Understanding the Generative AI Landscape

The making of the ‘thought engine’ and how far have we come, was a session theme at the recently concluded convention on AI. A team of eminent panelists discussed the challenges and opportunities brought in by AI.

Mr Sachin Premnath

Director-Business Consulting, EY

ChatGPT has generated a lot of interest both in the investment and technology communities. Business users are now asking us if we can apply any AI component to their work. This is the opposite of a year ago when we used to pursue customers to try Gen AI in their work. ChatGPT is the fastest to reach 1 million customers, and a lot of venture capitalist funding goes into it.

There are social, economic, and legal constraints in using ChatGPT, and laws have been passed in the US and EU regions regarding its use. Stanford has conducted a study on how various models are performing and how reliable they are. There are places where people have applied AI in their day-to-day work, such as preparing a speech, generating templates, wordsmithing, and writing emails to clients, which are perfect ways to use it. You could also use it to improve revenue opportunities and identify new revenue streams that did not exist in the past. However, one has to be judicious in using the various components of Gen AI.

AI: A Holistic Field

Over the last decade, people have been using the words AI and ML interchangeably. AI is a holistic field of study that encompasses machine learning algorithms. The more layers you train AI, the more it becomes deep learning. It has evolved from natural language processing and has been trained specifically from a tech standpoint. Today, unless you want to experience it, you don’t need to go to Paris and stand near the Eiffel Tower for a picture. You can use Gen AI to create a more realistic image of you by the Eiffel Tower. From a source code generation standpoint, many of my clients are leveraging this, especially for re-platforming from one language to another. It’s not limited to text anymore. You can create content, images, music, videos, and pretty much anything; it is possible now.

Any enterprise can be broken down into three layers: analysts, function heads, and CXOs. All three of them can leverage Gen AI, and there are various use cases for it. You can use it for searching knowledge databases. When you are on a call with a customer and need to refer to an SOP document before responding, Gen AI can search those documents and provide a relevant response. The function head can use the search functionality to gain insights from the document. Thus, it can be applied to every role.

Mr Vijay Karunakaran

Founder & CEO, Ingage Technologies Pvt Ltd

Even a year before ChatGPT hit the market, we had been using Gen AI for a lot of content generation. I also advise the Tamil Government for AI projects. We help the forest department monitor and track the movement of elephants so they don’t get killed on the railway track.

I am going to discuss the role of semiconductors and Metaverse. Semiconductors drive the Gen AI explosion. On average, close to 10 million people log in and send queries to ChatGPT. They get 8 billion searches every single day. That’s going to cost them heavily, and it is not going to be free for too long. They already charge for ChatGPT Plus. Even if you pay $20 a month, you’re restricted to a certain number of queries per day or per week. Why is that?

Mobile Computing & Apps

The real cost comes from computing. There are many semiconductors. We moved from the PC world in the 1990s to a smartphone mobile world in 2000. It is the CPU, the central processing unit, that is delivered by the Intels and AMDs of the world that drove the PC computing. Then came the mobile smartphone computing, also a CPU, but we had a new entrant called ARM processor driven by a bunch of semiconductor companies. Now, we are entering the new role of AI computing, which can also be called immersive computing or spatial computing.

Across the PC computing market, the economic share of the technology stack for semiconductor companies was close to 20-30%. That dropped to 10% in the smartphone era because the smartphone market exploded due to applications. It is the software that took a major market share of the entire economic value of mobile computing. Now, for the next five decades, as we are entering generative AI, it is a great opportunity for semiconductors to recapture 40 to 50% share of the entire technology stack. The semiconductor is the most important component of generative AI.

There are four layers in the AI stack:

• Layer 0 is the hardware and the cloud which will drive the AI explosion, and that is where the semiconductor comes into play.
• Layer 1 is Model Foundation which comes from the Open AI companies.
• Layer 2 is Integration, Orchestration, and Deployment Tooling, which includes all the frameworks and is provided by companies like Google and Microsoft.
• Layer 3 is AI applications like ChatGPT, image creation, and so on.

From CPUs to GPUs

There are various types of semiconductors. We have terms like FPGA, GPU, and CPU. GPU means a graphical processing unit, while CPU stands for central processing unit. The GPU was born from the requirements of the gaming industry. It is highly parallel processing. The company that makes that is Nvidia. There are only four companies that are beyond $1 trillion. The first is Apple at $2.6 trillion. India’s GDP is only $3 trillion. Then comes Microsoft, Google, and Saudi Aramco. This is followed by Amazon and Nvidia. Its market cap doubled in the last one year because it is the GPU of Nvidia that drives the AI industry. A CPU in a computer costs only $80 or $90, but an Nvidia chip costs $10,000.

