Jose Kuttan, Founder, aviel, Building products in AI and open source.
AI in software has been around for some time, but large language models (LLMs) have changed the landscape in consumer and enterprise software. For the first time, AI is truly a general-purpose computing system.
For businesses, LLMs can perform a variety of tasks that were impossible just a few years ago. It’s now possible to create a book by providing a two-page letter, an entire tutorial, or even a brief textual prompt. It is also possible to create and write code. Let’s consider some use cases from specific areas:
• A financial analyst may give an LLM an Excel file and ask for a summary of the data. Or the analyst may ask the data warehouse a specific sales-related question.
• In health care, LLMs understand the risks of a patient developing a condition.
• LLM can be used to answer a question that requires multiple steps – such as, for example, customer problems in a call center.
How can businesses prepare to integrate LLM
We are likely to see many more new applications of LLM in the coming months and years. With technology handling specialized technical skills, employees’ time and effort can be focused on better understanding the business and customer needs.
This will change the way many organizations work. For example, as more and more software incorporates LLMs, many job responsibilities will evolve into verifying and approving steps suggested by LLMs rather than having humans complete the tasks.
Because employee roles will be changing with the adoption of LLM, business leaders will need to work to help their teams adapt to these new roles. With this in mind, here are three considerations for businesses as they look to get the most from the new levels of productivity offered by LLMs:
Of course, the tricky thing about LLMs is the hallucinations – that is, making up facts – which can make them sometimes unpredictable and a little risky to adopt. If it can create facts instantly, can the multi-step plans and answers it provides be trusted?
There are researchers who are working on solutions—AutoGPT (React), Tree of Thoughts, and other frameworks—to improve the accuracy of LLMs in complex tasks using an approach called grounding and reason and action (or chain of thought). The idea is to have several intermediate steps to add real data and reduce the possibility of hallucinations to the LLM. It may still be hallucinogenic, but it can be made more accurate, stable, and reproducible in production systems.
In the meantime, one way to manage hallucinations while these outlines are being developed is some-shot prompting, a process in which you provide example answers in the prompt, which can sometimes be very effective. If you are using LLM in a domain-specific application, fitting the model with data can also help reduce hallucinations.
Regardless, an important design principle is to always keep the user in control and in charge. Ensure that the output of the LLM is always verified by the user. This brings us back to my earlier point that roles will move towards verifying and approving AI outputs. Business leaders must ensure that users can always verify the output and override it when needed. As models become better and applications become more stable, validation may become necessary.
A key difference between LLM and other AI models is that most other AI were built and trained for specific use cases. On the other hand, LLMs are often very flexible and general-purpose. You don’t need special techniques or training to integrate it into your system. It is an AI model that can be quickly shaped according to the user’s needs, not the other way around.
While LLMs are general purpose, you still need to create data pipelines to expose your internal data to the model. For example, if you are a law firm that is using LLM internally to automate basic tasks, you will need to ensure that all of your documents are correctly extracted and stored in a vector database. Are.
Accuracy will be incredibly important in some industries. For example, in the healthcare industry there is no margin for error. A lot of engineering effort will have to be done to ensure accuracy, although perhaps not a specialized army of ML engineers as before.
Finally, you need to ensure that your employees have the best LLM-enabled tools available. Code suggestion tools for developers, image creation tools for designers, text creation tools for your sales and non-engineering staff, etc. Help them embrace the change and prepare them to take advantage of the benefits to come.
That said, you also need to make sure they understand data security and the risks of sharing sensitive data with LLMs. Solutions such as anonymizing data sent to LLM are improving data security. By understanding the risks, developing best practices for sharing your company’s data with LLMs, and ensuring that the data is anonymized, you can reduce the risks involved in sharing data with LLMs.
Business leaders should be excited about how AI native applications will help increase productivity in our organizations. This decade will bring many changes to enterprises and enterprise structures as the way we work changes from verifying the appropriate set of tasks and figuring out the steps to executing them correctly.
The Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs, and technology executives. Am I eligible?