How companies use ChatGPT & Co
Are large language models used in commercial products?
Fooling around with ChatGPT is a lot of fun, and if we believe the news then by now it does most of the homework assigned to students. However, like most things large language models (LLMs) are not going to stay unless someone figures out how to make money with them - and clearly, this has already happened: a lot of companies have adopted ChatGPT in their products, substantiating that while LLMs are in a hype phase at the moment they already produce actual value (and profit). Unlike other past hype, this makes me confident LLMs are here to stay. I will go through some of the most important use cases.
Web search
Microsoft has probably been first to commercially adopt LLMs, specifically Prometheus (based on GPT-4) for their Bing Chat. This caused a lot of renewed interest in Bing, to the point where people saw the first serious threat for Google Search. Google offers a similar tool called Search Generative Experience (SGE), however, it is only available via the Google mobile app (Android and iOS) in the US and on Chrome desktop.
The main problem is costs: it is rumored that an LLM-powered search currently costs about 10x more than a standard keyword search, due to the large amount of computing resources needed - indeed, 10x may still be a conservative estimate. Estimates suggest that powering all searches with LLMs would cost Google about 3$–$6 billion dollars. This is the reason why Google only offers the functionality to a small fraction of its users, and why Microsoft announced that Bing Chat Enterprise will be priced at $5 / month in the future.
Code Generation
Arguably the LLM-powered tool increasing efficiency the most is GitHub Copilot (and Copilot X). Initially built upon OpenAI Codex (based on GPT-3), in the meantime a new model has been developed jointly by OpenAI, Microsoft and GitHub. In a nutshell, the tool allows to write software by simply instructing Copilot to write the code. A skilled programmer is still required to check and correct the provided code but typically this is a lot more time-efficient than writing the program from scratch. Already in February 2023, it was reported that Copilot is behind an average of about half of a developers’ code across all programming languages. In addition, Copilot is also able to explain and optimize code, and to automatically produce comments and commit messages.
Education
The new Duolingo Max offers two additional features for language learners: Powered by GPT-4, it explains to students why their answer to a question was incorrect, like a human tutor. And it allows chats with AI personas: imagine role-playing to practice certain situations like meeting a new person.
Chegg assists college students with assignments, initially by human experts. After ChatGPT arrived, interactions with the services drastically decreased and its stock fell by 40%. Now the platform offers Cheggmate, using GPT-4, to help students with assignments.
Quizlet offers Q-Chat, a ChatGT-powered tutor, which allows to study via discussions and produces quizzes on the material. Similarly, Udacity provides a GPT-4-powered tutor for personalized guidance: it helps students to solve complex problems by giving detailed explanations, summarizes content and translates material to other languages.
Collaborative platforms and social media
Salesforce is enhancing Slack with Einstein, an LLM based on ChatGPT. Slack ChatGPT can draft replies and comments in chats, summarize threads and search for outside information.
Similarly, Snapchat has added conversational AI powered by ChatGPT: The chatbot My AI is added to the contact list and users can chat with it. Just like ChatGPT, it will answer any kind of questions or just entertain you…
Ghost is an anonymous social app, allowing its users to share messages in a group chat without revealing their identity. The app also includes the possibility to ask ChatGPT questions within the group chat.
The India-based microblogging app Koo, rival of X (formerly Twitter), has introduced ChatGPT features to their product, assisting users with writing posts: by suggesting top news stories, finding quotes and even drafting the whole blog post.
Customer Relationship management and Customer Support
And there is more from Salesforce: AI Cloud also uses Einstein. For sales reps, it generates personalized emails to customers and for support teams, agent chat replies and case summaries. AI Cloud also generates personalized marketing content across email, mobile, web, and advertising and delivers insights and recommendations.
Octopus Energy, an energy supplier in the UK, uses ChatGPT for customer service. They claim it is now handling 44% of customer inquiries and doing the work of 250 people, receiving a 80% satisfaction rate - to be compared to 65% achieved by human workers.
Other companies that use ChatGPT to talk to their customers include Instacart and Shopify.
Travel agency
Expedia uses ChatGPT for conversational AI assistance: Customers can plan their vacations by chatting with an LLM instead of searching for hotels and flights. The app can also recommend content, like generating lists of hotels the customer might be interested in.
Observability
Dynatrace is a platform for observability and security for cloud-based company offerings, analyzing the health of the used infrastructure, the applications and any security threats, among other things. This produces a large amount of data related to application performance, security, user interactions, and much more, which the company using the software has access to. Recently, Dynatrace introduced Davis Copilot, powered by an LLM, to allow users to “ask their data” via natural language interaction. This way, queries, dashboards, notebooks, and even workflows can be generated.
Computer games
Unity provides tools to help create professional games and 3D content. With their addition of LLM techniques creators can use Muse to create real-time 3D animations, physical effects, code, and other content. In addition, Sentis allows the embedding of LLMs into games: non-player characters (NPCs) can be powered by LLMs - and interact with players more realistically. It is even possible to create new content and environments on the spot while users are playing!
The future
It is not hard to predict that the number of products with a Generative AI backbone will increase strongly in the future. This not only includes LLMs, but also the generation of other modalities like images and speech. The number of profitable use cases will further increase, in particular with costs of using LLMs still going down.