On the other hand, while it costs one billion dollars to build a data center built using CPUs, it costs just 100 million dollars (just one tenth of a CPU) for a GPU-based parallel processing data center. Hence, there’s a huge demand for GPUs of Nvidia. Intel is trying to catch up. In the last five decades, oil and gas was the currency for international diplomacy in geopolitics. For the next five decades, it is going to be semiconductor chips. Why does America defend and spend billions to save Taiwan? Is it because they like Taiwanese people or their noodles? No. It is because they have the world’s best top-notch cutting-edge fabrication company called TSMC, which controls 90% of all the advanced chips manufacturing.

Facebook changed its name to Meta because they think that the Metaverse is the next computing and it’s going to be the future. Over the years from the 1940s, we have come a long way in computing. We had workstations and floppy cards. Then we moved to PCs and laptops, and now we are in the smartphone era. The touch and feel interaction with the device is going to be immersive as we go forward. We’re going to use gestures and eye movements to do the computing. That is where we’re heading. Then we have the AI edge driven by augmented reality and Blockchain.

Web 1.0 to Web 3.0

We have also moved from Web 1.0 to Web 2.0 and Web 3.0. In Web 1.0, when we enter the URL, we get the data from the server—which is Read Only. That was the start of the internet revolution, and it was there for 10 to 15 years. Then came web 2.0. Currently, most of our applications are on web 2.0. It can read; it can write. When you order food on Swiggy, you give a lot of information about yourself, your location, and your credit card; it means that you are writing data from the device to the server. That’s what helped the entire mobility ecosystem.

But there are two drawbacks to web 2.0. Number one, it is centralized. That means anything you send hits the server and goes back. At the server, there are people like Facebook, who give the products WhatsApp and Instagram for free. They give it free because for them, you are the product. They store the data to analyze your behavior and offer products that you may not need. The second drawback is it is two-dimensional.

Web 3.0 is meant to solve that by doing decentralization using blockchain technology. That’s number one. The second is delivering immersive 3D content. Engagement delivered by augmented reality and virtual reality put together is called the Metaverse. That’s where the world is moving, whether we like it or not. Right now, the devices are on our head. At some point, it can become smaller and smaller. It’s more an enterprise game show now. But soon when Apple launches its app—Vision Pro, the market price may come down. 80% of the content of the Metaverse could be generated using generative AI.

Prof U Dinesh Kumar

Dean, Data Centre & Analytics Lab, IIM Bangalore

What are the problems with Large Language Models (LLMs)? How much can we trust large language models? LLMs use Internet data. The question is, can you trust the data on the internet, and is there bias in the data? Let me give you an example.

A Sensation, for the Wrong Reason

In an internet poll conducted in the year 2000, Jamie Pollock was voted the most influential man of the past 2000 years. Not many can recognize him. In that poll, Jesus Christ came second, followed by Karl Marx. You may wonder how he got the largest number of votes.

In 1999-2000, he was playing for ManCity, which had just won the Champions League and Premier League. It was a division one team. It was the last match of that season, and they were playing against another team called Queens Park Rangers (QPR). Both these teams were in a critical situation because if they lost the match, they would be relegated. In that match, Jamie Pollock scored a self-goal, and Man City was relegated to division two while QPR stayed in division one. This internet poll was done immediately after that match. All the fans of QPR voted for Jamie Pollock as the most influential man of the past 2000 years. That’s how he got a large number of votes.

The point I’m trying to make is that you must know where you are getting your data from for your large language models. What is the source of the data? I asked this question, “Who are the most influential men of the past 3000 years?” ChatGPT said, “Confucius, Julius Caesar, Leonardo da Vinci, Karl Marx…” and Jesus Christ was missing from the list because I said the last 3000 years. When I asked, “Did Jesus agree with Buddha’s idea?” it gave me some answer. It said, “Yes, there are similarities, but they never met at all.”

Look out for the Source

On YouTube, there is a BBC documentary titled, “Is Jesus possibly a Buddhist monk?” If you include YouTube data, the answer might have been different. The problem when we use large language models is that we have to be careful where the data is coming from. The second is, there could be bias in the data. For example, if you take any other form of medicine other than allopathy, Wikipedia will say they are all pseudo-scientific, whether it is homeopathy, naturopathy, or anything else. Wikipedia also says that 80% of the population of India and Nepal actually use something other than allopathy. Basically, it says that one-sixth of the population has used pseudo-scientific methods. Can you trust Wikipedia data which is heavily used in large language models (LLMs)? LLMs have a large number of limitations. They cannot predict.

