“Ideas are funny little things. They won’t work unless we do.”

This is the quote on a plaque in my living room. It’s the only heirloom I valued enough to bring from my childhood home in India. As you can see from the image, the plaque has gained round edges and chips from all the falls it has endured over four decades. My own learnings, shaped by the computational transformation of our lives, has been similarly molded.

I feel the need to create a new plaque that reads:

“Ideas are funny little things. They won’t work unless we leverage machines and do.”

I was a curious kid with a continuous barrage of questions: Why? How? What? This desire to know, however, was somewhat suppressed. I grew up in a society that encourages children to speak only when spoken to, and in a middle-class family, there weren’t many who had the time to entertain a child’s never-ending questions. This steered me toward becoming a more introverted boy who learned by observing, creating mental frameworks to make sense of the world, and developing systems thinking to connect things on multiple levels. This inward journey of hypothesizing and justifying outcomes became natural, though sometimes in ways that weren’t grounded in reality. In a way, I was like a modern LLM, living in my own world of hallucinations.

I was a curious explorer with a passion for cricket until I changed schools after the 8th grade. It was then that I became serious about studying for the engineering entrance exams, and I was fortunate enough to be accepted into an IIT. Throughout my undergraduate and graduate days, most of the coursework bored me. The one thread that continuously inspired me, however, was our growing ability to model and simulate the world around us. When you model something from first principles, you can answer questions like “Why?”, “How?”, and “What?” and perform all kinds of what-if analyses. I could finally build on my childhood explorations, but now they were grounded in solid scientific and theoretical underpinnings.

Fate would have it that I ended up working on one of the most challenging computational topics: the turbulent combustion of two-phase flows, a field at the intersection of fluid dynamics and chemistry. It was a humbling experience. I could only make a small dent in the field, and there has been only limited progress in the 30 years since.

However, this journey raised new questions in my mind. By working across the disciplines of domain sciences, applied math, computer science, and statistics, I developed a deep appreciation for what I call computational thinking. This is the ability to take a problem and convert it into a computational one that can be solved in-silico. This conversion requires a blend of skills: domain knowledge (physics, chemistry, biology), the ability to map that knowledge onto theories and equations with appropriate approximations (engineering), an underlying understanding of the algorithms to solve those equations (applied math), the knowledge of how to best solve it using massive parallelization (computer science), and robust analysis (probability and statistics). It’s also critical to ground this process in the specific question you’re trying to answer, the level of fidelity needed, and the constraints of time and money. By tackling one of the toughest computational problems, coupled with access to the largest supercomputers of the time and a background in math, CS, and physics, I started developing the skills for a career in computational science.

I was recruited from Georgia Tech while still finishing my Ph.D. to help a national lab respond to an important DOE initiative called SSI (Strategic Simulation Initiative), which aimed to develop a high-fidelity digital twin of an Internal Combustion Engine. I visited Sandia National Laboratories in Livermore even before officially joining Oak Ridge National Laboratory (ORNL) to help lay down a stake for the lab in this $70 million initiative. In a twist of fate, the project’s funding was eliminated by Congress. In retrospect, this was fortunate for me, as it forced me out of my comfort zone to explore applications beyond my niche. The bifurcation chart below shows my evolution from those early days at Georgia Tech to a Distinguished Staff Member at ORNL, and now, for the last 10 years, as a Research Fellow at SABIC.

At ORNL, I was a true explorer, working on a broad range of problems and across disciplines with some of the best minds in computational science. Later, I was recruited by SABIC to focus on a narrow set of capabilities I had developed at ORNL. Working in a corporate research organization helped me focus on a few key problems and deliver solutions. We developed and optimized a novel reactor entirely in-silico, de-risking it through extensive experimentation. If successful, it will displace the century-old steam cracker for converting hydrocarbons to olefins. We also developed unique solutions that leverage the low melting point of plastics to create self-quenching, safer batteries. Through work with the World Economic Forum and other petrochemical companies, I also gained exposure to policy and the need for broader stakeholder engagement to bring sustainability solutions to market.

Over the last three decades, many ideas have been brewing in my mind. I never had a chance to develop them further because I was limited by my own abilities or because they were adjacent to my core expertise, requiring dedicated time I didn’t have. Time is a finite resource, and it was difficult to translate those ideas into proofs-of-concept (POCs) and actionable plans.

Fast forward to today. I am having a blast turning many of those lurking ideas into POCs and actionable plans. This is possible thanks to the computational thinking I’ve honed over the last 30 years and my curious explorations over five decades.

What made the difference? LLMs, AI/ML tools, the open-source software ecosystem, and low-cost hardware have dramatically reduced the barrier to fleshing out ideas and creating prototypes. It is an exciting feeling to pour life into one’s ideas, weed out the bad ones, and advance the most promising. The surviving POCs serve as a starting point for feedback, allowing one to build, refine, and pivot as needed. We are bridging the gap between biological and machine intelligence to create solutions for society’s most pressing problems.

My plea to you is this: if you have ideas in the back of your mind or collecting dust on a shelf, now is the time to revive them. See if the current maturity of LLM tools and the computational ecosystem can help you advance them and let computers do most of the heavy lifting.

From my experience, using LLMs reliably is both an art and a science. Here’s what I’ve learned:

1. They are great at synthesizing information, but be cautious. You must ask, “How can I leverage all the knowledge on the internet to my benefit?” Our brains can only store so much information, and we forget more as we age. But watch out for these traps:

• Oversimplification: They often over-generalize and simplify complex information.
• Appeasement: They try to give you the answer they think you want. Question everything and have the LLM critique its own output.
• Confident BS: When the training data is limited on a specific topic, an LLM will still give an answer with confidence, even if it is completely fabricated.
• Incoherence: If a prompt exceeds the model’s context length, its output can become incoherent and nonsensical.

2. You are ultimately in control and responsible for the final product. Your reputation is on the line.

• Resist the prose: It is easy to get carried away by the beautiful and well-formed text LLMs generate. Stay critical.
• Hallucinations are real: Society’s tolerance for mistakes from a human expert is much lower than for the errors LLMs currently make. You don’t want those mistakes tied to your name. To err is human, but our reputation can only handle so many errors.
• Maintain control: Especially when coding, limit the control an LLM has so that it cannot do unintended harm.

3. Understand these two key concepts to know where LLMs can be most helpful.

• Transfer Learning: If you have limited data for your specific problem, see if the LLM has been trained on similar problems. Models trained on analogous tasks are better at capturing the essence of your high-dimensional problem, even with limited data.
• Latent Space Extrapolation: LLMs are good at extrapolating and interpolating in the “latent space” – the abstract, compressed representation of data. This might be related to how biological intelligence works, as the underlying mathematics of neural networks and biological intelligence share similarities.

4. Be aware of their fundamental limitations.

• No Embodied Feedback: We continuously adapt by reading a room or from physical feedback, like the pain of a burn or the sting to our ego when we make a mistake. Machines lack this constant, real-world feedback loop for improvement.
• No Integrated Context or Consciousness: They do not integrate experiences to build context and consciousness the way humans do.

Once you understand what LLMs can and cannot do, what is computationally possible, and have a little patience, you can begin to multiply your efforts. By engaging machines and their unique form of so-called intelligence, you can start turning your ideas into reality. I have posted some examples of my work in the past and will continue to post more to give you a glimpse of what I am able to do in my free time.