Generative AI is not a new technology. Even expert systems during the 1980s allowed for the creation of useful content. Yet it can take awhile for a technology to evolve and find the right product-market fit. And of course, this happened in a huge way with the launch of ChatGPT in November 2022. This system, known as a large-language model, showed tens of millions of people the power of generative AI.
Since then, this technology has become a priority for many enterprises. A survey from Enterprise Technology Research shows that about 53% of the respondents planned to evaluate or use OpenAI’s system. It was the highest ever for a new product.
So then, how does this technology really work – and what are the challenges and issues? And what about the real-world use cases? The answers to these questions are far from clear-cut. After all, generative AI is experiencing rapid change. Then there are the extreme levels of hype. But there are some key concepts to keep in mind.
The Power of Transformers
In 2012, natural language processing (NLP) appeared to have stalled. It was getting tough to find ways to improve the accuracy. Granted, the emergence of next-generation models like recurrent neural networks would help, which allowed for better processing of sequential data.
But these innovations still fell short. Deep learning remained much more effective with recognizing images.
This is why AI researchers started to explore new techniques. The focus was on finding ways to get a better understanding of the context of words and this was done by using attention models.
The monumental breakthrough came in 2017 with the publication of the “Attention Is All You Need” paper. It set forth a sophisticated system of neural networks – called a transformer — to understand the interrelationships of words. The paper demonstrated major strides with language translation.
Over time, researchers understood that the models generally got significantly better based on the number of parameters – say well over 100 billion. The training data set was publicly available content like from Wikipedia and Reddit.
“Generative AI, particularly LLMs such as ChatGPT create human-like text content by uniquely excelling at three main things,” said Asa Whillock, the VP and General Manager at Alteryx Machine Learning. “There is prediction of the sequences of words that people regularly use, maintaining memory to preserve context in a conversation, and the addition of warmth to responses by randomly varying words chosen. As a result, LLMs have been able to advance substantially at creating human-like speech or other content.”
The Use Cases
Generative AI has the potential to impact just about any industry. Allowing for natural language prompts is incredibly powerful. It means that anyone can use the system – even to do something like programming a computer. In terms of the applications, there are already plenty of examples where companies have enhanced their platforms with generative AI. This is often in the form of a copilot or virtual assistant. But in the coming months, there will likely be entirely new products.
So let’s take a look at some of the use cases:
Synthetic Data: Without quality data, it’s impossible to have effective AI. This explains the challenges with autonomous vehicles. The reality is that there are seemingly endless edge cases. But generative AI can help to create data for models.
“We can use generative AI to create lots of ‘silver’ — as opposed to human annotated ‘gold’ — data for training,” said Anju Kambadur, the Head of AI Engineering at Bloomberg and one of the authors of the BloombergGPT research paper. “More data usually means better performance of a model. Similarly, as training data – which is usually larger than evaluation data – can now be generated quickly and efficiently, model and product development is accelerated. In fact, there are scenarios where human labeling accuracy and speed can be improved by using generative AI in a ‘companion mode.’”
Marketing: Generative AI does have some serious limitations. One is hallucinations, which is where the model creates false or misleading content. Newer models like GPT-4 have made progress in mitigating this. But there is still much that needs to be done. This is why generative AI may not necessarily be the right choice where accuracy is essential, such as with healthcare.
But there are areas where the threshold is lower. This is certainly the case with marketing. Content creation was the first killer app for generative AI.
“In 2020, we followed the market pull which was helping people write their everyday copy,” said Chris Lu, who is the cofounder of Copy.ai. The company, which has raised about $14 million, is a pioneer in the category. Its solutions help to create engaging blogs, social media and emails.
Language Translation: This can be expensive and time consuming. But generative AI will be particularly effective with language translation at scale if it can sort out some issues with speed, cost and integration.
“Not surprisingly, translation works best for languages with a lot of data like English, German, French, Spanish, and so forth and less well for ones with little data,” said Dr. Arle Lommel, senior analyst at CSA Research. “Generative AI has two primary advantages. First, you can add directives about translation to your prompts, such as ‘translate the following to German using informal and a sixth-grade reading level.’ Next, it handles context-dependent translation better than current machine translation systems, which primarily process text one sentence at a time.”
The Future
According to Sequoia Capital: “Generative AI can make [knowledge] workers at least 10% more efficient and/or creative: They become not only faster and more efficient, but more capable than before. Therefore, Generative AI has the potential to generate trillions of dollars of economic value.”
Yes, this is the kind of technology that businesses cannot dismiss. Generative AI represents an inflection point, similar to that of the emergence of the Internet.
But again, it’s important to understand its real capabilities, the risks and the practical use cases to realize the true value.
