Beyond Chat: What Else Can Large Language Models Do?

Of all the subsets of the Transformer architecture introduced in 2017, Large Language Models (LLMs) are the architecture du jour of the current AI boom. LLMs are quickly becoming integrated into many companies’ product offerings because of their superior capabilities relative to previous iterations of neural network-based models. They offer the ability to perform reasonably well with only a few (or no) labeled samples of data (known as “zero-shot” or “few-shot” learning) and the ability to scale adequately with a proportional increase in parameters and training data.

With a high frequency of academic and industrial breakthroughs along with consistent compounding investment, it’s no surprise that companies like Google and OpenAI keep hitting home runs with models by the likes of BERT, GPT3, LaMDA, etc. The market is getting more crowded, and interest is building. Tech giants desperately want in on the action and have plenty of capital to deploy toward developing novel models. In addition, AI companies have been making their models more accessible to the public via public APIs or open-source distribution. This has created a positive feedback loop which produces various applications built on top of these models. These core applications will continue to further drive adoption and diffusion of LLMs, and more broadly, help increase the ubiquity of AI technology in today’s world. 

Core Application Areas for LLMs

If you’re familiar with LLMs like GPT-3, then you know how they can generate human-like output by inferring context from user inputs, or prompts. To the layperson, this is probably the most recognizable function of an LLM: A black box to generate logical textual output. However, LLMs are capable of more than just pure text generation and completion. A single model is actually incredibly versatile and can perform many other natural language processing (NLP) tasks. 

Natural language processing is a decades-old field of computer science, but LLMs appear to be a more effective solution for many of NLP’s core problem areas. Let’s dive in.

Summarization

The task of automatic text summarization dates all the way back to the 1950s. Methods were largely extractive and rule-based, scanning for keyphrases in the input text and arithmetically calculating which phrases should directly comprise the summary. Research has shifted towards more abstract methods which generate original summaries derived from the base text. 

Enter LLMs. With their Transformer architecture and large corpus of trained text data, LLMs are able to better infer context and significant relationships in a body of text. They’re able to better perform abstract summarization and generate meaningful overviews of passages.

Companies such as Open AI and Cohere.AI have developed models that can not only summarize texts but also perform sentiment analysis, topic classification, and more. GPT-3 allows its users to summarize a body of text if prompted to do so. Certain professional fields built on massive corpora of connected documents—like legal practice, academic research, and medicine—could really benefit from using LLMs to surface insights from their core texts. LLMs boost productivity by allowing users to quickly navigate a dense repository of information.

Translation

Another notable application of LLMs is machine translation. Previously, translation methods were—like summarization—rote and syntax-based. Researchers quickly saw a degradation in performance when trying to translate larger passages of text. It’s a much harder task to factor in linguistic nuances like metaphors, idioms, etc. Even Google Translate was culpable of messy translations before it switched to a neural-based zero-shot system. LLMs possess the power of fast and accurate translation.

Recently, researchers at Meta built a model called No Language Left Behind-200 (NLLB-200). NLB-200 boasts being able to translate 200 languages with an average of 44% higher accuracy than previous neural network techniques. In the digital age, translation facilitates communication between different communities and cultures. Imagine online education being taught in a preferred language of choice, or having business meetings with global stakeholders with a reduced language barrier. LLMs like NLLB-200 help enable language-agnostic communication through translation, further connecting more of the world together. 

Conversational AI

Conversational AI refers to an AI-enabled system that engages in natural language dialogue with users. Like the aforementioned applications, conversational AI is not a new concept. Historically speaking, applications like chatbots have been around for a while now. Programs like ELIZA, created at MIT back in 1964, and Cleverbot were some of the first tangible instances of conversational AI. However, these more antiquated instantiations of conversational AI were very limited in scope, only being able to perform a few tasks based on keyword recognition. Here, LLMs again leverage their architecture to offer ever more expansive conversational abilities. 

LLMs have the capability to preserve context over the longer course of a dialogue flow while also recognizing user intent to craft meaningful, nuanced responses. OpenAI’s ChatGPT has captivated the world. This smaller, finer-tuned variant of GPT-3 has demonstrated its ability to emulate a conversation regarding almost anything (barring current events) that the user prompts with its interface. 

From helping you draft a resume to finding bugs in snippets of code, ChatGPT shows just how engaging LLMs can be. ChatGPT attracted a million users in just five days because of how easy it is to use and how many different ways it has to respond. It’s only a matter of time before Microsoft’s partnership with the AI titan spawns personal ChatGPT bots that help users across all Microsoft platforms, from Azure to Windows to its Office suite. ChatGPT’s viral success is a milestone for LLMs, demonstrating their efficacy when applied to conversational AI.

Text-to-Code

In terms of LLM-assisted programming, one new use of these models is to make code directly from text input by the user. Text-to-code generation is a lofty goal that’s juuuuust starting to come to fruition thanks to LLMs. Models like Codex by OpenAI and CodeT5 by Salesforce are trained on public code repositories so that they can use natural language inputs to generate functional code. Conjuring Python scripts and React components has never been easier. 

Beyond text-to-code generation, these models are also capable of more assistive functions like code summarization and autocompletion. Github’s Copilot, a pair programmer powered by Codex, empowers engineers by inferring code from comments and vice versa. It’s also capable of producing code, suggesting alternative methods, writing up well-covered unit tests, and many other productivity-enhancing features. 

Copilot is a prime example of how LLM applications extend beyond human language generation. What if there was a way to use LLMs in a similar fashion to handle even more complex engineering tasks? Could LLMs be used to automate tasks like DevOps or full stack development to save time and avoid duplication? This speculative line of thinking regarding second-order, outside-the-box applications reveals many divergent paths—all aimed at the same ultimate goal—for future research and development efforts. 

What’s Ahead for LLMs?

New language models are a compelling replacement for that which has come before. Their multifaceted nature and sheer enormity allow them to serve as the foundation for many applications looking to tackle problems in the NLP space. While problems such as noisy hallucinations still exist with LLMs, their size has been increasing tenfold every year. The fast pace of LLM development is likely to keep up, and new breakthroughs will continue to show what these models can do. 

With the hotly-anticipated public rollout of models like OpenAI’s GPT-4, Anthropic AI’s Claude, and Deepmind’s Sparrow, among others, it’s probably safe to assume that the trend of paradigm-pushing LLM applications will continue. More and more previously unfathomable applications will be built utilizing LLMs in original, creative fashion. As Large Language Models grow even larger and become more sophisticated, their capacity to serve as human-computer interfaces for almost any task becomes a potential reality