Amazon plans to host the company’s custom generative AI Models

Amazon’s cloud computing field, AWS, wants to become the place where businesses host and improve their own generative AI models.

AWS launched Custom Model Import today as a preview feature. Bedrock is a new service in AWS’s suite of generative AI services, designed for businesses. With this feature, businesses can import and use their generative AI models as fully managed APIs.

Companies can use the same infrastructure as other generative AI models in Bedrock’s library, such as Meta’s Llama 3 or Anthropic’s Claude 3, once they import their models. Furthermore, Bedrock will provide them with tools to improve their learning, skills, and safeguards to mitigate their biases.

Vasi Philomin, VP of generative AI at AWS, said, “There have been AWS customers that have been fine-tuning or building their models outside of Bedrock using other tools.” “This Custom Model Import feature allows them to add their models to Bedrock and see them next to all the other models that are already there. They can also use these models with all the workflows that are already on Bedrock.”

Custom import models

A recent poll by Cnvrg, Intel’s AI-focused subsidiary, found that most businesses are using generative AI by creating their models and making them better for their needs. According to the poll, businesses say that infrastructure, such as cloud computing infrastructure, is the thing that stops them from deploying.

With Custom Model Import, AWS hopes to meet that need and keep up with other cloud providers. (Amazon CEO Andy Jassy hinted at this in his most recent letter to shareholders.)

Vertex AI, which is Google’s version of Bedrock, has let users upload generative AI models, change them, and serve them through APIs for a while now. Databricks has also been developing tools to host and change custom models for a long time. Its own DBRX was recently released.

When asked what makes Custom Model Import unique, Philomin said that it, and by extension, Bedrock, provides more custom model options than its competitors. He also said that “tens of thousands” of users are currently using Bedrock.

Philomin said, “First, Bedrock gives customers a lot of different ways to deal with serving models.” “The second thing is that we have built a lot of workflows around these models, and now customers can see them next to all the other models we already have.” Most people like this because it allows them to test out different models using the same workflows before sending them to production from the same location.

What are the mentioned model customization options?

Philomin talks about Guardrails, a tool that allows Bedrock users to set limits on what kinds of content models can produce, such as violent content, hate speech, and private personal or business data. (AWS’ models have been no different from other generative AI models in that they have gone off the rails in bad ways, such as leaking private information.) Besides that, he talked about model evaluation, a Bedrock tool that allows users to check how well one or more models meet certain criteria.

After being available to select users for a few months, both Guardrails and Model Evaluation are now generally available.
Just so you know, Custom Model Import only works with three model architectures right now: Mistral’s models, Hugging Face’s Flan-T5, and Meta’s Llama. Another thing is that Vertex AI and other similar services, such as Microsoft’s AI development tools on Azure, offer similar safety and evaluation features (see Azure AI Content Safety, Vertex model evaluation, and so on).

The Titan family of generative AI models from AWS, on the other hand, is only available in Bedrock. Since its release, Custom Model Import has undergone several significant changes.

Updated Titan models

AWS released Titan Image Generator as a preview in November and has now made it available to everyone. As before, Titan Image Generator can use a text description to make new images or change the way existing images look. For example, it can change an image’s background while keeping the people in it.

This version of Titan Image Generator in GA can make pictures with more “creativity,” according to Philomin, who didn’t go into more detail.

When the model first came out in November of last year, AWS wasn’t clear about what data it used to train the Titan Image Generator. Few vendors are willing to share this kind of information. They see training data as a competitive advantage and keep it, as well as information about it, secret.

Training data details could also lead to intellectual property lawsuits, which is another reason not to give away too much. Several cases currently going through the courts don’t accept vendors’ fair use arguments. These cases say that text-to-image tools copy artists’ styles without their permission and let users make new works that look like the originals of artists without paying the artists.

Only Philomin would say that AWS uses both first-party data and licensed data.

“We license a lot of data as well as have our independent data sources,” he said. “We do pay copyright owners license fees to use their data, and we have agreements with a number of them.”

We learned more about this now than in November. However, I believe Philomin’s response may not satisfy everyone, particularly content creators and AI ethicists who desire greater transparency regarding the training of generative AI models.

Rather than being transparent, AWS maintains its indemnity policy, which safeguards customers if a Titan model, like the Titan Image Generator, replicates a training example that may be subject to copyright protection. (Several competitors, such as Microsoft and Google, have similar rules for their image-making software.)

AWS says that images made with the Titan Image Generator will have a “tamper-resistant” invisible watermark, just like the preview. This is to protect against deepfakes, which are another serious ethical threat. Philomin states in the GA release that compression and other image changes and tweaks are less likely to damage the watermark.

As a less controversial topic, I asked Philomin if AWS is looking into video generation like Google, OpenAI, and others because there is a lot of interest and money in the technology. Philomin didn’t say that AWS wasn’t, but he also wouldn’t say much else.

“Of course, we’re always looking to see what new features customers want,” Philomin said. “Video generation is something that comes up when we talk to customers.” “Please pay attention.”

Titan Text Embeddings V2 is the second generation of AWS’s Titan Embeddings model. This is the last piece of Titan news. This model transforms text into numerical representations, enabling search and personalization applications. The first-generation Embeddings model also did that, but according to AWS, Titan Text Embeddings V2 is faster, cheaper, and more accurate overall.

Philomin said, “The Embeddings V2 model cuts the total amount of storage needed by the model by up to four times while keeping 97% of the accuracy.” This makes it “outperform other models that are comparable.”