April 17, 2023
Pip Courbois, Head of Innovation and AI Solutions
With the launch of ChatGPT by OpenAI and all the fanfare and hype, there is a current of panic going through the NLP community.
This fear is primarily fueled by the perceived impact of ChatGPT on non-technical communities; there is a perception, whether true or not, that these algorithms are going to change everything. Certainly, ChatGPT changes everything. In this short blog post, I want to make the argument that although ChatGPT and its offshoots are technological marvels and will change what we do going forward, there is no need to panic. We are still years away from it replacing the innovations and functionality in the wide area of natural language processing (NLP).
The panic is well-founded. GPT LLM is completely changing the playing field, and we’re in the infancy of these AI algorithms, but this is a change akin to the launch of Google or the personal, desktop computer. Within the data science community, I think of this as the scaling of gradient-boosted machines (GBMs). Once GBMs are scaled to production systems, it questions why anyone would use linear methods such as logistic regression (in the classification use-case, the same argument applies to regression) for prediction except in very specific cases where interpretation, speed of scoring, or inference are necessary.
ChatGPT is a version of a Generative Pre-trained (GPT) Large Language Model (LLM). While there is not enough space in this short blog to demystify these algorithms, at a high level, these models predict the next word of a given input (usually a question) over and over until the sentence, paragraph, conversation concludes. The models are trained on a large corpus of natural language; the latest version of ChatGPT, GPT-4, was released just a couple weeks ago and is trained on the "internet," and essentially on all English-language information out there. By predicting the next word (over and over), the algorithm produces a natural-sounding written response. Also, by training on the internet, the responses are essentially a popularity contest. Although, instead of finding the top web pages like most search engines, the algorithms are finding the next word.
Natural Language Processing (NLP) is a branch of data science focused on the analysis and processing of natural language, whether spoken or written. The definition of NLP is often more specific and synonymous with artificial intelligence, which is giving computers the ability to understand and produce written or spoken words in the same way humans do. We prefer the broader definition because there are many useful NLP algorithms that do not rely on neural networks, deep learning, or other brain-like algorithms. Here at Blend360 we have been developing expertise and algorithms in NLP since our inception. The figure on the right shows the algorithms we deploy with our clients. Many of these algorithms rely on deep learning algorithms, but others do not.
These algorithms still have their place and solve very important problems our clients are asking for. LLMs will not replace them all, and in many cases, a LLM will be severe overkill and over-engineered for the use case.
GPT4 and most other LLMs, trained on the internet, are good replacements for a Google Search or in general conversation but cannot, for example, help a company automate its customer service. Not out of the box. Training a LLM on a proprietary data set is called domain-adaption. To automate their customer service, a company needs a large corpus of customer service data to train the LLM. Traditional chatbots are used in this regard and are a combination of models and business rules. LLM that predicts the next word rather than the entire response will revolutionize customer service but will still need to be trained for specific domains.
We at Blend360 are exploring ways that LLMs can be leveraged for our clients in practical and appropriate ways. We treat every client’s challenge as unique and believe that out-of-the-box, one-size-fits-all solutions are ineffective except in the most generic situations. As such, we are exploring methods to use LLMs in unique situations in a cost-effective way at scale, and to train LLMs on our clients’ data and launch them within our clients’ ecosystems.
LLMs are revolutionizing the NLP space but should be thought of as another tool in our box (albeit a very important one) and not a replacement for the entire toolbox. We have actively added this tool to our toolset and are working with clients on deploying it.
If you want to know more, please let us know by leaving a comment below and/or sending us a message at NLP@blend360.com.
Blend360 co-creates data science solutions with clients to achieve their business goals. Clients are the heroes in our story, and we provide the tools and expertise needed to succeed. We use advanced analytics ,machine learning, and AI to enable data-driven decisions that drive growth and innovation. As a leader in the data science industry, we collaborate with businesses of all sizes to create innovative solutions that make a difference.