Is ChatGPT relevant to your service? — Flitto’s look into the cutting-edge chatbot
Artificial intelligence (AI)-based technologies surround us. From the voice recognition in our phones’ accessibility services, automated chatbots, and home appliances that can understand our verbal commands and provide just the right temperature for users — it would take the entire page to exhaust the list.
To top the list, ChatGPT is probably a familiar name for readers who stay up to date on what’s new in the AI scene.
ChatGPT is a super-advanced chatbot operated by OpenAI, an AI research and deployment company. Having drawn over 1 million users within the first week since its launch, the service is estimated to have over 100 million active monthly users as of January 2023. It has truly taken the IT scene by a storm.
With its success, concerns have started to arise for the fields of work previously deemed exclusive for humans. The clever bot has already gone on to write codes, co-author academic journals, and pass the bar exam. It is to no surprise that Google CEO Sundar Pichai has announced “Code Red” as a reaction to the new bot’s threatening emergence.
As a natural language processing and data solution company, Flitto has delved into the chatbot and looked into the ways it could be relevant to our corporate readers.
What is ChatGPT, and why the hype?
ChatGPT uses the GPT-3.5 model to create complex and natural sentences that closely mimic texts written by humans. What makes the chatbot distinct is its ability to answer queries based on the gigantic amount of texts it has been fed and its capacity to ping-pong a conversation with its users.
The bot can even write creative stories and summarize it.
Moreover, ChatGPT does not remain docile when it is asked morally dubious queries and actively refuses to answer them.
The bot also manages to object to questions based on false premises (sometimes it can’t, but we’ll get to that later). It even cleverly admits to its own shortcomings, making it all the more human.
But can we trust ChatGPT on its own?
The new chatbot is undeniably smart. It’s been able to provide some of the most coherent lengthy texts as answers compared to many other chatbot services that came before it.
However, readers who have tried conversing deeper with this clever AI engine probably have noticed some questionable details to the answers it offers. As its creators have also disclaimed, we can’t currently fully rely on its intelligence yet.
For instance, ChatGPT tends to make up for a lack of training data for questions it finds tricky by “hallucinating” answers. This causes the bot to announce unverified, fictional information in its signature reliable tone, like in our case below.
There is no book titled “Th Rspctabl Phonbk” nor an author named Ladybaby Tyler. Meanwhile, a good example of a novel without the letter “e” could be “Gadsby” by Ernest Vincent Wright.
The GPT model also tends to have a difficult time with mathematics. Below, we have asked ChatGPT a simple 4-digit by 4-digit multiplication equation.
And here is the right computation provided by the Microsoft calculator.
Moreover, while the current GPT model boasts a massive 175 billion trainable parameters, these parameters have been collected until 2021. This means that GPT’s knowledge is relevant until the year 2021, and it cannot be relied on for more recent topics.
OpenAI recognizes this issue; and to make up for the potential misuse of its service, its researchers have announced their plans to introduce a watermark technology for the text created by its chatbot. In the future, as AI-generated texts become smarter, it would be increasingly difficult for humans to distinguish between what’s been written by robots and humans. The company’s solution aims to prevent the chaos such a situation could bring.
The technology behind it — how is it relevant to other services/corporate readers?
OpenAI’s Generative Pre-trained Transformer (GPT) models are what’s behind the positive (and negative) assessments on ChatGPT. GPT models use both natural language generation and natural language processing as their key features. The latest GPT-3.5 model has been fed with 175 billion parameters of text scraped from the web. This allows it to predict which strings of words would be most statistically appropriate when given a prompt.
Last December, Google CEO Sundar Pichai issued a “code red” as ChatGPT emerged with its explosive market presence. While the tech giant has dominated the search engine market for over 20 years, it seems to take the potential threat very seriously.
This incident triggers a follow-up question: Will chatbots, when powered by a good language model, usurp search engine services?
From the users’ perspective, the crucial factor would boil down to the efficiency, accuracy, and convenience experienced.
Chatbots can be highly convenient when trained with a colossal amount of information like GPT-3. They can serve as a personalized assistant that provides customized results for each user and query. Personalization is noted to be a crucial feature with growing importance among modern consumers as technology evolves.
