Balancing creativity and precision: Can generative AIs achieve both?
Even before the superficial success of ChatGPT by OpenAI, numerous tech companies have continuously developed their own large language models (LLMs) for text generation. Remarkable advancements have been achieved by related technologies, especially when it comes to the model’s mechanisms.
Different LLMs each have distinct characteristics. While their compositions, performance, and sizes may vary when it comes to supported languages or data sizes, their end goals have a common thread: to offer values to a group of end-users.
Meanwhile, not a single LLM provides a one-for-all solution to all end-user needs. Unreliability that is in discrepancy with its seeming cleverness has been an ongoing problem for generative AIs available so far. Would it be possible to achieve a foundational LLM that can give both reliable information and creative outputs? If so, how can we achieve it? (Note: Focusing mainly on the text aspect, this article uses “LLMs” and “generative AIs” interchangeably.)
What is a large language model (LLM)?
A large language model refers to a massive structure of language data that runs on a specific algorithm (model) which enables it to process, predict, and generate text. Each company that develops one has its own unique recipes to the model as well as data composition.
The performance of LLMs is often defined by the size of the parameters. The parameter refers to the value in which the model can diversify its output and process inputs more efficiently. The amount of parameters is considered to be proportional to the size of the text data LLMs are trained on; the bigger the text dataset, the higher the parameters the models are able to process.
Some LLMs, in order to scale these parameters to train on, utilize massive information as their initial unsupervised training dataset. For instance, these datasets could have been scraped from the internet. LLMs that were trained using this method can produce human-like output solely by the sheer amount of human-generated text it has learned.
However, not all information available in the world wide web is factual or informative. As goes the popular saying in the field of machine learning, “garbage in, garbage out,” the above-mentioned training method is in part what also limits the model. The resulting model is also left with parameters that contain biased, harmful, or completely wrong information that can be detrimental to the end-user experiences.
To counter this, LLMs can be modified and trained for specific use cases through fine-tuning. The popularized ChatGPT is one such example. It was trained with a supervised fine-tuning method called Reinforcement Learning from Human Feedback, or RLHF, which involves humans who actively teach the model to answer appropriately. Whether or not it would be possible for relatively superficial procedures such as fine-tuning to completely block out unhelpful answers rooted in the foundational composition of the model is another question.
Use cases for “creative” text generation via LLMs
One of the biggest assets for LLMs trained on massive text datasets is their capacity to produce natural results. In particular, end-users can utilize such generative AIs for broad domains that does not require much fact-checking.
Fictional content, diverse genres of writings, internet memes and colloquial conversations can amplify the AIs’ capacity to respond with such use cases where users are looking for initial inspirations, diverse and creative outputs.
Some of the values offered to end-users by generative AIs that are focused on creative use cases include:
- Inspiration for users who seek casual informative references to kickstart their intended activities
- Companionship for users who are looking for someone (or something) to talk to
- Efficiency for users who need to process long blocks of information into more digestible versions
- Entertainment for users who are fascinated by the concept of AIs in general
Through fine-tuning, these values can be adapted to different domains and industries, where the end-users can be artists, computer scientists, marketers, educators, and more. While it cannot be fully relied on yet on its own, it can serve as a good aid for users as long as users are willing to verify them by using other trustworthy resources.
What approaches are available for an “accurate” generative AI?
For AIs to be able to offer more than just inspiration, it needs to provide reliable information.
Specifying domain usage
An approach toward an accurate generative AI for end-users is to fine-tune it in a clearly specified usage domain. Targeting a specific user segment makes it possible to set a realistic goal for data collection and application. By making sure that the dataset contains comprehensive factual information on a particular subject, the risk of hallucination caused by the lack of or inappropriate information would be mitigated.
Clarifying the intended purpose of the LLM-derived product will enable users to clearly understand the limitations of the model, even if the LLM has been trained with data that has not been duly assessed or filtered.
Rigorous fine-tuning
Another option would be to fine-tune the model to a point where it can refuse to answer prompts to which it cannot provide accurate answers.
However, it must not be overlooked that generative AIs’ answers are not based on its active awareness of what is factual. Hallucination and factual errors are based on multiple factors including noisy data points and even forced training.
Using or building LLMs trained exclusively on factual information
Some generative AI services take a different approach when it comes to training large language models. To address the challenge of unsafe responses, some companies that aim to develop accurate generative AIs rely on carefully curated datasets from trusted sources, such as academic journals or verified databases. They may also use techniques such as natural language processing to help filter out inaccurate information from their training data.
For instance, generative AI startup Writer develops their own large language model with an encoder-decoder architecture designed to value accuracy over creativity. Beyond the architectural level, the startup mentions the importance of data, ensuring only accurate, real-life data have been used to train its LLM. Depending on the purpose of using the generative AI, one can actively opt for such model design that focuses on accuracy rather than creativity.
Moving toward a data-centric approach
The usage cases for these text-based generative AIs are nearly limitless. Their adaptations are offered as a product to promote workplace efficiency in various domains like marketing, law, computer science, and even education. Some of them show exceptional abilities to offer creative outputs, and the factor of entertainment has definitely been integral to their wide success in the general public.
The impressive potential has become possible through the continued model-centric approach on generative AI technology. LLMs have developed to achieve notable levels of accomplishments when it comes to model architectures, particularly after transformers.
Meanwhile, there is a discrepancy in the level of accomplishment that can only be solved on a data level. Many of the problems that inhibit generative AIs from being reliable, including hallucination and problematic tone-and-manner, can be addressed by feeding the right data and continuing good maintenance of the model so as to avoid being outdated.
Beyond providing entertainment, usage safety must also be carefully considered when deploying or adapting artificial intelligences. For this, it is about time to more actively discuss the promotion of a data-centric approach for artificial intelligence development.
