BY DR SHAWN CUNNINGHAM
15 NOVEMBER 2023
On 30 November 2022, OpenAI released its ChatGPT model as a chatbot. The chatbot quickly went viral on social media as users shared examples of their queries and the witty answers of the service. Examples included asking ChatGPT to summarise the works of Shakespeare in a few bullets, writing a movie script or explaining history. Within a few days, the chatbot attracted over one million users.
ChatGPT is an example of a Large Language Model (LLM). It is a form of generative artificial intelligence that uses a statistical language model trained on massive amounts of data. Large Language Models are not new, but the scale and open access of the ChatGPT technology is what caught the market by surprise.
When ChatGPT3 was launched many of the leading software development companies did not have a ready technology to launch into the market as a standalone service, even if their software had already included machine learning functionality. Furthermore, the backing of Microsoft with its global processing infrastructure gave OpenAI a huge advantage.
In March 2023, OpenAI released the application programme interface (API), which then mobilised thousands (if not millions) of open-source and other software developers to use the OpenAI technology in their solutions. Many large software developers had to suddenly reprioritise their investments to try and catch up or, at the very least, not be left entirely behind. OpenAI also offered developers and users a first subscription fee to use the more advanced features of their software.
Unfortunately, the weaknesses of Large Language Models also made the headlines. These so-called “hallucinations” occur when these models make up missing information that sounds plausible. Many media pundits immediately assumed that the technology had no future. Yet, the technology development announcements kept on coming on at a staggering pace.
Only eight months after launching ChatGPT3, OpenAI released ChatGPT4. ChatGPT4 offered significant improvements over ChatGPT3, like real-time internet connectivity, and more advanced multimodal input capabilities. For instance, ChatGPT4 could now recognise images and sounds and could offer more advanced problem-solving. ChatGPT4 could also collaborate with users on creative projects, and could learn from feedback.
By the middle of the year, several competitors had released their own models with different degrees of success. For instance, Google launched its Bard service but did not receive the same media or developer attention that OpenAI and its main investor, Microsoft, had received. Adobe launched its Sensei AI and Firefly image generation technologies that could fill in blank spaces on a canvas, or that could blend new images into backgrounds by following user prompts. By October, Adobe announced that it in future AI technologies would be central in all future software releases.
Meanwhile, the functionality of ChatGPT was being incorporated into various applications, like the Bing browser, Microsoft Office suite and other applications. E-commerce vendors like Shopify and Instacart, and Grammarly, the popular language and spell checker, released software using technologies licensed from OpenAI.
It is tough to track the uptake or impact of LLM technologies because they are being deployed in the back end of many other technologies. A proxy indicator for the growth in AI is to track the growth of chipmakers like Nvidia, which manufacture the vast majority of the processors used to compute technologies like LLMs. Another proxy indicator is to follow the investments into large data centres[i] like those operated by Teraco Data Environments[ii] or Data Sciences Corporation[iii].
Even people who are critical of AI or generative AI technologies cannot avoid this phenomenon anymore, as the technology is being incorporated into social media, communications and even office productivity software. Just quick glance at the technologies powering the applications on my smartphone shows me that LLMs are now already deployed to our phones, computers and favourite websites.
So what is next?
Currently, the rumours are that ChatGPT5 will be released in December 2023. The patents for the release were already filed with the United States Patent and Trademark Office (USPTO) in July 2023. In the meantime, the updates and the new features of a handful of software developers are keeping the global technology community on their toes as everybody tries to figure out how to adapt or leverage new capabilities.
There are some optimists that believe that the holy grail of general artificial intelligence is upon us. However, I doubt that LLMs will take us there, it is only one kind of artificial intelligence.
If you are curious who I follow to stay informed about the developments and potential of AI and in particular LLMs, then look up the following resources:
Professor Ethan Mollick (@emollick on X) is constantly trying new LLMs tricks, and
Professor Yann LeCun, Chief AI Scientist at Meta, is an important critical voice on the benefits, risks and the policy environment around AI.
For insights into what is happening in AI beyond LLMs, look up Andrew Ng (Co-founder of Google Brain) and Demis Hassabis of DeepMind.
[i] https://www.datacentermap.com/south-africa/
[iii] https://datasciences.co.za