Moreover, with chatbots, there is no more need to bump into cumbersome advertisements or sponsored content either. And if your work involves a lot of writing and thinking, it will summarize its findings based on its stored information. It will even write its own content like promotional articles and marketing copies for your convenience.
Most importantly, conversational models (like ChatGPT) that use GPT are able to emulate human conversation, and this is able to positively affect customer sentiments. There have been suggestions on how large language models could make for an attractive replacement for search engines. Analysts in Google also proposed a radical redesign of the current search engine model into a synthesis that incorporates LLMs.
ChatGPT is only one of the many ways the gargantuan language model can be utilized. OpenAI’s examples page demonstrates the myriad of ways GPT APIs could be utilized, including translating programming languages and classifying tweets into sentiments.
Furthermore, OpenAI allows anyone to experience and utilize the power of GPT and implement a human-like chatbot service to our own websites. Commercial usages are also possible as long as appropriate payments are made; the OpenAI APIs are currently paid on a token-based system.
To summarize, the huge success of ChatGPT proves the potential of the commercial usage of large language models (LLMs). The appropriate utilization of the related solutions would make work processes more efficient and provide streamlined values for customers.
A follow-up question now would be how to apply the technology in the best way possible.
Why data supplementation matters
As we have previously pointed out, ChatGPT is certainly an entertaining playground where users can converse with an eerily clever artificial intelligence interface. Nonetheless, its application in the actual service requires more careful consideration.
The GPT-3 model was previously criticized for producing racially biased or offensive texts. OpenAI creators also acknowledge the rooms for improvement within their GPT API embeddings solution, clearly specifying its limitations:
Limitation 1: The models encode social biases, e.g. via stereotypes or negative sentiment towards certain groups.
Limitation 2: Models lack knowledge of events that occurred after August 2020.
ChatGPT, in itself, is a result of thorough fine-tuning and specific training. It was trained to avoid producing problematic answers that can be derived from the GPT model. However, fine-tuning is a rigorous process that involves hours of human assessment. This effort is still insufficient to guarantee a sufficient level of accuracy for every domain or industry-related query.
For this reason, OpenAI also supports the customization of some of its models via fine-tuning. Fine-tuning allows the API users to specify certain appropriate answers for prompts. This makes commercial applications of GPT effective, especially for industries that demand highly specialized information.
An additional training dataset must be ready to fine-tune a model. To experience visible enhancements, OpenAI suggests that the training dataset contains at least a few hundred high-quality examples. Ideally, these examples should have been manually quality-controlled by human professionals. This would ensure that useless or offensive contents have been filtered out.
On a larger level, data supplementation in terms of different languages could be the next step for GPTs to evolve. For instance, there is a vast difference between the quality of answers for English queries and Korean queries as shown in the examples below:
The English press release article demonstrates great fluency as expected. Now, let us request the same prompt in Korean:
We can observe that the article has become significantly shorter. Lastly, we requested the same prompt to be written in Korean:
Not only is it short and contains the words within the prompt almost exclusively, the tone is inappropriate for a press release article.
Similar problems are most likely to be prevalent across many other non-English languages due to the discrepancies among the number of parameters fed to the model.
Conclusion
What’s really fascinating about the whole topic is not really the massive size of the GPT model per se. We don’t mean to downplay the hours of processing the data and several months of training it — But LLMs will most likely continue to grow with more training, though less explosively, with the sheer amount of information that is generated every day.
The huge dataset is bound to contain undesirable content too, which has caused the previous GPT to give false or biased responses before.
Therefore, it’s the mode of moderation demonstrated by ChatGPT that makes it the hot service of today. The chatbot is able to thrive thanks to the additional hours and effort to fine-tune and overwrite the problematic parts of its model. And to achieve this, advanced natural language processing technologies with an active participation of human experts who can vet contents are necessary.
NLPs have become more relevant than ever to the utilization of AI technology, as demonstrated by ChatGPT’s success and the formidable potential of LLMs like GPT. With exponential growth expected in the field, it would be ideal to explore further criteria and regulations regarding these AI technologies: Notably, copyright problems and ethical issues.
Flitto DataLab hopes to keep the conversation going on these topics through more blog posts to follow.
