IBM led some of the 20th century’s most pioneering projects in artificial intelligence and supercomputing, including the development of the Deep Blue chess-playing system that became the first to defeat reigning world champion Garry Kasparov, in 1997. Then, in 2010, the IBM Watson AI system for answering questions in natural language was put to the test on the US TV quiz show Jeopardy and won top prize.
But, in the years that followed, a failure to follow through on these projects meant the company fell behind. Now, chief executive Arvind Krishna is seeking ways to use AI to deliver labour-saving solutions to large enterprises, as well as making powerful and error-free quantum computers a reality.
Here, he tells the FT’s west-coast editor, Richard Waters, why IBM struggled to turn its Watson breakthroughs into a business success, and how he is trying to capitalise on a new era in AI.
Richard Waters: I want to jump straight in on AI. ChatGPT [OpenAI’s artificial intelligence chatbot] has focused everybody’s minds . . . Has something significant happened and — if so — what is that?
Arvind Krishna: I think the marketing moment offered by ChatGPT is incredible. We’ve seen these moments before: a company called Netscape brought the web browser to everybody’s attention. I mean Netscape was not an eventual winner but the internet certainly was. I think what ChatGPT has done [is] helped make AI real to lots of people who kind of were aware of it but didn’t maybe quite see what the power of AI would be. So full credit to what they did there.
To be a bit more geeky, the underlying technology is based on two things: one is large language models and the other is generative. When we look at these, I’ve got to give credit to a lot of people. Google has worked on these for probably close to a decade . . . Facebook has worked on this, many universities have worked on it from Stanford to MIT to Berkeley to Illinois, I can go on and on — including some, Richard, from your home country and Cambridge and Oxford, probably others around the globe as well.
And we have worked on them also. The use cases we work on are not consumer, so consumer is a lot easier to explain: they type in something and, some number of times out of 10, you get an interesting, intriguing, and in the right ballpark answer. What nobody can quite say is how likely is it to get a completely incorrect answer, as at least one of the two demonstrations has shown.
If you’re using it for consumer search, fine. If I’m using it to answer a question on somebody’s financial transaction, that’s actually quite a problem. Or, if it’s being used to answer somebody’s question on what healthcare treatment they might seek. So, for enterprise use cases, I still think that there is a massive opportunity that is outside the pure consumer space.
But, to answer your top question, AI I believe is really important and I’m not new in saying that. Within AI, I think we’ve gone through these phases . . . from machine learning to deep learning . . . I actually think the large language model is the more important piece.
The real advantage of these models is you train the model once and the first time is quite expensive but, then, from that, you could generate 100 or 1,000 models very easily. That’s a massive cost advantage and that’s a massive speed advantage. I think enterprises are going to understand that very quickly and work with those who can lead them down that path.
RW: When you talk to businesses and CEOs and they ask ‘What do we do with this AI thing?’ what do you say to them?
AK: I always point to two or three areas, initially. One is anything around customer care, answering questions from people . . . it is a really important area where I believe we can have a much better answer at maybe around half the current cost. Over time, it can get even lower than half but it can take half out pretty quickly.
A second one is around internal processes. For example, every company of any size worries about promoting people, hiring people, moving people, and these have to be reasonably fair processes. But 90 per cent of the work involved in this is getting the information together. I think AI can do that and then a human can make the final decision. There are hundreds of such processes inside every enterprise, so I do think clerical white collar work is going to be able to be replaced by this.
Then I think of regulatory work, whether it’s in the financial sector with audits, whether it’s in the healthcare sector. A big chunk of that could get automated using these techniques. Then I think there are the other use cases but they’re probably harder and a bit further out . . . things in like drug discovery or in trying to finish up chemistry.
We do have a shortage of labour in the real world and that’s because of a demographic issue that the world is facing. So we have to have technologies that help . . . the United States is now sitting at 3.4 per cent unemployment, the lowest in 60 years. So maybe we can find tools that replace some portions of labour, and it’s a good thing this time.
