Understanding recent developments in artificial intelligence.
In February of 2022 Rob Toews, an experienced AI commentator, wrote that, ‘Put simply, to solve language is to solve AI.’
In the same article he added, ‘The technology is now at a critical inflection point, poised to make the leap from academic research to widespread real-world adoption.’
This was pre ChatGPT. Although others had previously had similar thoughts, these words were insightful indeed.
Since it was released in November 2022, ChatGPT, the chatbot developed by OpenAI, has become the most talked about piece of innovative technology since the iPhone. It has been tested and analysed to exhaustion and filled thousands of pages of both online and paper publications. Perhaps most significantly, it has lit the spark of AI curiosity for those who were previously uninterested.
Great strides for AI
Scientists have long been searching for the ultimate artificial intelligence, that is to say a machine that can think and make decisions in similar ways to the human brain. In 1950 Alan Turing’s renowned paper ‘Computing Machinery and Intelligence’ which examined the question of whether or not machines could ‘think’ and formulated the still relevant ‘Turing Test’ to assess machine intelligence, initiated science’s quest to develop human-parity AI. Over the last 75 years it has occupied some of the greatest minds, claimed billions of dollars of investment and is on-going.
But that isn’t to say that great strides haven’t been made. The advances have been quite breathtaking.
As anyone of a certain age will know, the digital age has dawned fast. Computers have only been part of our lives for a few decades and all but the youngest generations will clearly remember the times when mobile phones weren’t glued to our hands. Many different forms of technology have been heartily welcomed and then quickly discarded. Punch cards, floppy discs, pagers and MP3 players were all short-lived, made obsolete by speedy technological advances.
AI systems have gone from being able to visibly distinguish cards on the left from those on the right in the 1950s to beating humans at backgammon in the 1990s to being able to recognise images of objects in the 2020s.
In the last 5 years however, the advances have been staggering, especially in regard to language.
Why ChatGPT, why now?
Artificial neural networks, a type of AI that is able figure out relationships within datasets without relying on the input of specific code, have been around for a while. But it was the development of the transformer neural net by Google in 2017 that sparked the recent revolution in AI and language.
Transformer models are able to mimic certain processes of the human brain and power natural language processing. They are essentially the bedrock of AI language models like BART and the GPT family, which includes of course, ChatGPT, the behemoth that dominates the conversation today.
GPT stands for ‘generative pretrained transformer’ which is important because it tells us that this model has the capacity to generate or produce output and is trained on existing text data that it consumes via the internet. This ‘training’ consists of absorbing language patterns and recognizing which words are associated with one another. Because the quantities of data involved are huge, it gives GPT a better knowledge of context in language than other previous types of AI.
What can ChatGPT do?
It’s a simple question but the answer is that we’re still finding out.
ChatGPT is really good at responding swiftly to text prompts in natural-sounding human language. It understands multiple languages, has encyclopaedic knowledge (up to 2021) and recreates any style. You can ask it to compose pretty much any form of text, from a sonnet in the style of Shakespeare to a shopping list to a complex project management plan. It could help you write an email, your resume, a social media post or even a novel. It can also write code and create Excel formulas.
ChatGPT is, to put it mildly, impressive.
Of course, what is super exciting is how much time and energy can be saved by using ChatGPT and in a commercial setting everyone is still getting their heads around what that means.
The possible tasks that ChatGPT could take on to ease human work pressure are many. It could:
- produce marketing copy,
- enhance customer service,
- adapt text to a brand voice,
- check grammar and punctuation,
- create a feedback survey,
- strengthen SEO,
- summarize meeting transcripts,
- write schedules.
And of course, much more. The possibilities are pretty remarkable.
And now for the ‘but’. ChatGPT despite its impressive output still retains the imperfections of other AI language models, albeit to a lesser extent.
The problem of inherent bias in AI applications has been well documented. Using data it finds on the internet as a way to learn about the human world means that AI can also absorb the biases found in this data. Humans themselves have many prejudices, many of them unconscious and these are often repeated in information on the world wide web. Cultural, gender and racial biases are all at risk of being copied by AI and ChatGPT is no exception.
ChatGPT also has a tendency to invent responses that sound truthful and balanced but are in fact completely made up. These are commonly referred to as ‘hallucinations’. Often these inventions are easy to spot, especially if you know the facts yourself, but sometimes they sound realistic and you could easily be fooled.
The creators of ChatGPT are fully aware of these limitations and have stated that they are actively working to resolve them. It’s worth noting that ChatGPT does have a much more robust safeguarding system than previous AI models and frequently refuses to provide an answer to queries on sensitive issues.
ChatGPT has the potential to provide hackers and cyber criminals with their best tool yet. Where once phishing scams were relatively straightforward to spot with their poor grammar and clunky English, ChatGPT’s fluency of expression and often faultless spelling provide enhanced scamming prospects to lawbreakers.
And what about all that information you feed into ChatGPT? Where does it go? Until recently this data could have been used to continue training OpenAI’s models and meant personal or confidential information was unprotected. However, the company has implemented a ‘history disabled’ feature which limits the storage of conversations with ChatGPT to 30 days and helps bring it in line with current GDPR law in Europe. It has also announced a business subscription option ‘for professionals who need more control over their data as well as enterprises seeking to manage their end users.’
Part of the future
There is no denying the big steps forward that artificial intelligence has made in the last few years, particularly in regard to how it manipulates human communication. Large language models like ChatGPT that can generate natural-sounding and informed responses as well as engage in intelligent conversations with the people using them, are without a doubt, here to stay.
The drawbacks and dangers of these AI models are being rapidly understood and will hopefully be controlled by sensible and careful management. Whether we are any nearer AI that has reached human parity is still highly debatable but what is abundantly clear is that AI and models like ChatGPT are going to be part of our future. Our priority must be to make sure we use these tools in ways that benefit us all.
If you’re interested in discussing the possibilities of ChatGPT with us here at t’works and how it can add value to your language projects, we’d love to hear from you. Get in touch here.