GPT3 could revolutionise how business can use AI
OpenAI’s language prediction model that produced an essay could be used to deploy cloud resources and improve the quality of chatbots, among other business use cases
Artificial intelligence (AI) is bounding ahead. Powerful tools such as the Generative Pre-trained Transformer 3, also known as GPT3, have near limitless access to computing power using scalable cloud infrastructure.
Software developers have rapidly climbed the ladder from simple “if this, then that” type automation through to powerful chatbots that can mimic human responses and intelligence in attempts to deliver general AI.
OpenAI, a Silicon Valley non-profit founded by Elon Musk and other entrepreneurs, created GPT3 – a powerful natural language processing (NLP) AI that uses deep learning to produce human-like responses to text queries and to solve problems using natural language.
The Guardian recently published an essay written completely using GPT3, asking the software to convince people that robots come in peace. Although the article probably doesn’t have aspiring writers quaking in their boots, such an undertaking would have been impossible just a year ago and the stuff of fantasy just a year or two before that.
But this powerful AI can do so much more than produce essays of dubious quality.
GPT3 has been trained with a huge amount of natural language from the internet – almost 500 billion words in order to train the system with the largest pool of knowledge possible.
The training cost, estimated to be as high as $12m, delivered a model that uses 175 billion parameters – significantly larger than its predecessor, GPT2, which boasted a mere 1.5 billion parameters. And that was a 10-fold increase from the first iteration of GPT.
It i’s worth noting that OpenAI held back the release of GPT2 because of fears that it could be misused by extremist groups to spread misinformation. That gives some indication of the power such software can wield.
NLP has broad applications for business. According to Gartner, the strongest and most immediate use cases for NLP relate to improved customer service, such as responding to customer and employee queries and the automation of tasks. For example, legal tasks such as contract analysis and compliance enforcement could be undertaken by AI, freeing up people to work on more complex and higher-value tasks.
Very powerful model
Min Sun, chief AI scientist at Appier, says GPT3 is a very powerful model for generating languages designed to solve tasks specified by the task description or a few question-answer pairs.
“For instance, English to French translation tasks where an English sentence is the question and the corresponding French translation is the answer,” he says. “It also goes beyond human language to programming languages such as generating program instructions when given English descriptions.”
Using GPT3, one developer has built a tool that translates natural language into SQL queries. Another creates application layouts by simply asking the software to build the design you want.
GPT3 can also be used to do everything from creating spreadsheets through to building complex CSS, or even deploying Amazon Web Services (AWS) instances, a neat trick given that OpenAI has signed a deal to build GPT3 exclusively on Microsoft Azure.
Such novel applications are possible because GPT3 is designed to perform well off-the-shelf without needing to fine-tune the model with extra data, says Sun.
“Other than the common tasks in NLP such as summarisation and chatbots, I can see GPT3 becoming a very important tool to make traditional robotic process automation [RPA] much more widely applicable,” he adds.
The applications for RPA are interesting. While that emerging technology has found a place in many enterprises for automating many routine tasks, the injection of an intelligent AI could further boost it by enabling more complex automations to be created.
For instance, GPT3 could be used to interpret complex text in a document and launch an action or create an alert. Or it could be used to identify recurring text across an organisation and suggest automation opportunities that might not be obvious to human operators.
Read more about AI in APAC
- Not content with his stellar achievements in artificial intelligence, Chih-Han Yu wanted to create an impact with his work and formed one of Taiwan’s most important unicorns.
- Singapore’s DBS Bank is doubling down on its use of AI and data analytics in a bid to provide retail customers and investors with more personalised services on their mobile devices.
- Fujitsu has developed a computer vision model that recognises hand-washing gestures to ease the enforcement of stricter food handling rules in Japan.
- Alibaba is among a growing crop of technology companies that are rising to the challenge of solving the toughest problems in natural language processing.
Leigh Pullen, CEO of Australia-based RPA outfit CiGen, says: “RPA is an established technology and we are already using machine learning, natural language and AI to extend capability on existing automation cases and detect new use cases to give businesses the power the automate more complex processes.”
Miguel Grinberg, a developer at Twilio, notes that GPT3 can also stand in for a customer service agent and converse about a variety of topics in a natural, human way.
“It can be easily trained to provide information specific to an event, a product line, or even specific customer details, such as reservations,” he says. “It can also assist agents by pre-generating one or more possible responses to a customer prompt, which the agent can review and edit as necessary before sending, saving them time and effort.”
However, GPT3 is not without its barriers and challenges. Sun notes that although GPT3 has demonstrated that it can be grammatically correct and even create text that resembles the informal or idiomatic ways that humans communicate, it has not fully grasped the nuances of the reality of the human world.
The technology is also costly, even when it is used without being fine-tuned, and needs significant optimisation before it can serve a broad range of business usage needs at scale, according to Sun.
Twilio’s Grinberg adds that although GPT3 can generate text that is often indistinguishable from that of a human, it does not understand any of it. And so, in the course of a conversation, it may contradict itself, or mention concepts or ideas in perfect English, but not based on facts.
Then there is the problem of “deep fakes”, which are underpinned by generative AI technologies such as GPT3, according to Gartner. Although deep fakes are not yet pervasive among the fake content and news spread across the internet, this will change rapidly in the next five years.
With GPT3 being the first commercial release of this technology, Grinberg reckons the rapid pace of development between the initial release of GPT through the second version and GPT3 could slow down. “My hope for future iterations of this technology is that its training is continuously updated to cover current events as they happen,” he says.