AWS opened the door for developers to import tailored AI applications into its cloud-based Bedrock managed service in April. Custom Model Import for Amazon Bedrock is available under preview in the hyperscaler’s U.S. East region.
The feature lets enterprises upload models built on open architectures and fine-tuned outside the platform for domain-specific use cases, the company said. The Bedrock multimodel marketplace already offers access to commercial models trained by AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI and Amazon along with guardrail and evaluation tools.
“Custom model import is a sneaky big launch that satisfies a customer request,” Amazon President and CEO Andy Jassy said Tuesday during a Q1 2024 earnings call.
The feature was designed for “companies who prefer not to build models from scratch but rather seek to leverage an existing large language model or customize it with their own data,” Jassy said.
AWS and its hyperscaler competitors are using generative AI technologies as a magnet to draw business to cloud.
The strategy is paying off.
“We're seeing strong AWS demand in both generative AI and our non-generative AI workloads, with customers signing up for longer deals and making bigger commitments,” Jassy said.
Amazon’s massive cloud division reported a four-percentage point quarter-over-quarter revenue bump during the three-month period ending March 31.
In the last year, the race to deploy generative AI technologies has created a cloud ecosphere teeming with embedded copilot tools and massive LLMs versed in everything from poetry to Python. But customizable open models with more modest compute demands may be the more practical business solution as enterprises winnow pilots down to scalable use cases.
“Customers want the right model for the right use case,” Arun Chandrasekaran, Gartner distinguished VP analyst, said in a CIO Dive interview. “Enterprises are saying, ‘These enormously powerful models are a great demonstration of the potential of AI, but that's overkill in my environment.’”
Enterprises are less hesitant to feed proprietary data into an LLM when they have visibility into the model, he added.
Opening up LLMs
Currently there are no clear standards for what constitutes an open-source model. Meta, for example, has yet to release the research documentation for its nominally open-source Llama 3. The LLM’s license grants users the right to “use, reproduce, distribute, copy, create derivative works of and make modifications to the Llama Materials.”
Meta also placed commercial limits on Llama 3, requiring developers to request a license for model applications that log more than 700 million monthly users. Chandrasekaran believes most enterprises won’t be hamstrung by that restriction.
At the enterprise level, increased portability and deeper transparency unlock the door to domain-specific customizations. An open model fine-tuned with a company’s data is likely to be far more efficient than a large general-purpose model with broad functionality, Chandrasekaran said.
The downside of model customization is that it takes resources and know-how at a time when technical skills are in high demand and short supply. Even the White House was pushed to make a public effort to recruit AI talent.
Bedrock’s model import function is only helpful if a company has the requisite chops to build AI, Forrester Principal Analyst Tracy Woo told CIO Dive.
“AWS has given you lots of LEGO blocks to build whatever you want in Bedrock,” Woo said. “But if I’m a bank or I run a grocery store chain, I may not have access to the technical people who can do that.”
Hyperscalers should also be wary of the mounting costs of generative AI workloads when cloud bills are already an enterprise pain point, according to Woo.
“Generative AI can be really expensive and that’s one of the things that all the cloud providers are scared of,” Woo said.
Chandrasekaran agreed. “The mom and pop enterprises are not going to be doing a lot of customization for many reasons,” he said. “It’s expensive, there's a GPU shortage you have to deal with and it’s just very complex.”