But having said that, LLMs can contribute in multiple ways—in text generation, code generation, image generation, speech, video, 3D, and other forms of data. We can create data that would otherwise be difficult. When AI came into discussion about 100 years back, people wanted to create something that would imitate humans, but all that has changed. What we see today is that humans actually imitate Gen AI systems. In the last 10 or 15 days, the entire media world was focusing on Oceangate, the submarine that imploded in the Atlantic. The problem with these innovations is that we have very limited data about the materials that were used and to what extent they can survive the deep-sea environment.

People try to see how Gen AI can help in treating neonatal sepsis, in which newborn babies can die because of sepsis. Sepsis is a bacterial infection, and there is a very high chance of death. In India, the mortality rate is 35. We are looking at non-invasive procedures to solve this problem. Can we look at the baby’s parameters and mother’s parameters and predict the chance that the baby may get into sepsis? Asha workers in India carry a lot of data. Can we pull that data together and come up with a model that can help reduce neonatal sepsis deaths?

Thanks to Gen AI, we can create solutions where data-driven decision becomes very easy and accessible to everybody. For example, if somebody asked me a question, “Should I invest in Infosys stock or TCS stock?” At the back end, I can run AI models and generate data, but GenAI can’t predict things. So I have to integrate other machine learning algorithms to do enhancers kind of things.

There’ll be a lot of jobs that will be created through AI technology. ChatGPT is an information synthesizer. Unless you integrate many other technologies, it is not going to be a thought engine. We are far away from that, but there are efforts to make it a thought engine.

India has been trying to get into the chip manufacturing space. When is it expected to see the light of the day? How much investment is required for India to become a leading chip manufacturer?

Vijay: In semiconductor manufacturing, there are three distinct phases. The first phase is the design, which is done using PCs and software. Once the design is completed, it is sent for manufacturing where the fab comes into play. After the first few prototypes become available, they undergo the post-silicon validation process. India already has a large presence in the first phase, i.e., the design phase. All the major players in the semiconductor industry, such as Intel, Qualcomm, and others, have a presence in Bangalore and other parts of India. Semiconductor manufacturing requires close to $10 to $15 billion in investment. The money is not that difficult for a country the size of India. What is crucial is the technology transfer. India has reopened the PLI window, and Vedanta and Foxconn came up with a deal for some manufacturing, but it got stalled because they couldn’t find a technology partner.

The technology we are considering is very outdated, probably 28 nanometer thickness, whereas the current chips come in 4 nanometers size. India must get close to America and try to influence TSMC to invest in India and bring the technology transfer. It is no longer just a money problem.

How do we see the manufacturing industry five years down the line, using AI?

Sachin: AI has already been implemented in manufacturing, at least for the last five or six years. I have worked extensively with AI in manufacturing, especially for pharma clients, to manufacture high-quality soft gels. Specific computer vision and AI use cases have been implemented. We can detect, from a safety standpoint, if people are wearing helmets and other safety gear. AI will make further inroads and grow much more to improve efficiencies and reduce costs.

Prof Dinesh Kumar: There are a large number of applications of AI in manufacturing. We have worked on projects in manufacturing systems to predict whether there will be a fire in the next 10 minutes using IoT sensors. We have also worked with many automobile companies to predict failures, warranty claims, and detect any fraudulent claims. AI is being used in almost every sector.

Is it possible to translate more than one language at the same time?

Sachin: Yes, the Government of India has a specific initiative called Bashini (Basha Interface) using which a lot of chatbots are being developed, and the language accuracy is extremely good.

Prof Dinesh Kumar: In the UN Council, members from more than 150 countries speak different languages. Whenever a speaker speaks in any language, it automatically gets translated.

Dinesh: Another good application is a 3D avatar of a person delivering a speech.

With respect to Web 3.0, the data will be more personalized. What happens to the data science professions, which are more dependent on data?

Vijay: Obviously, upskilling is required. There won’t be any job loss. There will be realignment or reassignment of various job roles, and people will upskill themselves.

Prof Dinesh Kumar: Data Science will not go away because ChatGPT cannot predict. It’s a language model. Data scientists predict. People use LLM because it’s a fun application today, but soon they will realize the limitations and need to bring in AI/ML and other models to integrate them with large language models to solve problems.