Amazon’s vision: An AI model for everything

Matt Wood, vice president of product for Amazon Web Services, is at the tip of the spear of Amazon’s response in the escalating AI battle between the tech giants.

Much of the internet already runs on AWS’s cloud services and Amazon’s long game strategy is to create a single point of entry for companies and startups to tap into a rapidly increasing number of generative AI models, both of the open-source and closed-source variety.

Wood discussed this and other topics in an edited conversation below.

Q: Microsoft and Google are both nipping at your heels by offering these huge AI models. How does AWS view this market?

A: I have not seen this level of excitement and engagement from customers since the very earliest days of AWS. We have over 100,000 customers today that routinely use AWS to drive their machine-learning capabilities and these generative AI systems.

One of the interesting differences with these generative models is that they make machine learning easier than ever before to use and apply. We built a capability that we call Bedrock, which provides the very easiest way for developers to build new experiences using this technology on AWS. You just provide a prompt, select which model you want to use, and we give you the answer.

Where we kind of think of things a little differently is that it doesn’t seem that there’s going to be one model to rule them all. As a result, our approach is to take the very best, most promising, most interesting models and to operationalize them so customers can really use them in production. Customers can combine models from Amazon and from third parties in ways that are interesting and novel.

Q: How many models are there now?

A: On Bedrock, we have models from Amazon we call Titan. We provide models from Anthropic, AI21 Labs, which has great support for different languages. We’ve got models from Stability AI, and we’ll have more coming in the future.

Q: So you’re basically curating the best models out there?

A: Indeed. But there’s an old Amazon adage that these things are usually an “and” and not an “or.” So we’re doing both. It’s so early and it’s so exciting that new models are emerging from industry and academia virtually every single week. But some of them are super early and we don’t know what they’re good at yet.

So we have SageMaker JumpStart, which has dozens of foundational models, many of which have already been trained on AWS so they’re already up there. That’s where we’ll have a kind of marketplace that customers can just jump on and start using them. For example, the Falcon model, which is on the top of the leaderboards in terms of capability, was trained on AWS inside SageMaker and today is available inside SageMaker JumpStart.

You can think of JumpStart as the training field for these models and the ones that prove to be really differentiated and capable and popular, we’ll take those and elevate them into Bedrock, where we’ll provide a lot of the operational performance and optimization to deliver low latency, low cost and low power utilization for those models.

Q: If AWS is offering all of these open-source models on its platform, is there concern that AWS could be held responsible for the safety of those models?

A: No. It really helps if you think of them less as this very loaded term of artificial intelligence, and more just applied statistics. It’s just a statistical parlor trick, really. You can kind of think of them a little like a database. We want to enable those sorts of capabilities for customers. And so if you think of it that way, it makes a lot more sense as to where the responsibility lies for their usage. We’re engaged in all of those policy discussions and we’ll see how it plays out.

Q: Who do you think will use SageMaker JumpStart and who will use Bedrock?

A: There’s going to be people who have the skills, interests, and the investment to actually go and take apart these models, rebuild them, and combine them in all these interesting way. It’s one of the benefits of open source models. You can do whatever you want with them.

But then for the vast majority of organizations, they just want to build with these things and want to know it has low latency, is designed for scale and has the operational performance you would expect from AWS, and that the data is secure.

Q: There’s a debate over which is better: Fine tuning/prompting very large, powerful and general large language models for specific purposes, or using more narrowly focused models and fine tuning them even further. It sounds like you believe the latter option is going to win.

A: You can’t fine tune GPT-4. What we found is that in the enterprise, most of those customers have very large amounts of existing private data and intellectual property. And a lot of the advantages and the opportunity that they see for generative AI is in harnessing that private data and IP into new internal experiences, or new product categories or new product capabilities.

The ability to take that data and then take a foundational model and just contribute additional knowledge and information to it very quickly and very easily, and then put it into production very quickly and very easily, then iterate on it in production very quickly and very easily. That’s kind of the model that we’re seeing.

Q: Can you give me any customer examples that stand out?

A: It’s super early and we’re still in limited preview with Bedrock. What has struck me is just the diversity and the breadth of the use cases that we’re seeing. A lot of folks are using these in the kind of unsexy but very important back end.

So personalization, ranking, search and all those sorts of things. We’re seeing a lot of interest in expert systems. So chat and question-answer systems. But we’re also seeing a lot of work in decision-making support. So, decomposing and solving more complicated problems and then automating the solution using this combination of language models under the hood.

Q: What’s the vision for who will best be able to take advantage of these products? Do you see a possibility that startups could basically just form a company around these APIs on Bedrock?

A: There are going to be waves and waves of startups that have an idea or an automation that they want to bring into companies, or an entirely new product idea that’s enabled through this.

An interesting area is larger enterprises that are very text heavy. So anywhere there is existing text is fertile soil for building these sorts of systems.

And what’s super interesting is that we’re seeing a lot of interest from organizations in regulated fields that maybe traditionally don’t have the best reputation for leaning into or being forward-thinking in terms of cutting-edge technologies. Banking, finance, insurance, financial services, healthcare, life sciences, crop sciences.

They are so rich in the perfect training data. Volumes and volumes of unstructured text, which is really just data represented in natural language. And what these models are incredibly capable at is distilling the knowledge and the representation of that knowledge in natural language, and then exposing it in all of these wonderful new ways that we’re seeing.

Q: Were you surprised about how quickly this all happened?

A: ChatGPT may be the most successful technology demo since the original iPhone introduction. It puts a dent in the universe. Nothing is the same once you see that, I think it causes, just like the iPhone, so many light bulbs to go off across so many heads as to what the opportunity here was, and what the technology was really ready for.

The real explosive growth is still ahead of us. Because customers are going through those machinations now. They’re starting to find what works and what doesn’t work, and they’re starting to really move very quickly down that path.

Q: AWS already changed the world once by enabling the cloud, which enabled everything from Uber to Airbnb. How is the world going to change now that internet companies have access to potentially thousands of powerful models through the cloud?

A: I can’t think of a consumer application or customer experience that won’t have some usage for the capabilities inside these models. They’ll drive a great deal of efficiency, a great deal of reinvention. Some of that’s going to move from point and click to text. These things are just so compelling in the way they communicate through natural language.

Then the question is how do these models evolve? By using these models, they get better. They learn what works and what doesn’t work. These capabilities will get better much more quickly than we’ve ever seen before.

Q: These models create the possibility for a new kind of internet, where instead of going to Google and getting a set of links, you’re kind of talking to different agents and behind the scenes all these transactions are happening. And a lot of that is going to happen on AWS. That has huge implications for search and advertising. Is that how you see this evolving?

A: You’re still going to have to go to a place to start your exploration in any domain, whether it is where you want to buy, or what you want to learn about. I don’t know if there’s going to be some big, single point of entry for all of this. That seems very unlikely to me. That’s not usually how technology evolves. Usually technology evolves to be a lot more expansive. I think there’s a world in which you see not one or two entry points, but thousands, tens of thousands, or millions of new entry points.