Peter Griffin is a feature writer and technology columnist for the NZ Listener.
OPINION: If, like millions of people around the world, you’ve been experimenting with the free ChatGPT service, you’ll now have a handle on its strengths and weaknesses.
After the novelty factor of asking it to write limericks and birthday speeches for you wears off, you’ll typically find it becomes a very studious, but not very intuitive assistant. It also makes a lot of factual errors as it is simply pulling information from the web.
But ChatGPT has a more intelligent sibling, ChatGPT Premium. As the name implies, you need to pay to use it, a fairly eye-watering $37 a month. I’ve been signed up for two months and although the bill shock will soon see me unsubscribe, I’ll miss the more nuanced and coherent conversations I have with it.
That’s down to the fact that ChatGPT Premium is “10 times more advanced” than the free version, according to its creator, OpenAI. Whereas ChatGPT uses 175 billion parameters, which are the weights and biases it applies to craft relevant responses to your text prompts, the premium version, based on GPT-4 (generative pre-trained transformer 4), is estimated to include trillions of parameters.
ChatGPT is based on a neural network, which is designed to replicate how the brain works. With GPT-4, we are dealing with a bigger brain. But having more parameters makes it a more expensive service to operate, which is why OpenAI is charging a fee for access.
However, it raises the prospect of an AI divide where those who can afford a subscription can access the tools that deliver better AI-powered answers and perform tasks to a higher level. The United Nations has pointed to this as an emerging problem on a global basis.
“Pricing will constitute a relatively higher barrier in the global south due to lower average incomes,” its Industrial Development Organisation pointed out in February.
“As these technologies become proprietary, countries may well start developing their own AI infrastructure to enhance workers’ productivity.”
The UK is moving to do just that. In March, the British Treasury committed £900 million to build an exascale computer – which would be more powerful than all but a handful of the world’s supercomputers.
It intends to use “BritGPT” to develop foundational models for AI applications, including large language models for conversational AI uses. In effect, it wants sovereign control over AI, rather than relying on profit-driven Big Tech companies to deliver it.
Would that work for a small country? Can we build a KiwiGPT? It would be expensive to try to replicate ChatGPT, but we may not have to completely reinvent the wheel.
Many local organisations have already developed their own large language models (neural networks), but are supplementing these by buying a licence to access the technology underpinning ChatGPT.
It means we could have a conversational chatbot that gives greater weight to trusted sources of local information, but also draw on the power of ChatGPT to improve the experience. Iwi could collaborate on developing a model that reflects Māori culture and knowledge in AI chatbot responses.
The government should also work with OpenAI and other generative AI systems providers such as Google and Amazon to offer free access to these services in the education and community sectors, in the same way that nonprofit organisations get free or discounted access to software packages.
What else will I miss about ChatGPT Premium? It has the ability to provide answers based on images you feed it. For instance, you can show it a photo of the contents of your fridge and it can respond with recipe suggestions. It can remember longer conversations, and although it still draws only on training data up to September 2021, you can instruct it to use the Bing search engine for more up-to-date answers.
The price of admission currently limits it to business uses. The priority now is to make these more-powerful AI tools accessible to everyone.
