The race among companies to adapt AI is evolving. Companies are now turning their attention to learning how to effectively deploy these powerful tools, rather than simply striving to be first. This development is due to poor creation of prompts (sets of instructions used to instruct the AI ​​to perform specific tasks) and the use of unspecialized models leading to inaccuracies and inefficiencies. It happened when companies realized that they were creating.
There are many examples of this evolution. Companies like Johnson & Johnson are creating libraries of prompts to share among staff to improve the quality of AI output. Other companies, including Starbucks, have gone further by creating internal models.
For context, it’s helpful to know that with so-called large-scale language models such as ChatGPT, you need to give the AI ​​a simple prompt: “Please summarize this story.”
“There are two parts to a prompt. One is to properly explain what you want them to do,” said Professor of Computer Science and Linguistics at Stanford University and director of the Stanford Institute for Artificial Intelligence. said Christopher Manning. luck. “Then a lot of tinkering ensues because people quickly realize that some prompts work better than others. We found that they tend to have positive effects.”
A prompt library can be as simple as a collection of conversation starters, like the one Johnson & Johnson uses to reduce friction for employees using its internal generative AI chatbot. can do.
“What we use is [our chatbot] You can upload internal documents to create summaries or ask ad-hoc questions,” a Johnson & Johnson spokesperson said. “We created an instant library of thought-starters to help employees explore potential use cases related to different areas of the business.”
Other prompts, on the other hand, are designed to minimize the risk of hallucinations (a term used to describe the frequent occurrences of AI generating facts that sound plausible but are not true) or to answer in the most efficient way possible. It is intended for formatting.
“We have different prompt libraries depending on the use case and expected output,” says Christian Pierre, chief information officer and partner at creative agency GUT. luck. “[Our] Strategists and data analysts share libraries and “keyword cheat sheets” that contain specific keywords that can significantly change the output. For example, we’ve found that by simply adding “Specify as Boolean Query,” ChatGPT creates a Boolean query that can be used in social listening tools. ”
Undesired output is often caused by a lack of knowledge of the relevant datasets. For example, a language model might provide an answer to a prompt asking why John was hit by a car, even though it has no information about John or the accident in the first place.
“The tendency of all these models is to use the facts if they exist,” Professor Manning said. [it will] Write something plausible that has no basis in truth. ”
Therefore, to create the perfect prompt, you need to provide a strong level of context, fine-tuning of keywords, and an accurate description of the desired format. The people who take pride in making these call themselves prompt engineers.
further evolve
Unfortunately, even the most optimized prompts may fall short of what large enterprises are looking for.
“These large language models are trained very generically,” said Beatrice Sanz-Saiz, global consulting data and AI leader at Ernst & Young. luck. “What we’re trying to accomplish is really bringing in the best talent, for example tax professionals, and really fine-tuning, retaining and retraining them.”
Ernst & Young has created an internal AI platform called EY.ai. Microsoft gave the company early access to Azure OpenAI to build secure and specialized systems. This speeds up the system, protects sensitive data, and most importantly, allows EY to tune the model to the desired outcome.
“If there’s one task you want to do, like read an insurance claim and write down what we’re doing with the claim and why,” There are quite a few examples: Professor Manning explained, “You can tweak the model to make it particularly good for your business.”
Fine-tuning is done by people with machine learning experience, not by impromptu engineers. At this stage, companies may decide to reduce the dataset by removing unnecessary things, such as the ability to write haiku, in order to lock the model into a specific function.
Ernst & Young further specialized the system by creating embedded libraries.
think about ” [embeddings] as an additional dataset to add to the model,” Saiz said. “By bringing together tax knowledge, national regulatory and even sector knowledge, we can connect all the dots.”
By plugging in these additional datasets, the model becomes highly specialized for its purpose. Companies are finding the best AI recipes that involve building on controlled datasets, injecting embedded libraries, and running queries with customized prompts.
“What we can now do is evaluate clients based on the collective knowledge that EY has built over many years, rather than the expertise of individual tax teams,” Saiz explained. “That he does not just within one jurisdiction, but across multiple jurisdictions and around the world.”
Sáiz believes that personalization of AI models through fine-tuned in-house models and built-in libraries is critical to the future of companies using AI. She also predicts that prompts will become less important as AI gets smarter.
But Professor Manning believes the future will be complicated. Specialized systems exist for high-volume tasks, but there is also room for a generic model that requires designed prompts for irregular tasks, such as writing a job ad.
“These are great tasks that you can give ChatGPT these days,” Professor Manning told Fortune. “I think a lot of companies would do well to hire someone who could learn a little bit and become perfectly good at writing prompts and get great results from ChatGPT.”