AWS announces Amazon Bedrock and Titan models

Background

In April 2023 AWS set out to give its cloud customers one managed place to reach foundation models from several companies, rather than a single provider's models.^[1]

What happened

On 13 April 2023 AWS announced Amazon Bedrock, a managed service offering API access to foundation models. The announcement was written by Swami Sivasubramanian.^[1] Launch model providers were AI21 Labs, Anthropic, Stability AI and Amazon itself.^[1]

Amazon itself was one of the providers. AWS announced two Amazon Titan models, a generative LLM and an embeddings model, in preview with a few customers.^[1] The Titan models were Amazon's own entries in the service, alongside the models from the outside companies.

The launch was limited. Bedrock started in limited preview, and it is serverless, with customer data kept private and encrypted.^[1]

Key details

The same day, AWS made two kinds of hardware generally available. One was the Trainium-based Trn1n instances, which provide 1,600 Gbps of networking. The other was the Inferentia2-based Inf2 instances.^[1] Trainium and Inferentia are Amazon's custom AI chips, and these instances let customers run and train models on that hardware.

Bedrock became generally available on 28 September 2023.^[2][3] At that point the service had grown beyond its first set of providers, and AWS also gave the Titan Embeddings model its general release at the same time.^[2]

Impact

The launch set out an approach in which AWS offers models from several companies through one managed service. It placed Amazon's own models next to those from partners such as Anthropic.^[1]

For readers, the event shows how a major cloud provider chose to organise AI model access: as a managed platform that gives customers a choice of models, with privacy and encryption as built-in features, and with the company's own models as one option among several.^[1][2]

Some terms in the announcement may be unfamiliar. An API, or application programming interface, is a way for one program to send requests to a service and receive results, so a developer can use a model without running it on their own hardware. A managed service is one where the provider runs the underlying infrastructure, such as servers and updates. Serverless means the customer does not have to provision or manage the servers that run the model.^[1]