Since the release of ChatGPT in late 2022, AI has received large and increasing amounts of attention and investment. We believe this is entirely warranted – AI in various forms is poised to change the way that businesses work. But one consequence of the ChatGPT release being the catalyst for this wave of attention is that people equate AI with large language models (LLMs), and they equate LLMs with chatbots. If you ask a random sample of people for diverse use cases for LLMs, you’ll get a list like this:
Interactive Google search + Wikipedia replacement
Question answering (especially in retrieval augmented generation)
We love chatbots – ChatGPT and others in its class are amazing tools – but, as an AI consultancy with a long history of projects in the space before the current mania, we’re sensitive to the conflation of LLMs and chatbots. Many of the most exciting potential uses for LLMs have little to do with the chatbot interface, and we think those should get more attention. Here are just a few examples, to excite your imagination.
Universal Data Interface
Right now, a huge amount of effort goes into preparing data for consumption by software systems. Whole businesses specialize in taking unstructured or mixed-format data and producing normalized data from it – for instance, extracting phone numbers from unstructured text, standardizing address information, or identifying non-exact duplicates in lists. But, with a bit of prompting, LLMs can handily do this kind of work.
More abstractly, software engineers spend a lot of time writing “glue code”, which reformats the input/output behavior of two software components to match each other. Early examples of consumer products in this category include the Rabbit R1 and the Humane Ai Pin, which grant access to the capabilities of a smartphone through a unified LLM interface. Though these have achieved only mixed success so far, we’re anticipating more activity in this direction.
Routine data work is no less important at the individual level, and is comparatively neglected because the scale is small. This is probably clearest in the health domain – you can get a CSV file from MyChart with information from all your medical visits, and you can get a CSV file with health information from your Apple Watch, and yet another from your food logging app. Three years ago, you would have needed a software developer or data analyst to get much value from those CSVs – today, an LLM can help you combine them and extract meaning from them in an afternoon.
Text Personalization
When you open a webpage in a foreign language, most browsers will offer to translate it to your preferred language. This is a form of personalization embedded into the application, and in fact it’s a feature built on an LLM. Translation is an important form of personalization, as it dramatically increases the scope of information you have access to on the web, but it’s a coarser form of personalization than we’re most interested in here.
With access to deep information about a user, language personalization could go much further. Reading any document, your browser could automatically know what you want to read and how you want to read it. This could mean making text simpler and easier to parse, or more condensed and technical to take advantage of your expertise. It could translate culturally specific metaphors into a form you understand, or it could adapt workouts or recipes to your preferences and constraints. On a whim, you could read an entirely new novel that’s the best (for you) version of the best (for you) story that could be written. And just maybe we could be rid of the scourge of book-length recipe introductions.
The pedagogical opportunities are also enticing: want to expose your kid to the great epics, but he won’t read anything that’s not sci-fi? You can now create a version of any story you choose in a sci-fi setting, at his reading level, that responds to the themes he's most interested in.
Copilots
After chatbots, this is the category of LLM application that has seen the most real-world progress. We’re all familiar with the suggested phrase and sentence completions in Gmail and Microsoft Outlook – these are simple LLM-enabled copilots. In 2023, Microsoft extended this sort of AI assistance to the creation of other documents – LLMs to help you respond to messages, animate Powerpoint slides, and manage projects.
In the world of software development, Github showed in 2022 that coding copilots can speed up development by at least 55%. Given the current, longstanding shortage of software engineers, that 2x increase in factor productivity could enable the production of a lot of much-needed software.
There’s a lot of promise in this space, and we’re looking forward to increasingly seamless integrations of copilots into composition tools of all kinds.
Simulations and Games
Flight simulators are used extensively by pilots during training. They offer a cost-effective way to develop essential skills without assuming any of the risk that accompanies flying a real plane. Similarly, for any skill involving language production, LLM-powered simulations will be a key tool for practice.
Many business skills would benefit from practice in dynamic but controlled environments – job interviews, sales conversations, and negotiations of any kind. With a simulation, varied, extensive practice becomes possible. In the context of education, Ethan Mollick at Wharton has written extensively about using contemporary chatbots as tutors and coaches.
But beyond the chatbot interface, people are actively working on whole-world simulations…otherwise known as games. While open-world games like Nintendo’s 2017 The Legend of Zelda: Breath of the Wild have impressive physics and chemistry engines, the social environments of these games have been quite limited. When every interaction has to be scripted ahead of time, a lot of human resources are required to populate a world, and interactions are necessarily stereotyped, whether cooperative or antagonistic. But in a world of characters animated with LLMs, non-player characters could develop and pursue their own interests in the background, and nuanced, mixed-motive interactions become possible.
Agents
If an LLM can act effectively as an agent in a simulated world, why not put it in charge of a system in the real one? There are plenty of systems which would benefit from reducing the amount of human interaction required.
For one example, consider navigation of the various bureaucracies we’re immersed in. If these organizations aren’t sufficiently well resourced or focused on customer service, it can take a lot of time to perform simple tasks like returning a purchase, filing a claim, or scheduling an appointment. With enough context, LLM-powered applications could perform these tasks with little intervention.
More speculatively – autonomous vehicles are very effective in routine driving, but act unpredictably when presented with unusual scenarios. The problem is that they don’t know what to do in complicated situations where construction workers, pedestrians, and other vehicles are behaving in ways they haven’t seen before. But an LLM could act as the “brains” in such a situation, reasoning its way through the unfamiliar situation.
Right now, most people think of AI as something that’s outside their lives until they fire up their browser and navigate to the ChatGPT homepage. That’s poised to change. LLMs are going to be a new layer in the fabric of the world we live in.
— Liban Mohamed-Schwoebel, ML Researcher @ Hop