Dr Ramarao Kanneganti is a computer scientist who has a PhD in computer science. He studied computer science at IIT Madras, part of the first BTech CS batch there, and then went on to work in different places including Bell Labs.
I joined IIT Madras in 1982. At that time, I think only one institute was offering computer science — IIT Kanpur. So, when we went to Madras, we didn’t know exactly what computers were. All we knew was that all the toppers were taking it. There were only 18 seats in all of South India.
In the mid-1980s, artificial intelligence became a really big thing. The rumour was that the Japanese were building supercomputers, which would become super intelligent. So, America started investing quite a bit of money in AI. In 1984, we had one of the first computer conferences in India, right in Madras. People were talking about supercomputers and AI. But that AI was entirely different from the current AI. It’s like a calculator versus computer.
One of the BTech projects that I did was in AI. But I wouldn’t call it AI these days. It was about how we can teach computers to do things. Historically, how we wrote programs was to give directions. You would think that’s easy but it is not. As an experiment, give somebody precise directions on how to go from your bedroom to the kitchen. How many steps do you have to take? What angle do you need to turn? That is the precision that was required in regular programming. The precision was holding these people back.
Instead, you could give these computers rules. And when we give the rules, we don’t tell you which rule to apply. I’m going to give you the rough layout and based on the rules, you take the steps. So in the 1980s, what we had was rule-based systems in some chemical plants.
In normal programming, it is very difficult to capture all these kinds of rules. You just throw a bunch of rules and the system somehow figures it out. So that is what happened in the 80s. But the complexity: How many rules? 1,000? 10,000? A million rules? That is what the research focus was at that time. You have these rules. But which rule to apply where and how?
So they reduced the problem to arranging blocks. If you go back to literature in the 1990s, you’ll find it funny. Why are these people obsessed with arranging blocks on a table? Turns out that you can reduce any planning problem to this problem. But no useful thing came out of this at all. So the funding dried up in the 80s. They call it AI winter. Nothing happened.
Only a few things came out. One is actually the rule-based systems that I talked about. It was really useful in one area: mortgage applications or loan applications in banks. There are so many government rules. And on top of it, there are heuristic rules. If this person has a job for more than three years, a stable job, then you can give a loan. If he’s changing jobs frequently, that’s a warning sign. These rules are very reminiscent of AI. Except it’s not AI. In fact, most people in Infosys and similar companies are familiar with this system of rules.
Deep Blue that beat Garry Kasparov in chess, was it similar to this?
Yes, but it came much later. Without going into too much technical detail, I’ll tell you what happened. After the 1980s, 90s this all died down.
All of you would have heard of a physicist called Richard Feynman. Do you know Feynman worked in a computer company? He was a professor at Caltech, a Nobel Prize winner, the most defining physicist. In fact, we studied his physics books to get into IIT. For a brief period, he actually worked in a computer company called Thinking Machines. It came out of MIT. It was entirely built ground up just for AI. But the company collapsed. Nothing came out of it.
Then the interest died down, the internet happened and people gave up on the idea of AI completely. But there was one problem that was nagging most people. I worked in Bell Labs. It used to bother the people there. What is it? Speech recognition. By the way, India has a very good research programme in speech recognition. Incidentally, Prof
B Yegnanarayana, one of the founding members of IIT-Hyderabad and IIIT Hyderabad, is considered the father of speech recognition in India. Anybody who came out of India in signal processing or speech recognition would have studied under him or under his students.
They were trying to solve a general purpose problem. The AT&T problem was very simple — namely the phone operator problem. Press one for this, press two for this … or having a conversation. Which department do you want? Sales. So recognition of a few technical words is what they wanted. They built one system. It is not exactly AI but is a precursor. Instead of the rule-based system that I talked about, they used some mathematical structures. If this happens, then I go this way, if this happens, I go that way — like a complicated network kind of a thing. It is called hidden Markovian chains. It got tremendous response but had only limited use.
IBM had a similar programme then. Deep Blue is an adjacent effort to that. So it is not exactly a rule-based system but it is not modern AI either.
The same scientists that worked on speech recognition, they realised what they are doing is looking at the signals and looking at the kind of a relationship with the signals and going to the next state and capturing what kind of decision making they need to do, some mathematical structure. So naturally where did they go? To Wall Street. The most successful investment company ever came out of IBM’s speech research. That is a company called Renaissance Technologies. Every year it returned 70% for 20 years!
But that was not solving other problems. So they found mathematical uses and that renewed the interest in understanding… let’s get away from this kind of a rule system. Instead let’s get into understanding mathematically how we can model it. That gave rise to modern AI. Ultimately, what they are doing is they’re taking a large amount of data and somehow creating this mathematical structure. You don’t know how, but you give the input, it can tell you whose face it is. Give the input, it can tell you what the next word is. So that is the kind of AI that they have built.
It’s kind of like creating algorithms?
The funny thing is that algorithm actually means there is a way of doing it. A, then B, then C. This one is not that. This one is a black box. For example, if I ask you, let’s say you turn this side. If I ask you, why did you do so, you cannot tell me why. It is just something that the brain made you do. You cannot account for all your actions, even the voluntary actions.
Our brain too is a black box. We think we understand what’s going on. But little do we know why we make the decision.
If we give AI a task … the argument is that we don’t know how it is going about it. This is the reason why some people are warning that this is going to be dangerous.
Yes, I completely understand. We are not able to rationalise it. Therefore, we are worried about it. I’ll give you a simple example. Suppose there are two people. You ask them mathematical questions. 1+1, what is it? One person says two. The other person, you ask him a series of similar questions, he gets it wrong 98% of the time. Usually 1+1 = 3, 2 + 3 = 7. The first person gives correct answers most of the time. Except a couple of times, he says that 1+1 = banana. A few times, he is behaving ‘irrationally’ in your mind. You feel more comfortable with the second guy than the first guy. Why? Because you understand this guy’s problem. You understand his mistakes. The first guy, you have no clue. Why does he say 1+1 = banana? You have no way of modelling that.
