In this blog post
Since late 2022, generative AI technology surged, impressing business leaders and investors with its ability to create human-like text and images. OpenAI’s ChatGPT gained an astonishing one million users in just five days, outpacing Apple’s iPhone adoption. Facebook and Netflix took months and years, respectively, to reach the same user base. Domains like finance and language preservation are embracing generative AI’s novel capabilities, enabled by neural networks trained on vast data and using attention mechanisms to understand context and generate original content.
As generative AI systems are being created and utilized, a fresh value chain is arising to facilitate the training and utilization of this potent technology. At first glance, it might appear quite akin to the conventional AI value chain. In essence, out of the six main categories—computer hardware, cloud platforms, foundation models, model hubs and machine learning operations (MLOps), applications, and services—only the inclusion of foundation models is novel.
Generative AI systems like OpenAI’s GPT-3 rely on vast knowledge—trained on approximately 45 terabytes of text data—which demands specialized hardware. Traditional computers are insufficient for these workloads; instead, large clusters of GPUs or TPUs with accelerator chips are necessary to process the massive data across billions of parameters in parallel.
GPUs and TPUs are expensive and limited in availability, making it impractical for most businesses to own and manage them on-site. Consequently, companies turn to cloud-based solutions to access and utilize computational power efficiently, allowing for greater flexibility and cost management.
Generative AI relies on foundation models—large pretrained deep learning models designed for specific content creation tasks and adaptable to various applications. Like Swiss Army knives, these models, exemplified by OpenAI’s GPT-3 and GPT-4, generate human-quality text, powering applications like ChatGPT, Jasper, and Copy.ai. Training foundation models involves vast datasets from public and private sources, necessitating expertise in data preparation, model architecture selection, training, and fine-tuning to enhance output quality.
Model hubs and MLOps
To utilize foundation models for building applications, businesses require two key elements: a storage and access platform for the foundation model and specialized MLOps tools for adaptation and deployment within their applications. Model hubs serve this purpose, offering access to closed-source models through APIs and licensing agreements, with MLOps capabilities provided by the model developer. For open-source models, independent model hubs are emerging, offering a range of services, from model aggregation to end-to-end MLOps support, catering to companies seeking generative AI technology without extensive in-house resources.
Foundation models possess the versatility to handle diverse tasks, but it’s the applications built upon them that enable specific functionalities, like customer support or marketing emails. These applications may come from new entrants aiming for unique services, existing providers adding innovative features, or businesses striving to gain a competitive edge in their industry.
Applications built from fine-tuned models stand out
Applications built on fine-tuned models excel in generative AI, falling into two categories: those using foundation models with slight customization and those leveraging fine-tuned models with additional data and parameter adjustments for specific use cases. Fine-tuning is cost-effective, faster, and accessible to many companies. Data for fine-tuning can come from industry knowledge, proprietary sources, or user feedback loops. Staying updated on generative AI advancements is crucial for developers to assess the benefits of adopting newer foundation models with enhanced capabilities.
The emergence of dedicated generative AI services is inevitable as companies strive to enhance their capabilities and explore business opportunities and technical challenges. Established AI service providers will likely adapt and expand to cater to the generative AI market, while niche players may enter with specialized expertise in applying generative AI to specific functions, industries, or capabilities, like customer service workflows, drug discovery, or feedback loop implementation in various contexts.
Evolving Service Catalogue for Specific Verticals & Operations
- Clinical notes and Report generation: Improves clinician efficiency by automating tasks like generating clinical notes, letters, and responses to patient queries.
- Medical Imaging Enhancement and Generation: Leverage medical imaging advancement by producing high-quality images such as MRI and CT scans, facilitating early disease identification and precise diagnostics.
- Anomaly Detection: Learn what normal medical data looks like and then flag anomalies or outliers in patient data, which can aid in the early detection of diseases or unusual conditions.
- Patient Data Augmentation: Generative models can augment datasets used for training machine learning algorithms, thereby improving the accuracy and robustness of these algorithms.
- Know your Client: Customer Data Platform to generate a summarized and unified customer profile with risk and credit score profiles, intelligent document processing in KYC process.
- Fraud Detection and Prevention: To produce synthetic data mimicking genuine and fraudulent transactions, enhancing fraud detection algorithms’ accuracy and adaptability to changing fraud trends.
- Personalized Financial Advice: Enables personalized financial advice by analyzing individual financial situations, goals, and risk preferences to generate tailored recommendations, facilitating informed and prudent decision-making for customers.
- Credit scoring and loan approval: To automate credit assessments by analyzing customer credit and financial data to predict creditworthiness, enhancing loan approval efficiency.
- Financial Document Generator: To streamline financial document creation by producing diverse forms like investment reports, tax documents, and insurance policies, leading to time savings and decreased human errors.
- Text generation: Improving clinician efficiency with administrative tasks such as assisting with generation of clinical notes, letters, responding to patient queries (e.g., autogenerating responses to EHR inbox messages) and producing patient information and educational material.
- Text summarization: Enabling clinicians to easily find the information they need, for example through summarizing information within a patient’s EHR or large volumes of medical literature. Providing patients with specific summaries of their health information.
- Question answering: Enhancing patient-facing chatbots and conversational assistance supporting activities such as triage, care navigation and administrative questions (e.g., billing). Enabling clinical decision support through answering clinical questions, for example, differential diagnoses and treatment options.
- Text classification: Medical domain specific LLMs enable classification of the large volume of unstructured text within the EHR for multiple purposes. This includes making it available for research and data analysis, enabling identification of patients for clinical trials and facilitating clinical coding for billing purposes. Text classification could also be used for sentiment analysis of patient feedback and reviews.
As organizations move forward, they must focus on understanding how generative AI will impact their industries and consider strategic choices to exploit opportunities and manage challenges. Additionally, understanding the value chain of generative AI, including computer hardware, cloud platforms, foundation models, model hubs, MLOps, and applications, is critical for organizations seeking to leverage the technology effectively.
- Exploring opportunities in the gen AI value chain | McKinsey
- The CEO’s Guide to the Generative AI Revolution | BCG
- Generative AI Technology in Business | Accenture
- Quick Answer: What Healthcare Provider CIOs Need to Know About LLM Applications Such as ChatGPT | Gartner