GitHub recently released the technical preview of the first ever app powered by Codex, OpenAI’s latest GPT (Generative Pretrained Transformer) language model. The app is every bit as ambitious as the name suggests: it is described by its creators as “more than autocomplete”, using its vast training set of public source code and advanced contextual learning capacity to recommend tests, convert comments, and provide users with nuanced, varied suggestions for completing their code. Over time, it learns from users’ coding styles to make more precise suggestions.
In other words, its aim is to behave more like a second pair of eyes – with a deep understanding of the computational problem at hand – than a simple autofill tool. The technology behind it is an extension of the immensely powerful GPT-3 model into programming languages as well as natural languages.
The results are just as astonishing as GPT-3’s: Codex solved tens of thousands of competitive programming problems during evaluation. This incredible performance raises a natural question: how long will it be before ‘autocomplete-plus’ tools such as these render human developers largely obsolete?
There are certainly many tasks which can be made significantly easier using AI-powered autocomplete. A fairly obvious one is the repetitive, boilerplate elements of coding, or indeed writing: if some pattern or piece of information needs to be repeated across code blocks or messages, an AI model can ‘learn’ the pattern and reproduce it where appropriate far more quickly than a human could retype it.
This is an efficient way to reduce the time spent on repetitive lines to mere seconds, leaving more time for human developers and writers to work on more interesting tasks.
Many professional projects require teams to spend more hours than they would like on tedious, repetitive subtasks, whether these involve retrieving and copying useful code from a different project, rewriting the same closing paragraph of an email, or anything in between.
Cutting down the time spent on these steps can make a considerable difference to how long a project takes, as well as helping those working on the project to stay productive and engaged with their work.
However, advances in AI have allowed autocomplete tools to do more than just the most menial subtasks. A significant advantage of these state of the art models is their ability to draw from context.
The models can learn things like readability and the correspondences between comments and code from their training sets, then adapt this information to suit the context of the programme at hand and, over time, narrow down the options which are the most appropriate for a particular user.
The suggestion mechanism, wherein a list of possibilities is provided from which the user can choose, is a useful way for the model to learn from the user’s choices and improve itself.
This enables the models to produce suggestions in response to more complex tasks which is not only human-like, but becomes increasingly similar to something the user themselves might produce.
There are natural language autocomplete tools on the market that predate Copilot – such as TypeGenie, a Customer Service autocomplete – that work in much the same way as Copilot.
TypeGenie bases suggestions on the company’s previous messages to customers rather than open source code like Copilot. It can understand the context of a customer’s message and continually adapts to the agent’s style. This sensitivity to context means that a model can suggest contextually appropriate output most of the time, and that it will only get better at doing so over time.
Through this mechanism, advanced autocomplete tools can sometimes generate sentence suggestions that are better than what a human – especially an inexperienced human – could produce.
For a new Customer Service Agent, an autocomplete AI trained on more experienced staff’s messages could go some way towards substituting for a human mentor.
Before we get too caught up in the science fiction possibilities of AI mentors, we should, of course, acknowledge that this technology by no means replaces human-written code or natural language output as it stands. If writing an entire programme or message from scratch, even the most advanced autocomplete will perform worse than a human; given insufficient context, its output will be highly generic and will not do much at all.
The insights, creativity and sensitivity which a human can provide are still unmatched. Even so, the nuance and adaptability of modern autocomplete tools mean that they are an effective way to speed up repetitive tasks and optimise your written output: a time-saving, quality-assuring asset to any team.
This blog was brought to you by TypeGenie. TypeGenie is an auto-complete product for customer service agents. If you are looking to improve your customer service speed and quality, learn more about TypeGenie:Learn more about TypeGenie >>
A closer look at GPT-3 and the affect it may have on the future of business.
Do you want to know if agents find TypeGenie saves them time? Whether they find it distracting? If you are curious about what our users say about TypeGenie and how it helps them, read on.