ChatGPT: time to put the toys down and get back to work

It looks like 2022’s most popular Christmas toy was ChatGPT. The imitate-anything-barbie has had all the signs – teachers up in arms, frequently unavailable, and of course, eye-watering price tags.

This isn’t a hot take. Sam Altman, the CEO of ChatGPT’s maker Open AI, very quickly tried to apply the hype-train brakes, describing it as, ‘incredibly limited, but good enough at some things to create a misleading impression of greatness’.

Spend more than a few minutes and you’ll realize that it’s a confident idiot – telling you things (up to some point in 2021) that may or may not be true – but doing so in a range of silly voices, such as Nick Cave, Shakespeare, or a guy who can’t stop bragging about the size of his pumpkins.

Open AI’s launch of DALL-E 2 – the painting-by-words app named after a Disney character – grabbed people’s attention. But the allure of ChatGPT is that it’s a doll with our likeness, so it’s easy to see as something more beautiful.

And so, without a hint of irony, journalists, bloggers, and PR-folk across the world can’t stop talking about this new toy - stuffed with a mix of pseudo-knowledge, half-baked ideas, and average writing.

When will it become a real boy?

Remarkably for a toy, and unlike predecessors like Microsoft’s short-lived chatbot ‘Tay’ (recalled after metaphorically setting fire to itself) - ChatGPT really does stand at the forefront of AI capabilities.

Just over five years ago, a revolution started in the field of natural language processing (NLP); a branch of machine learning concerned with human texts. Fundamentally, machine learning requires representing data (in this case texts) as numbers – something which language had proven extremely difficult to adequately capture, given its varied rules, contexts, and other nuances.

The new ‘Transformer’ models (of data, not action figures), are able to scale so large that they can capture more meaning than ever before. Contained within 165 billion parameters, ChatGPT has hidden numerical patterns from the granular to the grand – representing ‘meaning’ all the way from words to concepts that it found within the 450GB of the internet that it was trained on.

Still, the ‘smoke and mirrors’ is that ChatGPT is simply parroting a stream of words that have a high likelihood of being found on the web – but it is mathematically based on a truly astounding, multi-dimensional representation of language.

What about the expansion packs?

As an isolated chatbot, ChatGPT can’t do a lot beyond spitting out questionable information (and eat the data you feed it). But there are a range of forms that ChatGPT, or other similar ‘Generative AI’ models, are likely to emerge in and be transformative:

• Accessing relevant information from a knowledge bank: the way that we search for information on the web hasn’t changed much since Google’s algorithms ranked websites better than Microsoft, so is long overdue a shake-up.
  
  Beyond that, though, there are lots of other datasets that these models could transform. Scientific literature, for example, has many different layers of meaning embedded– and in the future, models could be trained to extract data in similar ways to humans, such as systematic reviews, summaries, and linked datasets.

• Conversing in technical languages: Programming and law are two areas that are already seeing significant changes – with the possibility that much current work in both could be written in natural language and converted into code/ legalese by these models. Similarly, scientific or medical language could be made more accessible.
  
  As ever with technological breakthroughs, such capabilities are more likely to augment human abilities than replace them – in part, because the limitations (discussed below) are too big to leave it alone.

• Writing in volume: written content is so voluminous today because of the ‘attention economy’ - in which articles, social media posts and platforms themselves are competing for human focus; either directly or indirectly to sell something. The internet is, in many ways, a terrible place now – a swamp of highly-charged, potentially fake, ‘noise’ – and yet it could get a whole lot worse through Generative AI.
  
  At some point, the internet will have to respond to this – and the hope is that we may look back on the attention-vampire-internet in the same puzzled way that we do to the late 90’s ‘breaking the dial-up connection by picking up the house phone’ internet.

Don’t believe the hype

ChatGPT is an incredibly sophisticated sci-fi toy that points to the potential future of AI-assisted creative and communications work – but it's unwise to think it's ready to be used professionally as it is:

• Bias: Microsoft’s Tay got a product recall, because – like ChatGPT – its internal stuffing was ‘words from the internet’. A large portion of that is completely toxic, and much of the rest you don’t even know what it is.
  
  Sure, Open AI learned the lesson and sewed Chat GPT up extra tight – having protocols in place to shut down contentious subjects. It’s not yet spouting Nazi propaganda like Tay – but it’s impossible that it’s not leaking out race, gender and sexuality bias in its answers: because that’s what it's built from.

• Trust: Across the media hype, the only standard people seem to want to measure ChatGPT on is on the incredibly low benchmark of, ‘Does it sound like a human?’
  
  If Generative AI is producing statements that have value and meaning, there should be ways to assess accuracy, ways to assess reproducibility, and even have a way to explain how it came out with the things that it did.
  
  The rush to employ AI is unfortunately driven by those who don’t have the skills for the jobs they’re recommending it for. Any credible professional isn’t accepting, ‘does it sound like a human?’, as a serious threat to their abilities – because the history of technology is a trash can full of toys that other technologists and entrepreneurs thought would disrupt something they couldn’t even be bothered to understand.

• Data and privacy: For ChatGPT to do something that has relevancy for a professional task, it’s going to have to be fed some information that has value: some seed information, something to analyze, something to work with that isn’t, ‘generic press release about a new product’.
  
  'Of course, this isn’t a terrible idea from a data security, confidentiality, privacy and copyright perspective!’, ChatGPT responded when I asked it. But in a way that I think you’ll agree sounds almost exactly like a human.