生成AI

To gain a competitive edge, business leaders first need to understand what generative AI is.

Generative AI is a set of algorithms, capable of generating seemingly new, realistic content—such as text, images, or audio—from the training data. The most powerful generative AI algorithms are built on top of foundation models that are trained on a vast quantity of unlabeled data in a self-supervised way to identify underlying patterns for a wide range of tasks.

For example, GPT-3.5, a foundation model trained on large volumes of text, can be adapted for answering questions, text summarization, or sentiment analysis. DALL-E, a multimodal (text-to-image) foundation model, can be adapted to create images, expand images beyond their original size, or create variations of existing paintings.

These new types of generative AI have the potential to significantly accelerate AI adoption, even in organizations lacking deep AI or data-science expertise. While significant customization still requires expertise, adopting a generative model for a specific task can be accomplished with relatively low quantities of data or examples through APIs or by prompt engineering. The capabilities that generative AI supports can be summarized into three categories:

• Generating Content and Ideas. Creating new, unique outputs across a range of modalities, such as a video advertisement or even a new protein with antimicrobial properties. 
• Improving Efficiency. Accelerating manual or repetitive tasks, such as writing emails, coding, or summarizing large documents. 
• Personalizing Experiences. Creating content and information tailored to a specific audience, such as chatbots for a personalized customer experiences or targeted advertisements based on patterns in a specific customer's behavior.  

Generative AI systems are democratizing AI capabilities that were previously inaccessible due to the lack of training data and computing power required to make them work in each organization’s context. The wider adoption of AI is a good thing, but it can become problematic when organizations don’t have appropriate governance structures in place.

The Ethical Issues Tied to Generative AI Governance
As users experiment with these systems, there are serious ethical issues that need to be addressed:

• Unknown Capabilities. Large generative AI systems such as ChatGPT have exhibited a massive capability overhang—skills and dangers that are not planned for in the development phase and are generally unknown and unexpected even to the developers. This can pose a serious threat if the right guardrails are not in place to effectively manage unexpected usage. 
• Bias and Toxicity. Outputs from generative AI will be as biased as the data it is trained on. Many popular language models today are trained on the wilds of the internet, where there is plenty of bias—along with toxic language and ideas. 
• Data Leakage. Many companies have quickly put policies in place to forbid employees from entering sensitive information into ChatGPT, fearing that it could get incorporated into the AI model and reemerge in public.  
• Hallucination. ChatGPT can make arguments that sound extremely convincing but are 100% wrong. Developers refer to this as “hallucination,” a potential outcome that limits the reliability of the answers coming from AI models.  
• Lack of Transparency. Generative AI models currently provide no attribution for the facts underlying the content they generate, which makes it impossible to verify the correctness of generated claims—further increasing the danger posed by AI-model hallucinations.   
• Copyright Controversies. Since the data sets used by AI models are derived from the public internet, a legal question arises: Does the content those models create amount to duplications of copyrighted works?  

Types of Text Models

• GPT-3, or Generative Pretrained Transformer 3, is an autoregressive model pre-trained on a large corpus of text to generate high-quality natural language text. GPT-3 is designed to be flexible and can be fine-tuned for a variety of language tasks, such as language translation, summarization, and question answering.
• LaMDA, or Language Model for Dialogue Applications, is a pre-trained transformer language model to generate high-quality natural language text, similar to GPT. However, LaMDA was trained on dialogue with the goal of picking up nuances of open-ended conversation.
• LLaMA is a smaller natural language processing model compared to GPT-4 and LaMDA, with the goal of being as performant. While also being an autoregressive language model based on transformers, LLaMA is trained on more tokens to improve performance with lower numbers of parameters.

• GPT-4 is the latest release of GPT class of models, a large-scale, multimodal model which can accept image and text inputs and produce text outputs. GPT-4 is a transformer-based model pretrained to predict the next token in a document. The post-training alignment process results in improved performance on measures of factuality and adherence to desired behavior.
• DALL-E is a type of multimodal algorithm that can operate across different data modalities and create novel images or artwork from natural language text input.
• Stable Diffusion is a text-to-image model similar to DALL-E, but uses a process called “diffusion” to gradually reduce noise in the image until it matches the text description.
• Progen is a multimodal model trained on 280 million protein samples to generate proteins based on desired properties specificized using natural language text input.

Generative AI text models can be used to generate texts based on natural language instructions, including but not limited to:

• Generate marketing copy and job descriptions 
• Offer conversational SMS support with zero wait time 
• Summarize text to enable detailed social listening 
• Search internal documents to increase knowledge transfer within a company 
• Condense lengthy documents into brief summaries 
• Power chatbots 
• Perform data entry 
• Analyze massive datasets 
• Track consumer sentiment 
• Writing software 
• Creating scripts to test code 
• Find common bugs in code 

Generative AI has massive implications for business leaders—and many companies have already gone live with generative AI initiatives. In some cases, companies are developing custom generative AI model applications by fine-tuning them with proprietary data.

The benefits businesses can realize utilizing generative AI include:

• Expanding labor productivity 
• Personalizing customer experience
• Accelerating R&D through generative design 
• Emerging new business models 

Generative AI technology will cause a profound disruption to industries and may ultimately aid in solving some of the most complex problems facing the world today. Three industries have the highest potential for growth in the near term: consumer, finance, and health care.

• Consumer Marketing Campaigns. Generative AI can personalize experiences, content, and product recommendations. 
• Finance. It can generate personalized investment recommendations, analyze market data, and test different scenarios to propose new trading strategies. 
• Biopharma. It can generate data on millions of candidate molecules for a certain disease, then test their application, significantly speeding up R&D cycles.  

Executives should work with their data engineers to identify creative ways to discover new generative AI solutions and assess which solutions are likely to bring the most value to the company. Generative AI is still in its infancy and companies must think outside the box to identify unique or hidden applications that will provide unique competitive advantage.

To get started experimenting to find new use cases, leaders need to ask themselves four questions:

1. Where do we have underutilized data that is critical for our business functions? 
2. Can this data be easily used to fine-tune an existing generative AI model? 
3. Can we transform this data into another format (from numerical data to visual data, for example) to leverage existing generative AI systems? 
4. What outputs do we expect and where in our organization could these outputs be used?