Amazon enters generative AI. Instead of building AI models, it’s soliciting third parties to host them on AWS. Amazon Bedrock lets developers construct generative AI apps using pre-trained models from AI21 Labs, Anthropic, and Stability AI. In addition, Bedrock delivers AWS-trained Titan FMs (foundation models) in a “limited preview.”
“Applying machine learning to the real world — solving real business problems at scale — is what we do best,” Vasi Philomin, VP of generative AI at AWS, told TechCrunch in a phone interview. “We think every application out there can be reimagined with generative AI.”
AWS’ recent relationships with generative AI companies and expanding investments in generative AI app development hinted foreshadowed Bedrock’s launch.
Last November, Stability AI picked AWS as its preferred cloud provider, and in March, Hugging Face and AWS partnered to move the former’s text-generating models onto the AWS platform. In addition, AWS announced a generative AI accelerator for entrepreneurs and a partnership with Nvidia to produce a “next-generation” AI model training infrastructure.
Custom and bedrock models
Bedrock is Amazon’s most aggressive move into the generative AI sector, which Grand View Research believes may be worth $110 billion by 2030.
Bedrock lets AWS clients use APIs to access AI models from several sources, including AWS. Unfortunately, the facts are a bit hazy – Amazon hasn’t disclosed formal pricing. Instead, the company stressed that Bedrock is for major clients creating “enterprise-scale” AI products, distinguishing it from Replicate, Google Cloud, and Azure.
AWS’ reach or revenue sharing may have enticed Bedrock’s generative AI model providers. However, Amazon didn’t share the model license or hosting agreement specifics.
The third-party models hosted on Bedrock include AI21 Labs’ Jurassic-2 series, which are multilingual and can create text in Spanish, French, German, Portuguese, Italian, and Dutch. Anthropic’s Bedrock model, Claude, can talk and process text. Stability AI’s text-to-image Bedrock-hosted models, including Stable Diffusion, can create art, logos, and graphics.
Amazon’s Titan FM family includes a text-generating model and an embedding model, with more likely to come. The text-generating model, similar to OpenAI’s GPT-4 but not as powerful, can write blog posts, emails, summaries, and database extracts. The embedding model converts words and phrases into semantically meaningful numerical representations, called embeddings. Philomin compares it to an Amazon.com search model.
AWS clients may customize any Bedrock model by pointing the service at a few labeled samples in Amazon S3, Amazon’s cloud storage plan — as little as 20 is needed. No user data is utilized for training the underlying algorithms, Amazon maintains.
“At AWS … we’ve played a key role in democratizing machine learning and making it accessible to anyone who wants to use it,” Philomin added. “Amazon Bedrock is the easiest way to build and scale generative AI applications with foundation models.”
Again, given the unsolved legal problems surrounding generative AI, one wonders how many clients will bite.
Microsoft has experienced success with its generative AI model suite, Azure OpenAI Service, which bundles OpenAI models with extra capabilities oriented toward commercial users. Microsoft reported 1,000 Azure OpenAI Service clients in March.
However, litigants are suing OpenAI and Stability AI for using copyrighted material, especially art, to train generative models. (Generative AI models “learn” to generate art, programming, and more by “training” on randomly collected online pictures and text.) Another lawsuit tries to determine whether code-generating models without acknowledgment or credit may be sold, while an Australian mayor has threatened to sue OpenAI over ChatGPT’s mistakes.
Philomin’s refusal to reveal Amazon’s Titan FM family’s training data didn’t inspire trust. Instead, he underlined that the Titan models were developed to detect and eliminate “harmful” content in the data AWS customers submit for customization, reject “inappropriate” information users input, and filter outputs, including hate speech, profanity, and violence.
Of fact, even the greatest filtering systems may be evaded, as proved by ChatGPT. Prompt injection attacks against ChatGPT and related models have been used to develop malware, find open-source code exploits, and generate offensive sexist, racist, and misinformation material. (Generative AI models tend to exaggerate biases in training data or make stuff up if they run out of relevant training data.)
But Philomin shrugged aside their worries.
“We’re committed to the responsible use of these technologies,” he stated. “We’re watching the regulatory landscape… We have several attorneys helping us decide which data to utilize and which not to use.”
Philomin’s attempts at assurance aside, companies might not want to be responsible for all that could go wrong. (AWS customers, AWS, or the model’s inventor might be responsible for a lawsuit.) But individual consumers may — particularly if there’s no payment for the privilege.
GA launches CodeWhisperer, Trainium, and Inferentia2.
On the issue and coinciding with its huge generative AI push today, Amazon released CodeWhisperer, an AI-powered code-generating tool, free of charge to developers without any use limitations.
CodeWhisperer hasn’t caught on, according to Amazon’s move. Its major competition, GitHub’s Copilot, has over a million users as of January, thousands of whom are business clients. So CodeWhisperer has ground to make up, which it wants to achieve on the corporate side with the simultaneous introduction of CodeWhisperer Professional Tier. CodeWhisperer Professional Tier provides single sign-on with AWS Identity and Access Management integration and enhanced security vulnerability detection limitations.
In reaction to Copilot, the AWS IDE Toolkit, and IDE extensions, we introduced CodeWhisperer in late June. CodeWhisperer can autocomplete whole Java, JavaScript, and Python functions with only a remark or a few keystrokes after training on billions of lines of open-source code, Amazon’s codebase, and public forum documentation and code.
CodeWhisperer now supports Go, Rust, PHP, Ruby, Kotlin, C, C++, Shell scripting, SQL, and Scala and highlights and optionally filters the license of functions it suggests that resemble snippets in its training data.
GitHub highlights to avoid Copilot legal issues. Success depends on time. Philomin stated these tools make developers more productive. Developers struggle to stay current. Tools like these help them worry less.”
In less contentious news, Amazon launched Elastic Cloud Compute (EC2) Inf2 instances in general availability today, powered by its AWS Inferentia2 chips, unveiled last year at re: Invent. Inf2 instances accelerate AI runtimes by improving throughput and latency for higher inference price performance.
Amazon also released Amazon EC2 Trn1n instances powered by AWS Trainium, Amazon’s AI training chip, today. Amazon claims they offer 1600 Gbps of network capacity and 20% better performance than Trn1 for big, network-intensive models.
Inf2 and Trn1n compete with Google and Microsoft products like TPU processors for AI training.
“AWS offers the most effective cloud infrastructure for generative AI,” Philomin asserted with conviction. However, customers require the correct pricing for these models. It’s one reason many customers haven’t produced these models. Generative AI pushed Azure down. Will Amazon suffer the same fate? It depends.
