Get Instant Loan up to ₹10 Lakh!
‘Generative AI can be seen; it’s more cool than traditional AI’
Generative AI and transformers (that help predict the next word or a sentence, and even a para) were not invented to perform physical stuff. They were invented to understand languages better. Images and videos are only a by-product of that.
As global chief technology officer of Evalueserve, Rigvinath Chevala leads all technology teams at the analytics firm, and serves on its executive leadership panel.
In a video conversation from his North Carolina office, Chevala provided insights into the transformative and disruptive potential of generative AI (artificial intelligence) tools like ChatGPT. He also shared his views on how corporations could leverage these tools safely to boost productivity, and the impact on AI startups, besides employment at large. Edited excerpts:
People are overawed, yet scared of generative AI and fear the dystopian scenario of such tools becoming sentient and taking over the human race. What is your position?
Yes. That’s what’s happening right now. But if you take a step back, from a machine-learning standpoint, most do not realize that AI is behind the scenes in many applications we interact with. But it’s not cool since you don’t see it and don’t touch it.
With generative AI, it is now touchable. And you can see it. That is where the coolness factor comes from, and is the key difference. There is a technology differentiator, too.
In the past, our deep learning techniques, like RNNs and CNNs (recurrent neural networks and conventional neural networks), were primarily sequential—you had to go to layers and layers of networks to get to a clean output.
Consider how Tesla facilitates auto driving—they literally have an RNN or a network of RNNs which takes input from all cameras and decides how the steering wheel needs to turn. That’s the output on that network. Right now, generative AI does not let you drive a car , for which you will still have to rely on the traditional core AI techniques.
Please provide us with some examples...
Generative AI and transformers (that help predict the next word or a sentence, and even a para) were not invented to perform physical stuff. They were invented to understand languages better. Images and videos are only a by-product of that.
For instance, consider the following sentences: A fox did not cross the street because it was wide; the fox did not cross the street because it was tired; the fox did not cross the street because it was dark. The context of these sentences change in each of these examples—you are referring to the street because it’s wide, or the fox as it is tired, and you are referring to a time of the day that is not spelt out in the sentences. And, transformers can figure out every context as you attach the weights to each context. While this is a simple example it is certainly a game changer.
But generative AI is finding applications across various domains where everything is being treated as a language. As an example, it is aiding drug discovery, as well as finding applications for other domains.
Yes. With machine learning we classified patterns automatically but the accuracy was only about 70%, requiring human intervention. But with generative AI, the understanding of that language, and context is much, much better, and the output is amazing. In fact, it’s reducing our effort by close to 40% compared to the previous methods.
Generative AI can be used to help businesses with patent research and market research.
For example, if a company is trying to invent a new product, generative AI can be used to search through patent databases to see if the product has already been patented. This information can help the company avoid infringing on other companies’ patents. Generative AI can also be used to help businesses with market research. For example, if a bank is trying to create a new tradable security, generative AI can be used to research the market and identify the best players to invest in. This information can help the bank make more informed investment decisions. About 80% of the questions that financial institutions ask are easily solvable by generative AI, thus it has the potential to save businesses a significant amount of time and money.
Also, when you scan forms like your 10-K filings (reporting of company financial), etc., sometimes they’re not like native PDFs, and some are less than 300 dpi (fewer dots per inch translate into lower resolution images, making them blurry at times) in some cases, and most OCR (optical character recognition) software, even the best ones, do not have the ability to understand (these images) if they’re blurry. We augmented our tool with transformer technology, and our accuracy shot up to almost 98%.
Yet executives remain wary about using Generative AI in highly-regulated sectors like the banking, financial services and insurance (BFSI) and healthcare sectors, given that these tools like ChatGPT, Bing and Bard are still replete with inaccuracies?
I agree. There are definitely challenges with generative AI, it’s not all rosy. After the Samsung episode (Samsung employees reportedly leaked the company’s secret information to ChatGPT on at least three occasions after the semiconductor division allowed its engineers to use ChatGPT), I had a call with our CEO and told him that “I’m going to tell my developers not to paste one line of code in there”.
It’s really about what data you are dealing with -- there’s paid public domain data and there’s completely private and confidential data. I think it’s a safe bet to use public domain data for your use cases -- they are like library services even if it’s for the financial world and investment banking since there’s nothing proprietary there. But if you deal with mergers and acquisitions, you’re dealing with revenue data that’s not public, and private companies that you are trying to acquire. That is a bad use case to send to ChatGPT. I think the banks are just banning it because they’ve not figured it out yet. So, the easy thing is shut the gates first and then just open each door.
Having said that, it’s a temporary think. For data privacy, the LLMs (large language models) are sitting outside your data network or firewall. If you bring the LLM back into the firewall, would you be able to use it? It’s just a matter of time, and maybe ChatGPT itself will lend itself to putting it back in your firewall.
So, what would you advise senior executives across domains who are exploring the use of generative AI tools?
This disruption is real, so you need to embrace and not resist it. The moment you do so, you see a swath of use cases, and you will wonder where to start. And that’s when it is key to understand one’s prioritisation framework, which has to show some ROI (return on investment) impact for you as a business.
You must consider whether you are dealing with public domain data or confidential data, and then pick the appropriate technologies that are available. There are models that are very good already, which you can download and use for your private data. But there are also costs involved, which I don’t think people realize—it’s not free, and the costs add up pretty quickly. So as a business, you need to build your own architecture for legal reasons, regulatory reasons, and that will become part of your project plan and cost.
I think most companies will now look at it (generative AI) and say: How can I generate new revenue using this technology? Everyone’s going to start thinking of how we can build a tool that’s different.
Speaking about disruption, do you believe that natural language processing (NLP) startups will be forced to pivot or reevaluate their strategies with big technology companies providing these LLM platforms?
I absolutely concur. In fact, I’ve seen that trend. The differentiation is gone. In fact, it’s been leapfrogged. I think the good news is that startups can pivot very quickly. They’re not like giant companies that can’t move fast. So, most of these startups will either sink or swim. And the ones that decide to swim will go in a different direction -- literally in a matter of a few months, not even a year from now. You will soon see new startup categories or existing startups pivoting in that new direction. GPT is great with unstructured natural language. But the moment you give it quantitative data, it still struggles. If you can develop a solution that can actually read the dashboard, or answer questions on the dashboard, it will lead to the emergence of a new category of applications and products.
And what about the impact of generative AI on jobs?
If one of your jobs is taken, you’ll find a higher value job right away, and that cycle will continue. People doing mundane tasks just move on. I think it’s just the profiles of these jobs that will change. For example, prompt engineers is now a new job category. Prompt engineering as a science will be taught in colleges. That’s going to happen with GPT too.
Do you believe that generative AI is a step closer to artificial general intelligence (AGI)?
It’s a grey area. It’s (generative AI) not AGI by itself in the current form. I think as more and more enterprises adopt it, generative AI will become a lot more enterprise-centric and B2B (business-to-business) specific. I know there are alarmists who fear AI is going to take over the world. I don’t think that will be the future. There are going to be outliers—people who abuse this technology for hacking, deep fakes, etc. But there will be a large population that will use this more for domain-centric, domain-specific applications. That’s our bet, and that’s our positioning.
Catch all the Business News , Breaking News Events and Latest News Updates on Live Mint. Download The Mint News App to get Daily Market Updates.
MoreLess
First Published:19 May 2023, 11:55 PM IST
- Top Gainers
- Top Losers
- 52 Week High