Fireside chat recap: beyond ChatGPT for document processing

“ChatGPT, the fastest-growing app of all time, and Nvidia’s emergence as a trillion-dollar company are signs that AI is having its iPhone moment.” 

“47% of US jobs could be eliminated by AI over the next 20 years.”

“I do really think that we are in an incredible time right now, and the landscape in how we do business will be disrupted within the next few years. I think that very, very quickly, the world will be divided by those that are using AI tools and those who don’t.”

“77% of global companies are either using or exploring AI in their businesses, and this is only going to accelerate.”

“45% of executives reported that the publicity of ChatGPT has prompted them to increase AI investments.”

“I really think that we really need to rethink how we run companies. The intelligent  enterprise is a thing that needs to be achieved very quickly because, right now, AI is a competitive advantage, but very quickly, AI will be a competitive disadvantage to those who do not adopt it early enough…If you don’t have your AI story, your AI platform, and your AI use case together, you will be disrupted.”

“By 2025, worldwide data will exceed 175 zettabytes. With most of this information locked inside emails, texts, PDFs, and scanned documents, it poses a real barrier to automation and digital transformation.”  

IDC, “Intelligent Document Processing Demonstrates How AI Can Make a Practical Difference to Business Right Now”

“The data we need to process is only getting bigger…the problem really here…is that 80 percent of your enterprise data is unstructured, and that is the massive treasure chest.”

“If you think about all your business intelligence [assets]…your ERP investments, your ETL data integration data, and your warehouse investments…all of that is only scratching the surface. You didn’t even get to the core yet because that’s only 20 percent structured data. So the big question is, really, how can you quickly turn these [unstructured data] assets that you have into structured data, but then what kind of business use cases are unlocked around that?”

“Generative AI is basically the technology behind the tools that everybody is talking about right now, so the Midjourneys and the ChatGPTs that everybody wants to use. It takes an input, for instance, a text or a prompt…and then it generates the output, so it generates a new image…or an explanation, or a poem, or whatever you ask it to do. The special thing is…it can also do things in different styles, so you can ask it for a portrait to be very photorealistic or very abstract, and that’s basically up to the user…but there’s always an input, and then an output is generated based on what the AI has learned before.”

“There’s also this other way to work with AI, and this would be the discriminative AI approach. With discriminative AI, you basically give some guidelines, so it can only group these few specific classifications, or it can only give you the answer based on the input that you gave it…this is really what you want to do in an enterprise setting…so discriminative AI is safer, but where I see the real power of generative AI is in creative work, of course.”

“In general, technology is really good at stuff that we are bad at, like multiplying big numbers, but technology is really bad at things that are easy for us, such as talking, picking up a glass of water, and stuff like that. This is really difficult for a robotic arm to do, so basically, all of the AI researchers are really focusing on closing this gap.”

“Large language models are the bigger version of it, so if the language model has over one billion parameters, we call it a ‘large language model.’ It can learn not only the syntax and grammar of language but also the semantic relationship between words, and especially the very large language models can go further than that and learn concepts about the world and how to work with longer inputs, so yeah, it’s basically a tool to make language computable.”

“Generative Pre-trained Transformers. Transformers are basically the technology for an AI to keep context over a long string of words. That was a big challenge before. Generative AI models that generate output based on text input and pre-trainers meant that ChatGPT was pre-trained on web-scale data to learn concepts about the world.”

“The human brain has neurons that fire when they get inputs and we’re just trying to find a mathematical representation for that…there are a lot of learnings that come from the AI research for understanding how the human brain works, but technically we are trying to simulate mathematically how the brain works.”

“When we think about hallucinations in general, it’s always your brain or your neural network creating a different output than the real world, so there’s a difference between the real world and [the output] you’re seeing, so that’s how I would define hallucination.”

“I think it’s really interesting and creative work if the AI can come up with really creative things and hallucinate really great images and stuff like that, but for an enterprise setting where you want to know the truth about what’s actually inside your document, you don’t want the AI to make things up. This is a big danger that I saw very early on with generative text summarization models and stuff like that.”

“For instance, I was working with a legal text generator that was trained on court orders to generate summaries of court orders, and I just tried it out on contracts, and it was always trying to create a victim and somebody who did something wrong, even though it was just a contract. Of course, really big generative AI’s have some guardrails about that, but there’s still the danger that the generative AI can hallucinate such information on top of what is actually in the document or in the input that you give it.”

“Automation is the killer use case for AI and Enterprises. I think the AI use cases in the consumer world are very clear, and that’s challenging Google’s dominance in the information retrieval arena. But clearly, the automation use case is the golden goose for AI in enterprises. If we think about how an enterprise works, there’s an opportunity here to really get what every enterprise strives for: economic efficiency with low-cost, high-efficiency rates.”

“In the Bay Area, San Francisco has given licenses to multiple self-driving car companies. We do see self-driving cars every single day; it’s creepy; nobody’s in there, but they are real. When the error rate is so low that we can truly license them in global cities, it’s not a question of whether we will have self-driving cars or not. So if we have self-driving cars, we will also have self-driving companies or business processes, and we will have them very quickly.” 

“Companies run on documents. If we’re talking about autonomous companies, companies interact with other entities—either other companies or their customers—through documents. So, if we think about self-driving businesses or business processes, we need to think about how we get there and what it needs.”  

“We need an operating system for autonomous business processes, and we need it really quickly. If you want to peel that onion further, you need an all-tolerant platform that’s highly scalable and self-healing. One that is optimized for AI workloads that are quite different compared to traditional ETL or data warehousing workloads, with different AI capabilities available for less technical users to point and click. The key here is to quickly find technology that is purpose-built for these AI workloads, has these AI capabilities built-in, and allows people to almost curate business processes to get as close—and every day a little closer—to that autonomous business process. Those that do this will significantly outcompete the competition at scale, and with lower costs and higher customer satisfaction.”

“We’re seeing those impacts coming in from the changing nature of the workforce to some of the bigger trends out there that they need to address to handle productivity and so on. This all stems from the same underlying need: to make better access to information easier so you can do something with it.”

“I’m not sure all jobs will be gone. If we look into the history books, there will be a redistribution of jobs. So there will certainly be a lot of jobs that will change in that sense. So should people on the call be afraid to lose their job? Well, maybe not a job, but maybe the job. The information worker will either scale up very quickly and become a person that is able to use a new generation of tools such as AI, or you might be pushed out or not receive that salary raise that you were looking for. That will be the reality—the reality is always that there will be people who can adapt to the new technology, and they will be more valuable to organizations. It will be people who once said, ‘Ah, who needs computers? I can do this all on a piece of paper much faster than typing it in this computer thing.’ We saw how that went, right?”