AWS, Amazon’s cloud computing business, wants to become the preferred place for companies to host and refine their personalized generative AI models.
Today, AWS announced the launch of Custom Model Import (preview), a new feature in Bedrock, AWS’ suite of generative AI services for businesses, that allows organizations to import and ‘access their internal generative AI models as fully managed APIs. .
The companies’ proprietary models, once imported, benefit from the same infrastructure as other generative AI models in Bedrock’s library (e.g. Llama 3 from Meta, Claude 3 from Anthropic), including tools to expand their knowledge, refine them and implement safeguards to mitigate their biases. .
“Some AWS customers have tweaked or built their own models outside of Bedrock using other tools,” Vasi Philomin, vice president of generative AI at AWS, told TechCrunch in an interview. “This custom template import feature allows them to bring their own proprietary templates to Bedrock and see them right next to all the other templates already on Bedrock – and use them with any workflows also already there on Bedrock. .”
Importing custom templates
According to a recent survey from Cnvrg, Intel’s AI subsidiary, the majority of companies are approaching generative AI by creating their own models and refining them based on their applications. According to the survey, these same companies say they view infrastructure, including cloud computing infrastructure, as their biggest barrier to deployment.
With Custom Model Import, AWS aims to rush to meet the need while keeping pace with its cloud competitors. (Amazon CEO Andy Jassy foreshadowed this in his recent annual letter to shareholders.)
For some time, Vertex AI, Google’s Bedrock equivalent, has allowed customers to upload generative AI models, customize them, and serve them through APIs. Databricks has also long offered toolsets for hosting and modifying custom models, including its own recently released DBRX.
When asked what sets Custom Model Import apart, Philomin claimed that it – and by extension Bedrock – offers a wider range and depth of model customization options than the competition, adding that ” tens of thousands of customers now use Bedrock.
“First, Bedrock offers customers multiple ways to manage service models,” Philomin said. “Secondly, we have a whole bunch of workflows around these models – and now customers can sit right next to all the other designs we already have available. A key element that most people like is the ability to experiment with several different models using the same workflows and then put them into production from the same place.
So, what are the model customization options discussed?
Philomin points to Guardrails, which allows Bedrock users to configure thresholds to filter — or at least attempt to filter — model results for things like hate speech, violence, and private personal or corporate information. (Generative AI models are known to derail in problematic ways, especially when sensitive information is leaked; AWS’s are no exception.) He also highlighted Model Evaluation, a Bedrock tool that customers can use to test the effectiveness of one – or several – models. perform according to a given set of criteria.
Guardrails and Model Evaluation are now generally available after a multi-month preview.
I feel obligated to note here that importing custom models only supports three model architectures at the moment – Hugging Face’s Flan-T5, Meta’s Llama and Mistral models – and that Vertex AI and others Bedrock’s competing services, including Microsoft’s AI development tools on Azure, offer more or less comparable security and assessment features (see Azure AI Content Safety, model assessment in Vertex, etc.).
What East However, AWS’s Titan family of generative AI models is unique to Bedrock. And – coinciding with the release of Custom Model Import – there are several notable developments on this front.
Improved Titan Models
Titan Image Generator, AWS’s text-to-image model, is now available to everyone after its preview launch last November. As before, Titan Image Generator can create new images from a text description or customize existing images, for example by replacing the background of an image while retaining the subjects in the image.
Compared to the preview version, Titan Image Generator in GA can generate images with more “creativity,” Philomin said, without going into details. (Your idea of what that means is as good as mine.)
I asked Philomin if he had more details to share about how Titan Image Generator was formed.
When the model debuted last November, AWS wasn’t sure what data it was using to train Titan Image Generator. Few sellers readily reveal such information; they view training data as a competitive advantage and therefore keep it and related information close to their chest.
The details of training data are also a potential source of intellectual property-related lawsuits, which also discourages revealing much. Several court cases reject sellers’ fair use defenses, arguing that text-to-image conversion tools reproduce artists’ styles without the artists’ explicit permission and allow users to generate new works that resemble the originals artists for which the artists receive no payment. .
Philomin would only tell me that AWS uses a combination of proprietary and licensed data.
“We have a combination of proprietary data sources, but we also license a lot of data,” he said. “We actually pay licensing fees to copyright owners to use their data, and we have contracts with several of them.
It’s more detailed than from November. But I have a feeling Philomin’s answer won’t satisfy everyone, especially content creators and AI ethicists who advocate for greater transparency when it comes to training generative AI models .
In lieu of transparency, AWS says it will continue to offer a compensation policy that covers customers in the event that a Titan model such as Titan Image Generator regurgitates (i.e. spits out a mirrored copy) of a training example potentially protected by copyright. (Several competitors, including Microsoft and Google, offer similar policies covering their image generation models.)
To address another pressing ethical threat – deepfakes – AWS says images created with Titan Image Generator will, as in the preview, be accompanied by an invisible “tamper-proof” watermark. Philomin says the watermark has been made more resilient in the GA version to compression and other image modifications and manipulations.
Moving into less controversial territory, I asked Philomin if AWS – like Google, OpenAI and others – was exploring video generation given the enthusiasm for (and investment in) the technology. Philomin did not say that AWS was not it… but he didn’t want to say anything more.
“Obviously, we’re always looking to see what new capabilities customers want to have, and video generation is definitely coming up in customer conversations,” Philomin said. “I would ask you to stay tuned.”
In one last bit of Titan-related news, AWS has released the second generation of its Titan Embeddings model, Titan Text Embeddings V2. Titan Text Embeddings V2 converts text into digital representations called embeddings to power search and personalization applications. The same was true for the first generation Embeddings model, but AWS claims that Titan Text Embeddings V2 is more efficient, more cost-effective, and more accurate overall.
“The Embeddings V2 model reduces the overall storage (required to use the model) by up to four times while maintaining 97% of the accuracy,” claimed Philomin, “outperforming other comparable models.”
We’ll see if actual testing bears this out.
techcrunch
