Kush Varshney wants you to rethink computing again

Quantum computing and artificial intelligence are forcing technologists and businesses to fundamentally reimagine the role of computers in business. At IBM Research, one scientist wants them to think even bigger than that.

In a series of lectures this summer, IBM Fellow Kush Varshney urged his fellow scientists to use this inflection point to rethink the very technical foundations of computing, incorporating mythology, symbolic mathematics, media studies and much more. To Varshney, these disciplines are inspiration for completely new, as-yet-unimagined forms of computer logic.

Touching on everything from the works of information theorist Claude Shannon to the sacred Vedic texts of India to Joan Baez playing at Sing Sing, Varshney’s talks have already inspired his fellow IBM researchers to undertake a number of experiments, such as building an interactive illustration of a sign language for quantum computing and applying category theory to the construction of quantum circuits.

Varshney spoke with IBM Think about what a computer can be, and how unexpected influences can power new innovations. An edited and condensed version of the conversation follows.

I’m trying to look from different angles at what it means to think, what it means to create meaning with language and how there are many different languages and ways of communicating that lead to different constraints. Even though we have this new technology with LLMs, it’s similar to many things that humanity has dealt with in the past. We shouldn’t forget to stand on the shoulders of giants.

Researchers have been bringing up new things in our Slack channel or via email and saying it could be relevant for our work. At IBM, we need to do differentiated work. We can’t be doing AI or quantum in the same way that everyone else is.

The attitude that we’re running with right now is to be provocative. If you have an idea, just go and do it. There’s nothing stopping us from pursuing things that can potentially change the conception of what computing can be, so you might as well just bring other people into your mind and your worldview.

On the LLM side, I think about how there are these almost dueling ways that words have been used [in human minds]. One is that [to use language] you stop and think, and it’s challenging for the human mind, whereas the other one is that it’s natural, and it just happens.

The “System 1 and System 2” theory, which Daniel Kahneman has written about, is very well matched to neuro-symbolic AI. What we’re working on at IBM Research right now with generative computing is taking a slightly different tack: for the neuro- and the symbolic to come together.

We’re trying to combine both kinds of ways of thinking, to maybe do something grander.

One of the researchers who does quantum and AI was pointing out how the category theory language approach could be applied to constructing quantum circuits. I had no idea about this. The fact that she’s thinking it could be relevant to how we could develop quantum circuits as a language, rather than as just engineering, could be its own new idea to pursue.

During the second week of the lecture, when I was talking about sign language, one of the researchers put together a visual interactive demo of how a quantum-like sign language could work, with an animation and everything. It took him something like an hour or two using Claude Code.

Then, on the AI side of things, we have some ongoing work that touches on these things with Mellea, which is our generative computing package.

Mellea is a software development kit developed by IBM Research that is meant to implement the idea of generative computing.

Generative computing is the idea that having LLMs doesn’t mean we have to throw away everything we’ve learned over the last 80 years of computer science and computer engineering, but just to think about what happens when we have this one extra computing element. You can use an LLM when you need to and not use it when you don’t need to. Just use regular code when it’s most appropriate, and then when you do use an LLM, you can wrap it in a way to make it more reliable.

What we’ve been talking about in the lecture series goes one step above that, which is to ask, ‘What are LLMs good for?’ That’s a basic question that we’re trying to think through in the lecture series. The answer is that they’re good for narrative, understanding context and making things more personalized. What they’re not good for are things like Boolean algebra or commutative reasoning.

The basic premise is that by studying orality, or the structures of myth, or file-change semantics, et cetera, you can help figure out what should be done as regular programming and what should be passed on to the LLM to do, and how should they mix together to make a more robust, useful version of AI.

It’s able to process large amounts of knowledge and spin it into a coherent, fluent language product. That ability to spin things together is usually reserved for the best of humans, or it goes beyond the human capability most often. So the magic is that the LLMs are doing something that we thought only we could do.

I think that’s breaking the boundary between intelligence, or language processing, as a human-only activity.

At Cornell, I took this Sanskrit epics class. It’s now been 25 years since I took that class, and I learned about Albert Lord’s book The Singer of Tales, [as well as the work of American classicist] Milman Perry.

In the last few years here at IBM Research, since my work relates to trustworthy AI and our notions of what’s right and wrong in terms of what needs to come out of these models, I’ve been getting into moral philosophy and collaborating with external philosophers to determine the appropriate behaviors. What does it mean to have dignity? What does it mean to have care? What does it mean to be trustworthy?

One book that I was reading is called The Divine Economy. It’s about how religions as institutions can be analyzed economically. In that book, there was one line that mentioned The Singer of Tales. It reminded me of that class and made me think that we should be thinking about LLMs along these lines.

We have not reimagined computing in a fundamental way since [John] von Neumann, and even he didn’t want the world to stay stuck in his eponymous architecture. Now that we have these two technologies breaking through, quantum computing and generative computing, it forces us to reimagine it, because the businesses that exist today exist because of the first iteration of computing.

So now we have a chance to think again about these other paradigms. If we do it well, that can unlock all sorts of business opportunities.

###

For more of Varshney’s thoughts on the next era of computing, subscribe to the IBM Think Newsletter or catch him on the Mixture of Experts podcast.