KAVI PATHER: Why firms need to prepare for arrival of ChatGPT and large-language models

ChatGPT, a powerful, freely available tool that uses AI technology to generate text that mimics human writing, is a phenomenon companies need to quickly embrace or risk falling behind in a transformative new technology.

ChatGPT is not the only large language model (LLM), but it is one of the most significant advancements in the field of AI and language processing. Like other LLMs, it is trained on an enormous data set of text, allowing it to answer questions and generate text on a wide range of topics. 

In November, OpenAI opened ChatGPT for public testing and within a week it had more than 1-million users — a record-setting adoption rate. Now active usage numbers are reported to be more than 100-million. 

It demonstrates incredible skill in carrying out short conversations on arbitrary topics, it can complete several different tasks with emphasis on language generation, and it creates original answers to questions it could not possibly have been trained on. 

LLMs like ChatGPT have the potential to revolutionise the way businesses operate. ChatGPT can help companies to gain a competitive edge by improving customer service, increasing efficiency of employees and saving costs. 

For example, they can improve customer service by providing more accurate, direct and human-like answers to customer inquiries, either through AI-powered chatbots or dynamic AI assistants. It is important to understand it is going to hit every business whether it has chosen adoption or not. 

It is likely going to be embedded into everyday tools very quickly. Microsoft already has plans to embed it in its suite of business productivity software, including Microsoft Office (Word, Excel, Outlook and Teams). And the software giant is planning to build it into its search engine, Bing. Google is likely to follow suit with its own LLM. 

We have seen the impact of LLMs with a similar tool developed by GitHub, an internet hosting service for software development with more than 100-million users. Today over 30% of new code is written by AI on Github. It is likely we will see similar trends in office work. Business should therefore be thinking about how these kinds of technologies will make knowledge workers more efficient. 

Within the next five years it is expected that any knowledge worker will be able to use this technology to complete their everyday work. This is expected to include mining information, and then comparing, evaluating, rewriting, formatting and responding to it. Imagine having it read all your company’s policies and advise you on how they apply to the deal you’re currently closing, or automatically rewriting sections of your board report in a better style.

There are interesting SA developments in language models too. Last year Jeff Dean of Google AI funded a local start-up created by African academics — Lelapa AI — looking to implement these technologies in languages that have been neglected by AI advances, including many African languages such as Zulu and Sotho. The company has ambitions of deploying its tools across the continent, allowing businesses to serve customers more efficiently in their home languages.

However, the big opportunity for business lies in fine-tuning the models for specific uses. This is where the base model is customised and trained on additional data (text) to hone it for a specific use. This is likely to be key to unlocking the full power of these technologies. Select examples include: 

• Retailers can use LLMs to provide improved customer experiences through dynamic chatbots, AI assistants and more.

• Life science researchers can train LLMs to understand proteins, molecules, DNA and RNA.

• Developers can write software and teach robots physical tasks.

• Marketers can train an LLM to organise customer feedback and requests into clusters, or segment products into categories based on product descriptions.

• Legal teams can use LLMs to help with legal paraphrasing and scribing.

We are going to see a period of frenzied experimentation as companies, start-ups and researchers search for all the potential applications of the technology.

However, it is important to note that the technology cannot be relied upon right now — every answer we get from it must be reviewed, fact-checked and reworked carefully by a human. The technology is still under development and not yet ready for commercial use. 

The use of LLMs may raise legal and ethical concerns, such as compliance with data protection regulations, intellectual property rights, and fairness and accountability. Language models can be prone to errors and limitations in their output, especially in situations where they encounter novel or unexpected input. These errors can result in significant losses in business applications where accuracy is critical.

Language models are trained on existing text data, which may contain biases and prejudices. These biases can be perpetuated in the models and result in discriminatory outcomes when used in business applications. Managing and designing for these risks is critical for any business adopting the technology.

One of the most exciting properties of LLMs, such as ChatGPT, is the behaviour that emerges at large model scales. This is the ability of the model to learn and adapt to new tasks or contexts without the need for explicit retraining. 

So it is possible that the ability to answer questions factually will improve with larger language models, such as OpenAI’s GPT-4, which is rumoured to be nearly 600 times bigger than the current model. However, it is not possible to predict the capabilities of larger models. 

What is clear is that we are at the very early stage of LLMs and generative AI technologies. If ChatGPT adoption is an indicator, the space is going to move far more quickly than we have seen with other technologies. 

• Pather is EY Africa artificial intelligence & advanced analytics leader.