ChatGPT’s limitations and possibilities
ChatGPT, the latest product from OpenAI, has recently taken over the AI industry. It has dominated the discourse in both AI engineering and the business world. What are the benefits of this new OpenAI model – and what are ChatGPT’s limitations? We asked Bartosz Baziński, the founder of SentiOne and a developer with expertise in language models.
Benefits of ChatGPT
The colossal amount of data behind the ChatGPT model is its primary advantage and the main reason behind its success. Thanks to a large amount of input data, ChatGPT creates complex sentences and can describe abstract concepts beautifully.
Additionally, the model has a rather neat approach to optimisation, with the language model training on massive amounts of data with lots of parameters (context for each word contains 2048 terms). Another essential component of the process is a separate model scoring its responses (with multiple reference points). It allows ChatGPT to generate truly impressive replies.
ChatGPT’s limitations — and their importance for business
ChatGPT’s limitations are common to any generic language model. We call this type of chatbot a black box. They can generate beautiful answers but won’t truly understand the question. They have no grasp of reality and cannot put any information in context.
For example, after the World Cup semi-finals, internet users asked ChatGPT: `What will be the result of the Croatia-Argentina match?’. The bot answered: ‘I don’t know the outcome of the hypothetical match.’ Hardly a satisfying answer – it could at least try to joke about how the best always wins!
Another problem with ChatGPT is that it relies on raw, unverified data. As a result, it can mislead its users – providing detailed and engaging answers full of false information. As a result, while a conversation with this model might seem genuine, its replies can prove inaccurate, factually incorrect, or full of fake news once you decide to dig deeper.
Since bots don’t understand the meaning of their answers, they can readily deliver incorrect data. It means they can actively spread misinformation dressed in flawless syntax and grammar. And there have already been quite a few such cases. For example, Tay, a bot built by Microsoft on Twitter data, started spreading racist content very quickly.
A final drawback of ChatGPT is the cost of such technology. OpenAI doesn’t share how much time it took to train such a large model, nor how many hardware resources it required. Our experience shows that it would take numerous servers and a lot of computing power. Unfortunately, it is a significant barrier to commercial implementation. GPT-3, the previous model from OpenAI, was already referred to as ‘extremely expensive’. We can therefore assume ChatGPT to be even more costly.
Why would businesses need AI?
As a result, ChatGPT can be a cool novelty for individual users or a helpful tool for content creators and marketers, generating text in whatever style necessary (for instance, imitating a drunken English sailor). Unfortunately, as some of its responses would likely mislead or even misinform, it might be hard to imagine how such AI could benefit businesses.
Nowadays, conversational AIs like ChatGPT allow for customer service automation. The key to success is guiding the bot through a specific multistep process that would solve the customer’s problem. Such a procedure follows a strict definition by the company in question.
Banks, medical facilities, and insurers utilise this technology. These bots follow specific set patterns and provide only the requested information. They must understand questions with 100% certainty and always attribute queries to the correct intent or dialogue path.
Complete control over the process and the bot’s behaviour is crucial for successful commercial implementation. After all, as an extension of the brand image, a chatbot must be trustworthy and cannot create any risk for the company.
How to adjust AI bots to the business?
What ChatGPT lacks is domain knowledge. Bots can only benefit the business once they have learned datasets from a specific industry. That is, ultimately, what businesses need – AI bots based on expertise and precisely tailored training data.
Additionally, business implementation of the AI model needs to fully reflect the brand’s voice to be its trustworthy extension. Without such functionalities, ChatGPT can only be a technological novelty with no chance of large-scale commercialisation.
SentiOne works on a similar model based on domain data from our key industries – banking, insurance, and finance. We build our datasets by monitoring the internet – analysing millions of opinions and online conversations, which we then clean and annotate to make it an efficient training material for the bot.
Additionally, we provide our Natural Language Understanding engine with client-specific data: customer service call scenarios, helpdesk conversation recordings, etc. It allows us to get an even closer understanding of users’ intents. This approach helps us build and test bots faster (since they’ve fewer phrases to learn).
Looking at SentiOne’s resources – the datasets and the team – I am convinced there won’t be a better model for the Polish language than the one we created. That said, large-scale models (such as GPT) can improve the accuracy of intent recognition and push the entire industry forward.