The cloud provider AWS would like to support you in the development and integration of generative AI applications. To this end, the company launches several solutions aimed at both training and inference.

AWS doesn’t feel comfortable sitting on the sidelines. Over the past few weeks and months, Microsoft in particular, and to a lesser extent Google, have been running AI scoop after AI scoop. Microsoft also has preferential access to the company’s GPT models thanks to its multi-billion dollar stake in OpenAI. The dependency between the generative AI models and Azure is definitely a win for the manufacturer. So it’s high time for AWS to respond.

This comes in the form of four different services. AWS wants to help you train models and perform inference. The big absentee from the whole story is OpenAI and with it the GPT-4 model, although Amazon offers the necessary alternatives for it.

bedrock

First and foremost, AWS introduces Amazon Bedrock. This is a solution that allows you to develop AI applications at scale based on models from AI21Labs, Anthropic, Stability AI and Amazon Titan. Bedrock allows customers to train models from the list based on their own data. This is done based on APIs.

The available models are suitable for both text generation and images. Under the Amazon Titan banner, AWS itself is launching two new ones large language models (LLM’s). From AI21Labs, AWS puts the Jurassic 2 models in the spotlight. They are good for text and are particularly good with European languages ​​such as Spanish, French, German, Portuguese and also Dutch.

Through Bedrock, users can quickly customize models to their liking. All you have to do is point out a handful of examples to adapt the model to your company’s needs. AWS cites the example of a fashion retailer giving a model access to the slogans of the most successful handbag marketing campaigns in recent years. Based on this information, the model can then generate a new campaign that takes into account the success factors.

instances

Then AWS introduces new instances for both training and inference of generative AI. In conclusion, these are Amazon EC2 Inf2 instances built around the Inferentia2 chips. They are optimized for generative AI and are said to have four times higher throughput and ten times lower latency than their predecessors. AWS claims that the price/performance ratio has improved by 40 percent.

The Trn1n instances, on the other hand, are intended for training. These instances also have self-made accelerators on board: the Trainium chips. Users can combine instances into so-called Ultra Cluster with up to 30,000 such Trainium chips on board. This configuration accounts for more than six exaflops of AI processing power.

co-pilot competitor

Finally, AWS launches Amazon CodeWhisperer. This is a competitor for GitHub Copilot which, like Microsoft’s competitor, can help programmers by suggesting code. Of course, CodeWhisperer also uses generative AI and can already handle Python, Java, JavaScript, TypeScript and C#.

CodeWhisperer launched last year but is now generally available. With this broader rollout, AWS also offers compatibility with ten new languages, including Kotlin, Rust, PHP, and SQL.

Relevant Announcements

AWS’ list of announcements is broad and shows that the cloud provider is determined to miss the AI ​​train. AWS has many unique in-house capabilities, including the ability to custom build its own powerful accelerator chips. The instances for training and inference make good use of it. The announcements also come just in time, as AWS just announced an AI accelerator program.

Bedrock also seems to offer significant added value. Loved the API approach and the ability to customize models based on a limited number of samples. It doesn’t take much imagination to see the practical added value in it.

Additionally, AWS is right that OpenAI, with its GPT models, is not the only one with powerful LLM that can lay the foundation for generative AI. In this way, the varied offer offers added value, even if we can’t shake the impression that the offer would be more complete by working with OpenAI. ChatGPT is the hype of the moment. This company’s close collaboration with Microsoft and Azure currently gives this main competitor a bit more authority and credibility when it comes to AI training and inference in the cloud.