Summary. The use of AI-based models in business threatens to revolutionize the content creation industry, having a significant impact on marketing, software design, entertainment, and interpersonal communication. These models can produce both text and images, such as a blog.
The large language and image AI model, also known as generative AI, foundation models, or foundation models has created new opportunities for content creators and businesses. These opportunities include:
- Automated Content Generation: AI models that combine large language and images can be used to generate content such as blog posts, articles, or posts on social media. This is a great tool to save time for professionals and businesses who regularly create content.
- Content quality can be improved: AI-generated material is often of a higher standard than that created by humans. This is because AI models are capable of learning from large amounts of data and identifying patterns that humans might not be able to see. It can lead to more accurate and useful content.
- AI models can generate different types of content, such as text, images, and videos. Businesses and professionals can use this to create content that is more interesting and diverse and appeals to an even wider audience.
- AI models can generate customized content based on individual user preferences. It can be used by businesses and professionals to produce content that will be more appealing to their audience and more likely read or shared.
How well does this technology mimic human creativity? As an example, GPT-3 is a “large-language model” (LLM), created by OpenAI. It wrote the italicized version of the sentence above in response to our first sentence. GPT-3’s content reflects both the strengths and the weaknesses of AI-generated content. It is responsive to the prompts that are fed into it. We tried several different prompts before we settled on this sentence. The system can write well. There are no grammar mistakes and the words chosen are appropriate. Thirdly, the system could use some editing. We wouldn’t normally start an article with a list of numbers, for instance. It also came up with some ideas we hadn’t thought of. We would not have thought of the last point, about personalized content.
It is a good example of the value that AI models can bring to businesses. These models could have a major impact on the content creation industry, including marketing, software, entertainment, and interpersonal communication. It is not “artificial intelligence” as humans have long feared and dreamed of, but to casual observers, it may appear that way.
What is Generative AI (Generative Artificial Intelligence)?
Generative AI can already do a lot. It can produce images and text, including blog posts, code, poetry, and artwork.( has even won competitions (controversially). The software uses machine learning models that are complex to predict words and images based on word sequences or previous images. In 2017, LLMs were first used at Google Brain to translate words, while maintaining context. Since then, text-to-image and large language models have been developed by leading tech companies including Google (OPT-175B), Facebook (OPT-175B), OpenAI (a nonprofit where Microsoft is the largest investor), and Facebook (OPT-3 for speech, DALL E2 for images and Whisper for text). Other online communities, such as Midjourney, which won the art contest, and HuggingFace (an open-source provider), have created generative models.
The models are largely restricted to large tech companies, as they require massive amounts of computing power and data. GPT-3 was trained initially on 45 terabytes and uses 175 billion coefficients or parameters to make predictions. A single training run of GPT-3 costs $12 million. Wu Dao 2.0 is a Chinese model with 1.75 trillion parameters. The majority of companies do not have the budget or data center resources to create their models.
Once a generative algorithm is trained, the model can be “fine-tuned” to a specific content domain using much less data. It has also led to the development of specialized BERT models for a variety of purposes, including biomedical (BioBERT), French (CamemBERT), and legal (LegalBERT). NVIDIA’s , BioNeMo, is a framework that allows for the training, building, and deployment of large language models for generative chemistry and proteomics. DNA/RNA can also be handled by this framework.
You still need to involve humans at the start and end of the process if you want to use generative AI.
For a model to generate content, the human inputs a prompt. In general, prompts that are creative produce creative outputs. The “prompt engineer” will likely become a profession until the next generation of AI is developed. This field has already produced a DALL-E 2-image prompts book and prompt marketplace where users can purchase other users’ prompts for a small price. The majority of users will have to experiment with several prompts to achieve the desired result.
Once a model has generated content, the human editor will have to carefully evaluate and edit it. Alternative prompt outputs can be combined to create a single document. Image generation may require substantial manipulation. Jason Allen, the winner of Colorado’s “digitally altered photography” contest, with Midjourney’s help, told a journalist that it took him more than 80 hours to create more than 900 different versions of his art and to fine-tune the prompts. Then, he improved the result with Adobe Photoshop and increased the image’s sharpness and quality with another AI tool. He printed three pieces of art on canvas.
The models of generative AI are extremely diverse. They can be fed images, long text formats, emails and social media, voice recordings, code, structured data, or even longer texts. They can produce new content, translations, and answers to questions. They also provide sentiment analysis, summaries, and videos. We will describe a few of the potential business applications for these universal content machines.
Marketing Applications
These models can be used in a variety of ways, but the marketing application is perhaps the most popular. Jasper, a marketing-focused version of GPT-3 can create blogs, social media content, web copy, and ads. It claims that it tests its outputs frequently with A/B testing and that its content has been optimized for search engine positioning. Jasper fine-tunes GPT-3 models using the best outputs of their customers, according to Jasper’s executives. Jasper is used by many small and medium-sized businesses and some large companies. VMWare’s writers, for instance, use Jasper to create original content, whether it be for email, product campaigns, or social media copy. Rosa Lear is the director of product-led expansion at VMWare. She said Jasper has helped them ramp up their content strategy. The writers have more time to research, come up with ideas, and strategize.
