The Custom Model Program: Tailoring AI for Business Success
In 2024, AWS launched the Custom Model Program under the AWS Generative AI Innovation Center. This initiative aimed at providing rigorous support for companies looking to customize and optimize their AI models. Over the past two years, the program has partnered with both global enterprises and innovative startups across varying industries, including legal, financial services, healthcare, software development, telecommunications, and manufacturing.
These collaborations have yielded customized AI solutions that resonate with each organization’s unique data expertise and brand voice, resulting in more efficient operations compared to off-the-shelf alternatives. The success of these tailored models is evident in the significant cost savings on inference operations they provide.
The Move Towards Advanced AI Adoption
As organizations transition from basic applications like proof-of-concept projects and chatbots, there has been a marked shift toward advanced personalization and optimization strategies. This includes techniques extending beyond simple prompt engineering and retrieval augmented generation (RAG). Notably, businesses are now focusing on creating specialized models for targeted tasks and refining larger models into more compact, efficient versions. Additionally, enterprises are implementing deeper adaptations through mid-training modifications and optimizing hardware for improved throughput.
Such strategic investments pay dividends, as demonstrated by the results from Cosine AI. Working alongside the Innovation Center, the company fine-tuned Nova Pro, an Amazon Nova foundation model, using Amazon SageMaker AI. The results were impressive: a fivefold increase in A/B testing capability, tenfold enhancements in developer iteration speed, and fourfold improvements in overall project timelines. These outcomes are particularly compelling as companies shift towards agentic systems where performance and specific task adaptability are paramount.
Five Essential Tips for Maximizing Value from AI Customization
To maximize the return on investment (ROI) from training and tuning generative AI models, the Innovation Center recommends several key strategies:
1. Align Technical Solutions with Business Goals
Beginning with a technical mindset can prove detrimental. Instead, the focus should be on identifying specific business objectives first. The Innovation Center has worked with numerous clients, achieving a 65% production success rate for projects by aligning technical solutions with tangible business outcomes. Prioritizing measurable success and real-world business value can help avoid flashy experiments that may not yield positive results.
2. Choose the Right Customization Approach
Understanding established customization methodologies is vital. Customers are encouraged to start with baseline techniques, like prompt engineering or RAG, before venturing into complex solutions. For instance, lighter-weight approaches like supervised fine-tuning can effectively enhance model performance with less data and compute resources.
The spectrum of customization options ranges from supervised fine-tuning to domain-specific foundation model development, allowing organizations to optimize their AI models progressively.
3. Define Clear Success Metrics
Measurable success is non-negotiable, regardless of the approach used. Establishing metrics tailored specifically for business outcomes and technical performance is critical. Organizations should ensure their defined success metrics align with their unique objectives, such as relevance, clarity, and style.
A practical illustration can be seen through Volkswagen Group’s use of Nova Pro. They leveraged their marketing expertise to improve the model’s capacity to identify on-brand imagery, aligning with their broader vision of scalable, high-quality content generation.
4. Optimize Hardware for Performance Improvements
Utilizing managed services like Amazon Bedrock can allow users to capitalize on inherent optimizations. However, organizations operating at a more granular level of technology should consider additional strategies for performance enhancement. For example, TGS, which processes massive amounts of seismic data, rebuilt its AI models on AWS’s robust GPU infrastructure, achieving scalable performance gains while effectively processing petabyte-level datasets.
5. Embrace the Diversity of Model Needs
The principle of “one size doesn’t fit all” applies squarely to AI models. While some are exceptional for specific tasks such as document processing or code generation, larger models that cover a broader reach may not be the most advantageous when it comes to resource requirements. Depending on application complexity, it might be prudent to utilize various models for different tasks, ensuring that the architecture remains adaptable to new AI advancements.
Support from the Innovation Center
At the core of the Custom Model Program is the commitment to delivering expert support throughout a model’s lifecycle. The Innovation Center provides a comprehensive process, working backwards from customer business needs to devise generative AI solutions. This hands-on approach ensures that customers benefit from specialized teams that assist in training and tuning models while integrating them within secure customer environments.
For organizations looking to innovate using AI, reaching out to an account manager or engaging with the Innovation Center at AWS events can open new pathways toward transforming their AI strategies into actionable business outcomes. The collaborative efforts of the Center reflect a genuine understanding of how AI can drive efficiency and relevance in today’s fast-paced digital landscape, encouraging businesses to embrace the full potential of model customization.
