HomeArtificial IntelligenceAmazon desires to host corporations' custom generative AI models

Amazon desires to host corporations' custom generative AI models

AWS, Amazon's cloud computing company, goals to grow to be the go-to destination for corporations to host and optimize their custom generative AI models.

Today, AWS announced the launch of Custom Model Import (in preview), a brand new feature in Bedrock, AWS' enterprise-focused suite of generative AI services. This feature allows corporations to import and access their internal generative AI models as fully managed APIs.

Once imported, the businesses' proprietary models profit from the identical infrastructure as other generative AI models in Bedrock's library (e.g. Llama 3 from Meta or Claude 3 from Anthropic). They also receive tools to expand and refine their knowledge and take safeguards to mitigate their biases.

“There have been AWS customers who’ve refined or built their very own models outside of Bedrock using other tools,” said Vasi Philomin, vice chairman of generative AI at AWS, in an interview with TechCrunch. “This custom model import feature allows them to bring their very own proprietary models to Bedrock and display them right next to all other models that exist already on Bedrock – and in addition use them with any workflows that also exist already on Bedrock are .”

Import custom models

According to a recent Opinion poll According to Cnvrg, Intel's AI-focused subsidiary, most corporations approach generative AI by constructing their very own models and refining them for his or her applications. According to the survey, corporations see infrastructure, including cloud computing infrastructure, as their biggest barrier to deployment.

With Custom Model Import, AWS goals to satisfy this need while keeping pace with cloud competition. (Amazon CEO Andy Jassy indicated as much in his most up-to-date annual letter to shareholders.)

For a while now, Vertex AI, Google's analogue to Bedrock, has allowed customers to upload generative AI models, customize them, and deploy them via APIs. Databricks has also long offered toolsets for hosting and optimizing custom models, including its own recently released DBRX.

When asked what sets Custom Model Import apart, Philomin responded that it – and by extension Bedrock – offers a greater breadth and depth of model customization options than the competition, adding that “tens of hundreds” of shoppers use Bedrock today.

“First, Bedrock offers customers multiple options for coping with serving models,” Philomin said. “Second, we have now a complete suite of workflows around these models – at once Customers can are right next to all the opposite models we have already got on offer. An vital aspect that almost all people like about it’s the power to experiment with several different models and the identical workflows after which actually put them into production from the identical place.”

So what are the suggested model customization options?

Philomin points to Guardrails, which allows Bedrock users to configure thresholds to filter — or at the very least try to filter — model outputs for things like hate speech, violence, and personal personal or corporate information. (Generative AI models are notorious for going off the rails in problematic ways, including leaking sensitive information; AWS's models have been there no exceptions.) He also highlighted Model Rating, a Bedrock tool that enables customers to check how well a model – or several – performs on a particular set of criteria.

Both Guardrails and Model Evaluation are actually generally available after several months of preview.

I have to note here that custom model import currently only supports three model architectures: Hugging Faces Flan-T5, Metas Llama and Mistrals models. Additionally, Vertex AI and other services competing with Bedrock, including Microsoft's AI development tools on Azure, offer kind of comparable security and assessment capabilities (see Azure AI Content Safety, Model evaluation in vertex and so forth).

What makes Bedrock unique, nevertheless, is AWS's Titan family of generative AI models. And coinciding with the discharge of Custom Model Import, there have been several notable developments on this front.

Improved Titan models

Titan Image Generator, AWS's text-to-image model, is now generally available after launching in preview last November. As before, Titan Image Generator can create recent images from a text description or customize existing images – for instance, by swapping the background of a picture while retaining the topics within the image.

Compared to the preview version, Titan Image Generator in GA can produce images with more “creativity,” Philomin said, without going into details. (Your guess as to what meaning is nearly as good as mine.)

I asked Philomin if he could provide me with more details about how the Titan Image Generator was trained.

When the model debuted last November, AWS didn’t provide details about exactly what data was used to coach the Titan Image Generator. Few providers willingly disclose such information; They view training data as a competitive advantage and subsequently all the time keep it and the associated information at hand.

Training data details are also a possible source of mental property lawsuits, one other incentive to disclose much. Several cases before the courts reject providers' fair use objections, arguing that text-to-image tools reproduce artists' styles without the artist's express permission and permit users to create recent works, that resemble the artist's originals, for which artists don’t receive permission for payment.

Philomin would just tell me that AWS uses a mixture of first party and licensed data.

“We have a mixture of proprietary data sources, but we also license lots of data,” he said. “We actually pay copyright holders royalties to make use of their data, and we have now contracts with several of them.”

It is more detailed than in November. But I actually have a sense Philomin's answer won't satisfy everyone, especially the content creators and AI ethicists who advocate for greater transparency in generative AI model training.

In lieu of transparency, AWS says it’s going to proceed to supply an indemnification policy that covers customers within the event that a Titan model like Titan Image Generator regurgitates (i.e. spits out a mirror copy of) a potentially copyrighted training example. (Several competitors, including Microsoft and Google, offer similar guidelines for his or her imaging models.)

To counter one other urgent ethical threat – deepfakes – AWS says that images created with Titan Image Generator could have a “tamper-proof” invisible watermark as previewed. Philomin says that the watermark within the GA version has been made more immune to compression and other image editing and manipulation.

Moving into less controversial territory, I asked Philomin whether AWS, like Google, OpenAI, and others, is exploring video generation given the keenness for (and investment in) the technology. Philomin didn't say that, but he didn't need to imply anything more.

“Of course, we’re continuously on the lookout for recent features that customers want, and video generation definitely comes up in conversations with customers,” Philomin said. “I ask you to hold in there.”

In recent Titan-related news, AWS released the second generation of its Titan Embeddings model, Titan Text Embeddings V2. This model converts text into numerical representations, called embeddings, to enable search and personalization applications. The first generation Embeddings model did this too, but AWS claims that Titan Text Embeddings V2 is more efficient, cost-effective and accurate overall.

“What the Embeddings V2 model does is reduce the overall memory (required to make use of the model) by 4 times while maintaining 97% accuracy,” Philomin claimed, “outperforming other comparable models.”

We'll see if real tests confirm this.


Please enter your comment!
Please enter your name here

Must Read