Despite being a hot topic for several months, AI remains a complex concept for many to understand. To help organizations cut through the noise, our Jake Thorne sat down with two AI experts, Cohesity CTO Dr. Craig Martell and Greg Statton from the Office of the CTO, on a recent episode of Data Talks. The three talked about the state of AI today, what is real and what is fake, and how to use it to achieve business value.
Read the summary below for tips on how to get started with AI, or keep scrolling to watch the video.
As AI matures, so is people’s understanding of it
Recently, there has been a lot of buzz around AI, and in a panic to react, many companies have adopted it—without knowing exactly what it is. At the same time, desperate to show relevance, others have inserted the term in all of their marketing materials, leading to confusion among customers and investors alike. However, as AI technologies slowly mature, people are taking a step back and remembering what it was like to qualify systems being brought into enterprise organizations. Which leads us to…
Start by determining what you’re solving for
This is how many people evaluate technology, and AI should be no different. Ask yourself what use cases you want to address: for example, today, AI’s main applications for IT are improving operations and leveraging backup data for business intelligence. Or put it in the context of inputs and outputs. “A system is working well if you’ve clearly defined what you want for the inputs and outputs and you have the right data in place to evaluate whether that system is giving you the right outputs for the right inputs,” said Dr. Martell. Approaching it from this lens can help your organization cut through the noise and understand what value AI can bring.
Having access to the right data will make or break how successful your result is
Like any other data initiative, to get started with AI, you must determine where the data lives, what is relevant, and if it is clean. It’s a huge process and can be a major hassle if you don’t have the right data management systems in place. As Statton pointed out, “It’s kind of a whole bunch of give and take from either open source projects or maybe different silos of companies and products to help collate this down into a single initiative.” Then, you must involve the data stakeholders to determine what data to use in your AI model. Having the right data in place is critical to producing quality outputs, which is why these steps are important.
Words matter, especially when phrasing questions
Using Cohesity’s AI-powered conversational search solution, Cohesity Gaia, Jake demonstrated how your question phrasing can impact your response. His first question was, “What is the first thing I should do to see if my system is secure?” While it gave him an accurate response, Jake demonstrated how asking a summarization query or a compare and contrast query could refine your responses further. Lastly, he showed how the best type of question leverages AI to generate unique text (e.g., “Can you write me a PowerShell script to register my SQL server? AAGs as sources or always on availability groups?”).
No one model is perfect
Pointing to the highway behind him, Dr. Martell said, “Could you imagine what that was like pre-pandemic? It wasn’t that empty, right? So there’s an example where if you had built a model about traffic and the timing you needed for HOVs or the number of HOV stickers you wanted to give away for electric cars, that model would’ve predicted something very different than what’s needed because the world changed.” AI is driven by data from the past, but the world is constantly changing, which affects the quality of the outputs. To ensure your model works well, you need to monitor it constantly.
The more important the answer, the more important it is to check the answers
As mentioned above, AI models aren’t perfect. Dr. Martell gives the example of buying pants and AI recommending a shirt. “If you do like the shirt, then the model got it right. If you don’t like the shirt, then the model hallucinated or got it wrong, but in that case one, it didn’t matter. And two, it’s easy for you to determine that the model got it wrong.” However, with large language models, detecting if something is wrong is much harder. For this reason, if you’re using AI to make important business decisions, it is essential to check your answers and have experts actually interfacing with your data doing this (e.g., a paralegal if you’ve asked a compliance question).
Key takeaways
AI can be a powerful tool that saves your IT team time, but it is certainly not perfect—or at a point where it will take your job. If your team is thinking of implementing AI technology, remember our key takeaways from this episode of Data Talks:
- Start with your use cases.
- Evaluate AI technologies as you would any other.
- Understand that the accuracy of your model and its responses depends on many factors, including how you phrase your questions and what data you use.
For more insights on data security and management, watch our previous Data Talks episode on securing cloud data. And keep an eye out for our upcoming Data Talks episode—coming soon.