RW: Yes, I remember very clearly, six or seven years ago, hearing from lots of tech companies that we have mastered language, we have mastered the intelligent assistant. We saw a lot of chat bots at that time. We started to hear a lot about Siri and Alexa and Google Assistant — and how these would affect customer care, and some of the applications you’re talking about. None of that seemed to pan out.
Why didn’t that wave of AI work around language — and why should we believe that this will be any different?
AK: I might take a slightly different position. I think it did work but in a very narrow use case, not in the broad sense that people are getting excited about today. The excitement today is because we can ask it: ‘Should I go to Paris? In which month? Plan me a trip!’ — and it can do something. The prior generation really couldn’t do that.
But, if you asked it: ‘Can I get an appointment between 9am and 10am?’ it could look up if there was there a free slot, it could look up if are you eligible for that appointment, and it could ask you a couple of questions to clarify that, and then book you that appointment. So I do think that, in the previous generation, we had many clients where about half their call centre calls could be handled completely by AI. And, actually, in many cases, the people are happier.
AI always has a pleasant tone, doesn’t get sharp with you, it answers 24/7. So I don’t take it as: ‘It didn’t work’. I think it did work — but in the context of those AI systems.
Now, the scope opens up to answer much broader questions and maybe ill-formed questions. Because, previously, your question had to be pretty precise . . . This is kind of the next plateau.
In most systems that get deployed in real life, it’s not a single big bang. Those are very rare. Even if you take smartphones today. People used to walk around with pocket PCs and little devices. Everybody’s forgotten about the PalmPilot! People are not giving credit to [the fact that] in real engineering, you get these steps. Each step is a very big step.
And, then, finally, there is a step and everybody says: ’Wow that’s what I’ve been waiting for!” — forgetting that, actually, you needed those previous 10 steps because they each brought a piece of the puzzle closer.
RW: Yes, if we go back a while, IBM itself got a lot of attention for things like [the supercomputer] Deep Blue beating Garry Kasparov at chess and the AI computer system] Watson winning the Jeopardy game: these were very advanced demonstrations of AI at various stages. So what happened? Did AI go off in a different direction from what IBM was exploring? What changed?
AK: No, I think maybe we took a tack that was both too hard and the market was not quite ready . . . Maybe we were more than 10 years early.
Two things . . . didn’t go well for us. One was thinking that the world wasn’t ready to do it themselves, [so] we went down the path of trying to build ‘black box’ AI solutions. But those were too monolithic and the world wasn’t ready to swallow that. People wanted to know” ‘What am I building?’. They wanted the piece parts. It was more of a builder’s world.
The second is that, when we looked at some of those problems — for example around healthcare — I think that we were maybe a bit naive. You have to have the right expertise, you have to work with regulators, you have to actually understand what they need to get satisfied.
So, if I put the two together, going after healthcare is maybe a 15-year journey, not a five year journey. I actually believe it will happen and you can see some early instances of companies who are making the play but it’s probably still a few years away.
AK: Once we’ve pivoted to saying. ‘OK, let’s do more small things, let’s do customer care, let’s do it around audit, let’s do it around finding documents, let’s do it around reading documents, let’s do it around HR’ . . . then, suddenly, we’re going to have a lot more commercial success.
RW: How far away do you think we are from real practical uses of this technology rather than demonstrations?
AK: I think this is here and now. I just came back from a conclave in Washington DC and I’ve also had this question from at least a dozen people, all of whom are willing to do that work in reality. Because now it’s inside people’s imaginations of what’s possible.
RW: Do you think that we’re going to see winners and losers? And, if so, what’s going to distinguish the winners from the losers?
AK: There’s two spaces. There is business to consumer . . . then there are enterprises who are going to use these technologies. If you think about most of the use cases I pointed out, they’re all about improving the productivity of an enterprise. And the thing about improving productivity [is that enterprises] are left with more investment dollars for how they really advantage their products. Is it more R&D? is it better marketing? Is it better sales? Is it acquiring other things? We are opening up their investment appetite by making the fixed cost of an enterprise a lot more efficient . . . There’s lot of places to go spend that spare cash flow.
