for problematic content such as hate speech, violence, and sensitive personal or corporate information. Given the notorious tendency of generative AI models to produce problematic outputs, including the inadvertent leakage of sensitive information, Guardrails serve as a critical tool for mitigating risks. Additionally, Model Evaluation allows customers to test the performance of models across various criteria, providing insights into their effectiveness and potential areas for improvement.
Both Guardrails and Model Evaluation have transitioned from preview to general availability after undergoing several months of testing and refinement.
It’s important to note that Custom Model Import currently supports only three model architectures: Hugging Face’s Flan-T5, Meta’s Llama, and Mistral’s models. While competing services like Vertex AI and Microsoft’s AI development tools on Azure offer similar safety and evaluation features, what sets Bedrock apart is AWS’ Titan family of generative AI models.
Coinciding with the release of Custom Model Import, there have been notable developments within AWS’ Titan family of generative AI models, further enhancing the platform’s capabilities and offerings in this space.
Upgraded Titan models
AWS has announced the general availability of Titan Image Generator, its text-to-image model, following its launch in preview last November. This tool remains capable of generating new images based on text descriptions or customizing existing images, such as replacing backgrounds while preserving the main subjects.
According to Philomin, Titan Image Generator in its general availability state boasts increased “creativity” compared to its preview version, although specific details on this enhancement were not provided.
During the model’s debut last November, AWS was relatively vague regarding the training data used for Titan Image Generator. This opacity is common among vendors, as training data is often considered a competitive advantage and closely guarded. Additionally, divulging training data details could potentially expose vendors to intellectual property-related lawsuits, particularly in cases where generative AI tools replicate artists’ styles without explicit permission.
Philomin disclosed that AWS employs a combination of proprietary and licensed data for training Titan Image Generator, paying licensing fees to copyright owners for the use of their data. However, this level of detail may still fall short of satisfying content creators and AI ethicists advocating for greater transparency surrounding the training of generative AI models.
In absence of detailed transparency regarding its training data, AWS has reiterated its commitment to providing an indemnification policy to protect customers in case a Titan model like Titan Image Generator produces a potentially copyrighted training example. This policy aligns with similar offerings from competitors like Microsoft and Google, providing customers with reassurance and legal coverage.
To combat the ethical concerns posed by deepfakes, images generated with Titan Image Generator will continue to be equipped with a “tamper-resistant” invisible watermark, as implemented during the preview phase. Philomin emphasized that the watermark has been enhanced in the general availability release to withstand compression and other image manipulations more effectively.
Transitioning to a less contentious topic, Philomin was asked about AWS’ stance on video generation, given the growing interest and investment in this technology by competitors like Google and OpenAI. While Philomin did not explicitly confirm AWS’ exploration of video generation, he hinted at ongoing discussions with customers and encouraged observers to stay tuned for potential developments.
In other Titan-related news, AWS unveiled the second generation of its Titan Embeddings model, Titan Text Embeddings V2. This model converts text into numerical representations, known as embeddings, to support search and personalization applications. Philomin claimed that Titan Text Embeddings V2 offers improved efficiency, cost-effectiveness, and accuracy compared to its predecessor, reducing storage requirements by up to four times while maintaining 97% accuracy. However, real-world testing will ultimately determine the veracity of these claims.