What we do
TCS Research
TCS Research: Inventing to build sustainable futures
Highlights
- There is no denying the buzz that ChatGPT has created in the months following its launch in November 2022. This has heightened interest in the underlying technology powering these applications, called a large language model (LLM).
- Much of the current discourse in the public domain has focused on the conversational and generative capabilities of LLMs. However, wider developments point toward these models serving to break down barriers in AI adoption and innovation.
- While risks, challenges, and uncertainties prevail, many more capabilities are yet to be explored. A progressive framework is recommended, which suggests moving from low-risk and internal use cases toward higher-order external use cases, minimizing risks in enterprise adoption.
- Enterprises can further benefit from LLMs as they leverage their evolving capabilities—which include multimodality and reasoning—and enable customization and interfacing with external data and knowledge systems to innovate.
In this article
The big picture
LLMs are neural models with billions of parameters.
These models are pre-trained on extensive data and are capable of a wide range of natural language processing tasks.
With recent developments, LLMs are fast emerging as a vital enterprise technology that promises to create shifts in how businesses build, adopt, and use AI.
However, despite their promise and heightened interest among enterprises, concerns over security, risks implications, and potential societal impact remain when considering their use within an enterprise.
While there is euphoria around AI’s generative and conversational capabilities, it is important to step back and take a wider view.
Only enterprises that contextually exploit the potential of LLMs in a responsible and resilient manner will succeed in generating value from them.
Widening scope
With the benefit of large-scale pre-training, a defining characteristic of LLMs has been the ability to adapt existing models to new tasks and domains.
This is demonstrated in instances of code generation, answering medical-related questions, analysis of legal text, and so on.
Until recently, such domain adaptation involved selecting a pre-trained model and fine-tuning it on domain-specific data. However, the latest generation of LLMs has demonstrated unique abilities to adapt to new tasks, with just a few specific examples fed as natural language inputs through prompts.
Such developments obviate the need to build models from scratch and the need for substantial amounts of training data, both of which have been barriers to AI adoption.
With the emergence of generative models such as GPT-3, there is an increased interest in prompt engineering, with supporting technologies such as vector databases and prompt chaining that continually expand the scope of LLMs.
Alongside LLMs, prompt engineering tools and techniques have developed rapidly, enabling complex tasks beyond conversations. Techniques such as chain-of-thought prompting, which helps break complex tasks into logical steps, have improved the performance of LLMs in resolving tasks that need logical reasoning, with prompt-chaining tools making it possible to design and orchestrate multi-step workflows.
Supporting technologies that augment and extend standalone LLM capabilities, such as vector databases and plugins, are expanding. Connecting LLMs with external data and systems will help overcome inherent shortfalls and unlock new possibilities.
LLMs are fast emerging as general-purpose AI, with models becoming increasingly more capable. These will play a meaningful role in enterprise AI adoption and innovation.
Innovating with large language models
Access to proprietary and open-source platforms has made LLMs more customizable.
How? By using out-of-the-box APIs to power LLM capabilities in existing systems to further build, optimize, and host entirely customized use cases.
Such developments are now making it imperative for companies across industries to experiment with LLMs and plan for how to adopt the technology strategically.
While immediate enterprise experimentations with LLMs will focus on mature conversational and generative capabilities, from a futuristic perspective, enterprises must define a strategic roadmap that looks beyond first-order use cases such as conversational and predictive search prompts. These first-gen applications will eventually incorporate emerging capabilities in novel ways to explore potential innovation opportunities.
We present a framework to depict how enterprises need to explore and exploit LLMs to create value in a progressive manner. It begins with low-risk internal use cases, such as writing assistants or content generators, and progresses to complex external use cases powered by combinatorial possibilities.
A progressive framework of how enterprises can leverage the capabilities of large language models to innovate and scale
The horizontal progression shows how standalone models offer limited possibilities given their capabilities, but when interfaced and integrated with external databases, knowledge sources, and software systems, they become highly efficacious.
Their natural language capabilities and foundational knowledge can be combined, contextualized, augmented, and harnessed to power innovation. In the enterprise context, the novelty derived could be in the form of automation, intelligence, conversational interfaces, and unstructured data labeling.
At the same time, a vertical progression emphasizes the need for enterprises to leverage the evolving abilities of LLMs, such as multimodality, reasoning, and executing actions such as search and web navigation. Exploring possibilities along both dimensions will help enterprises identify opportunities to innovate and generate value.
Be mindful
Apprehensions prevail around the safety, security, costs, and performance of LLMs.
In many instances, they have yet to demonstrate their value over existing AI approaches.
There are varied technical challenges such as explainability, consistency, and reliability, with a tendency to hallucinate that could impede adoption. Moreover, uncertainties around emerging AI regulations and privacy and intellectual property risks are also factors to consider.
Internal low-risk use cases (as pointed out in the progressive framework), could provide a safe starting point for experimentation.
Furthermore, LLMs significantly impact decarbonization targets and sustainability strategies given the high resource consumption involved in their running.
There is no denying that newer, better, and more efficient models will emerge as LLMs evolve. For enterprises to adapt to new developments with agility, it is crucial that they architect back-ends and select technology service partners who will minimize lock-in risks. Choosing the right model will determine whether LLMs become an asset or a liability.