RW: We’ve heard for so long about technology improving productivity and yet it hasn’t shown through in the aggregate data . . . I don’t want to sound cynical here but why should this be different?
AK: It’s your job to question people like me on this. I don’t call it cynical, but I might disagree with your premise that it hasn’t shown through in productivity numbers. I think economists only measure what they know how to measure and they’re going to an aggregate. But those aggregated things may not be counting the people now who are flexible at work, a lot of people now are working 40 hours a week. It used to be 600 people to design a chip; it now takes about 200.
So I know that using all these techniques has led to about a three times improvement. How many people really do the audit function now, if you look at the big four auditors? At least half the tasks — if not more — have been automated away. So I’m a little bit sceptical that it doesn’t show up. What happens is that every enterprise now says: ‘Well, I’ve got these people, let me turn them into more revenue-generating roles as opposed to take them out’.
RW: Let’s switch to a different topic: China. There’s been a lot of talk at the moment about relative competitiveness with China. I wanted to start with IBM’s own position. If we went back 15 years, maybe 20, IBM was investing quite a lot in trying to build out the market there. Like a lot of American tech companies, it believed this was both a great market and a great opportunity to seed its technology. [But] you pulled back. So, what was IBM’s experience of China and why did it lead to a disappointment?
AK: I wouldn’t quite use the word disappointment . . . Yes, IBM went to China in the early 90s — and we still have a lot of people there, we still have a business there. Could our business do better there? I believe it could, but it does need a bit of willingness and co-operation. We have no intention of leaving China, we intend to be there as long as the government in China allows us. We would like to maintain and grow our business there if at all possible — that is the position that we’re in.
When people talk about two different tech sectors, that doesn’t help anybody, I don’t think it helps the United States and it doesn’t help China. So I think we want to be in one, always.
RW: Isn’t there going to be — if anything — a deeper schism over the next few years?
AK: Well we seem to go in fits and starts at this topic. I was pleased that at least President Biden and President Xi met at the G20 summit in Indonesia and they had a conversation. I think that helps diffuse some tensions. Both of them talked about the things that could return to normal. If I look at comments since then, other than a few areas around advanced semiconductors and supercomputing, other things are allowed. You just have to be careful to whom you’re selling them, and what you’re doing with them.
I think technology has reached the point where most nations think it is as important to them as financial systems. Countries want their own banks, they want their own currency, they want an ability to print and have their own monetary policies. I think that countries want that flexibility around technology systems, as well. As long as they believe they have some of that flexibility, they’ll use some of the underlying building blocks — and that’s what we would argue for.
RW: When you assess the competitiveness of the US tech industry globally, and you compare it with the very ambitious plans that have been laid out in China, how do you assess that landscape? For instance, you talked about supercomputing, and we’ve recently seen China leap ahead with the first exascale computers. The US now has one, but China wants to build many. Is the ground already shifting?
AK: Actually, I don’t think so, I think we have reached the point of diminishing returns on classic supercomputers. What problem could you solve on a two exascale classical supercomputer that you couldn’t on half an exascale? I don’t think anybody has named one. Because you’re going to reach the performance limits of what we can imagine using it for. So I think we have probably reached the point of diminishing returns.
But, to your broader question, I think it comes down to the actual deployment and embracing of new technologies. You’ve got to give full credit to China, they seem to run really quick and really fast on that. But, when we come back to systems that you can trust, that you can use, that have gone global, that help people with their lives, there aren’t that many examples.
You think about what Google has achieved . . . you list those who have done work . . . everybody from Deep Blue [the chess computer program] to [DeepMind’s] AlphaGo [Go computer program] . . . So I think that we are quite competitive. But the question is: are we willing to embrace and deploy at scale very quickly? I think that is the real question, not are we competitive.