Kris Ruby, owner of the public relations and social networking agency Ruby Media Group is now using text and image creation from generative models. She claims that they can be used to maximize search engine optimization (SEO) and, in PR, personalized pitches for writers. She believes that these new tools open up a whole new world of copyright challenges and help her clients create AI policies. She says that when she uses these tools, “the AI is 10% and I am 90%” because it involves so much prompting. editing, and iteration. She believes that the tools help to make writing more complete and better for search engines, and that image creation tools could replace stock photos in the future and bring about a creative renaissance.
DALL-E 2 as well as other image-creation tools are used in advertising. Heinz used an image with a label that looked similar to Heinz to argue “This is how ‘ketchup,’ looks to AI.” However, this meant that the model had been trained on a large number of Heinz bottle photos. Nestle created an AI-enhanced Vermeer painting to promote one of their yogurt brands. Stitchfix is a clothing company that already uses AI for recommending specific clothing to its customers. They are now experimenting with DALLE 2, to create visuals of clothing that reflect customer preferences in terms of color, fabric, and style. Mattel uses the technology to create images for toy marketing and design.
Code Generation Applications
GPT-3 has proven to be a very effective, even if it is not perfect, code generator. GPT-3 Codex, a program specifically designed for code generation, can generate code in many different languages based on a description of a “snippet” of code or a small program function. Microsoft’s Github has also developed a GPT-3 version for code generation, called CoPilot. Codex’s newest version can identify and fix bugs in its code, and explain what it does – at least sometimes. Microsoft does not intend to replace human programmers. Instead, it wants to “pair” tools like Codex and CoPilot with humans to increase their speed and efficiency.
It is generally agreed that LLM-based code creation works well with such snippets. However, integration into larger programs and particular technical environments still requires human programming abilities. Deloitte’s extensive Codex testing over the last few months has shown that it increases productivity for experienced developers, while also creating some programming abilities for those without experience.
A six-week pilot with 55 developers at Deloitte found that a majority rated the accuracy of the code at 65% or higher, and the majority of the code was from Codex. Deloitte’s experiment showed that code development speeds for relevant projects improved by 20%. Deloitte also uses Codex to translate from one language into another. The firm concluded that professional developers would be needed for a long time to come, but increased productivity may require fewer. The firm found that the more prompts they used, the better their output code was.
Conversational Language
LLMs are being increasingly used as the foundation of conversational AI and chatbots. They could offer a greater understanding of the conversation and context than current conversational technology. Facebook’s BlenderBot can, for instance, carry on long, contextual conversations with humans. Google’s BERT understands search queries and is a component of DialogFlow, the company’s chatbot engine. Google’s LaMBA was designed to be used for dialog. One of its engineers was convinced that the LLM was a sentient entity after interacting with it.
These LLMs are not all perfect conversations. These LLMs are trained to mimic past human language. These systems have been a work in progress. Although they are being improved, hate speech is still not completely removed.
Knowledge Management Applications
LLMs are being used to manage text-based knowledge (or possibly image- or video-based knowledge) within an organization. Many large companies have found it challenging to manage large amounts of knowledge because the creation of structured knowledge bases is labor-intensive. Some research suggests that LLMs are effective in managing knowledge for an organization when model training is tailored to a specific text-based body of knowledge within the organization. Questions could be used to access the knowledge contained within an LLM.
Some companies are exploring LLM-based Knowledge Management in collaboration with leading commercial LLM providers. Morgan Stanley is collaborating with OpenAI’s GPT-3 to refine training for wealth management content. This will allow financial advisors to search for knowledge within their firm and easily create content tailored for clients. Users of these systems may need assistance or training in creating prompts. The knowledge outputs from the LLMs could also need to be edited or reviewed before they are applied. If these issues are resolved, LLMs can help to rekindle knowledge management and make it more effective.
Deepfakes, Other Legal/Ethical Issues, and Concerns
These generative AI systems have already led to many legal and ethical questions. Deepfakes, or AI-generated images and videos, which appear to be real but aren’t, have already appeared in the media, entertainment, and politics. To create deepfakes, you needed a lot of computing knowledge. Now, anyone can create them. OpenAI tried to prevent fake images by “watermarking”, or adding a unique symbol, to each DALL-E 2 picture. In the future, more controls will be needed — especially as generative videos become mainstream.
The question of what is original content and proprietary content also arises with the advent of generative AI. The providers of these systems claim that the text and images created are theirs since they do not look exactly like previous content. They are derivatives of the text and images that were used to train the model. These technologies will be a major source of work for intellectual-property attorneys in the future.
These examples should make it clear that we have only just begun to scratch the surface of the possibilities of generative AI for businesses and their people. In the future, it may be common for these systems to create all or most of our written and image-based content, including emails, letters, or computer programs. They could also provide drafts for blog posts, presentations, or videos. The development of these capabilities will have unforeseeable and dramatic implications for intellectual property and content ownership. However, they are likely to also revolutionize creativity and knowledge. If these AI models progress at the same rate as they have over the past few years, it is hard to imagine the possibilities and implications they could bring.
