Generative AI on AWS

If you don’t have a clear use case in mind, it can be useful to start with a business objective and work backwards from there. This is how we typically structure engagements, and we’ve found it to be a powerful tool for identifying opportunities and ranking their value. To be perfectly transparent, not every problem is a good candidate for gen AI capabilities, but when you have one, the upside can be huge.

Consider the architecture you’re working with and the nature of the guarantees you need. Some model makers will offer that no data is stored once you are paying for an Enterprise plan, for instance. But for some data, you may find that the notion of sending it over the public internet is untenable. If that’s your situation, consider services like Amazon Bedrock, which allow you to keep your data and access the model entirely within AWS infrastructure and therefore maintain the associated security and privacy guarantees.

We can perform a Total Cost of Ownership analysis, to help you understand your infrastructure costs once your solution is up and running. But the key to developing a cost-efficient solution is to use the proper techniques for your problem. For example, retrieval-augmented generation can vastly reduce the size of prompts you need to use. You may also want to consider model fine-tuning, quantizing your model, and training it to your specific use case to avoid the overhead of running a large model if you don’t need that level of generalized power.

Yes, without training or providing your model with the exact information you're querying through a technique like retrieval-augmented generation, you may find your solution “hallucinates” an answer or doesn’t alert you that it doesn’t have the specific answer. How a model reacts in each case depends on how it was developed, the specifics of your prompts, and the surrounding solution. As part of our methodology when building a solution, we tune performance until you’ve reached an acceptable error rate, ideally zero. And we may also suggest developing with a different model if you need highly-specific outputs.

The honest answer is: not always. While models are highly flexible problem solvers and their capabilities are continuously evolving, not every problem is efficiently solved with a large language model as part of the implementation. With that said, when you have a good fit for this technology, the upside can be tremendous. In general, the best use cases involve enhancing your existing capabilities, removing human intervention or manual work from a process, accentuating customer experience, or quickly generating outputs for prototypes of work like code, text, or images.

Experimenting with prompts can be a great way to start early prototyping. AWS SageMaker and SageMaker Studio also provide a platform for trying different models, quickly launching infrastructure, and working with an interactive notebook to try solutions and integrate your data. This is a great starting place for building a proof of concept. But if that sounds too daunting or you’ve reached your technical limits with that approach, consider engaging with a partner like Mission. We’ll get up to speed on the state of your current work, evaluate your objectives, and help you build a roadmap for how to get there. If you want our help building the implementation, infrastructure, and algorithms, we can do that too!