The major difference between the first-generation AI and the current-generation AI is this: AI is becoming more orchestrating than answering all sorts of questions. Then a few things will happen. The 1+1 = banana problem will go away because well-known problems can be solved using well-known ways. When you see ChatGPT 4.5, fundamentally, not much of a change. But the user experience feels a lot better because it is putting these guardrails in place.
On the IT sector in India, what is the main reason that we are only into back-end operations, maintenance? Why is there no innovation? You have been in the US for so long. What is the difference? Is it the quality of education that we impart?
This is a multiple-dimensional subject. I also thought about this one because for the first decade and a half, I worked exclusively in the US — Bell Labs, some of the smartest people, 5,000 PhDs. two, three Nobel Prize winners. Unix was invented there. C programming, that kind of place. Then I came to India, it was a bit of a shock to me. I adapted to India, eventually. The same people, I take them out of here and put them in Google, they’re effective. The same people, I put them here, they’re ineffective. It’s not about intelligence.
So when you look at the effectiveness of the people, I see there are several different factors that are making people effective there. For example, take a look at Bengaluru. Roughly 20% of the productive time of a person in Bengaluru is spent in traffic. Straight away, 20% of productivity and economic output is gone. It’s the equivalent of 20% of people joining the workforce.
The second thing is education. We conducted tests for thousands of people in Panjab University. Multiple choice, open book, a week’s time. 60% is the cutoff. Can you guess what percentage of people passed, four years ago [This was before ChatGPT]? Only 11 people passed out of 2,000 or so. They were relying on Google. I made the questions non-Googleable. The only way to answer the question is to experiment. They were too lazy to experiment. They were not taught to experiment.
The kind of skills that businesses look for are not taught in colleges. What do we look for? Simply to be effective. That means to be able to work hard, to come on time. Standard discipline things. In college, somehow they think if I’m good at my subject, that is good enough. No, no, no. I want you to be effective.
You are good at your subject, how does it matter to me? Your knowledge is completely immaterial to me. Your output is what matters. So show me the output. That means being able to work with people, being able to communicate. All these things are highly valued in the US. And they are taught from the beginning there.
How difficult is it to float a startup here in India? Vis-a-vis America. In the US, you read stories about students dropping out of college and starting a company from a garage. You don’t hear such stories here.
Fortunately, I never started a company here. But from whatever I hear, it is really complex. For example, I remember when we were bringing laptops to India to work. I was surprised to find people could not take laptops home in Bengaluru. Because it was an export zone. It boggles my mind.
There are many other things. You cannot close down a company easily. Even after our company was acquired by HCL, we had to keep the entity open for a long, long time... In the US, it’s trivial. Go online, submit it, you have a company name. You know what is more complicated? Getting a domain name. Hiring and firing and other rules, it’s quite simple there.
Everybody is after computer science, particularly in the Telugu states, there is still that American dream. Should they be continuing with the same course? Should the courses need to be changed?
In a lot of ways, computer science by itself is useless. What is really useful is looking at other productive industries such as agriculture and seeing how computers can help them.
Basic computer science courses will not cut it any longer. AI is already programming, it will execute the programs too. It is just like using a calculator. You do not have a bachelor’s in calculators. There are some directions that one can go.
One is what MIT calls a bilingual education, that is computers plus some other subject. It can be any subject, that is computer science and logistics, computer science and biology, computer science and actuarial science. So that is the way the curriculum should be changed. I attended a course in MIT that is online, a biology course. More than 60% of the people were computer science people. That is bilingual education. It is basic literacy. You learned how to read and write, now learn how to use ChatGPT and how to execute programs. I hope it becomes even simpler, at some point it will.
Trend two is you basically up the game. Your challenge is not to write simple programs. You need to be able to provide end-to-end solutions. Usually, the biggest challenge is translating a problem into a computer. It is a practice and all the college assignments need to be re-engineered towards that. In our interview process, as I said, 90% of people are passing it [after ChatGPT]. The test has become completely invalid.
So we had to do something else. We bring people in, we say this is the problem we have. They need to go and figure out a way to identify the problem, translate the problem and come up with a solution that works for me. If they can get to that kind of an education, then pure computer science education will have a good future. On the other hand, if you say, give me the manual, I will read it and I code, ChatGPT will be faster and produce code better.
Finally, China has come up with DeepSeek, America has so many others. Should India have one sooner or later? Otherwise, wouldn’t our data may be compromised.
Do you know I am running DeepSeek on my laptop? It is free, open source. We don’t have to reinvent the wheel. We can enhance it.
I went to a college recently. All the students were studying AML. I asked them: how many of you run your own DeepSeek? Not even one. Not even the college. So, that is the change that needs to happen, not another DeepSeek. A land where a million people can actually program using DeepSeek can be transformed.
The bigger threat is that the country doesn’t have an operating system of its own, with everyone depending on Windows. How come nobody is saying we develop our own Windows? And all of these government officials use WhatsApp, which is controlled by one person, who is trying to get into the good books of Trump. Why is nobody worried about it?
This country’s sovereignty is not compromised by AI, it is already compromised by several other things. And that’s what they really should be focusing on. China has their own version of Linux... India should get into the bandwagon of Linux.
We have these many university students. How many of the people actually have enhanced open source software? Why not basically say if you want to get a degree, enhance one open source software. Students would learn how to work with people remotely. And the best thing, they would become a part of a bigger movement.