Number of qubits, or quantum bits, in IBM’s latest processor
RW: Watching IBM’s portfolio has always felt like watching the evolution of the tech industry. And you’re making a big bet on quantum computing. You’ve got a 400 qubit [quantum bit, equivalent to a binary bit] processor out now which, five years ago, we’d have said: ‘Wow, we must be getting into the quantum age!” But it still feels like it’s a very long way away. So how do you think about the timing of your investment and the timing of a return?
AK: Quantum reminds me a lot of the semiconductor industry in the 1950s. If you look it up, the big winners — with the exception of maybe one or two like TSMC — had their roots in that very beginning. They learned what to use and how to use it, they learned how to scale it, they learned how to manufacture it. It wasn’t just about advanced R&D, that happened in universities, in many other places, in addition.
And the learning experience of how to do things at scale is massive. That experiential learning your in-house engineers get is incredible. It’s what I call experiential knowledge.
So, now, to quantum. We’d set out a road map four years ago and we’re on that road map. We said: ‘OK we’re going to do 60-odd qubits, then we will do 100-odd qubits, then we will do 400-odd qubits, then we’ll go to 1,000 qubits’ — that’s by late 2023, early 2024.
So we’re on that road map that we set out four years ago. Now, you ask the question . . . when are we going to have somebody do something that is of true commercial value? [But] just like every other technology there has been from the mainframe to the iPhone, every couple of years you get something that is better and better. It’s not revolutionary but it is much better. If we look at it in 10-year arcs, it is dramatically better.
So your question is when is somebody going to do something . . . I think around maybe five years from now we’ll see somebody do something that is absolutely remarkable on a quantum computer.
These systems are going to remain noisy in the near-term future. If they’re noisy, not all quantum algorithms are going to run really well. So you begin to ask yourself: which [applications] could tolerate some noise? I think problems in materials, and problems in risk, can probably tolerate more noise than some of the other well-known problems that others are going after. That’s why we want to remain focused on a variety of use cases.
RW: I hear people use the phrase “quantum winter” more these days and I always think it reflects the capital markets: excitement gets ahead of itself then things retrench. But the result, nonetheless, is people starting to talk about a quantum winter. Is it true?
AK: I think that real hardware investments take significant amounts of money. I think it takes hundreds of millions a year to be able to make progress on that. If you think that’s a 10-year journey, that’s not made-up money, . . . you need $1bn real cash that you’re going to spend, maybe more.
Then you need all of the expertise around materials and electronics and cryogenics and quantum physics. So you need a team that is pretty big.
We’re still at the phase where people are debating which material for a transistor, in this case what kind of qubit . . . it’ll probably take another six, 10 years to play itself out.
RW: But some commercial advantage before the end of the decade and maybe substantially earlier than that?
AK: I think it could be much earlier than that. You wrote about the Chinese paper which said that with 370 qubits you can do something. We can all debate the veracity of that paper, but I don’t think it matters. What it says is that people are now thinking about what you do with hundreds of qubits. That is the real breakthrough
There are some people who claim nothing will happen until we have a million qubits and we’re completely fault tolerant and we are noise-free with error correction. But that would be like saying: ‘Let’s not send Chuck Yeager circling round the earth once until I can build a spaceship that can visit the next galaxy! I’m not going to spend a dime!’ What?
You’ve got to first just learn how to launch a rocket, then you’ve got to learn how to circle the earth once, then you’ve got to learn how to come back with the person safely. Then, maybe, you put them into geostationary orbit, then maybe we learn to launch a satellite, then maybe we send an unmanned rocket to the moon, then maybe we send a manned rocket to the moon that could come back.
So that’s why — to me — this idea that you must be able to do a million qubits error-free is ludicrous.
I think that once we’re in the hundreds, maybe thousands, maybe with better quality than today but not perfect, they’re going to come out with really innovative use cases.
The above transcript has been edited for brevity and clarity.