As AI adoption skyrockets, the limitations of large language models are becoming more apparent. While LLMs can generate text, code, and conversation at scale, they remain fundamentally probabilistic systems. They are essentially “stochastic parrots” that mimic patterns but lack true, consistent logic.
Enter neurosymbolic AI, which aims to achieve artificial general intelligence by combining machine learning with human-like symbolic knowledge. That’s exactly what Sachin Dev Duggal is working on now.
“You’ve got the symbolic side of the world, which is actually really old AI—knowledge graphs, concept graphs—and we’re going to bring them together,” he explains in a conversation with AIM.
Duggal, who co-founded AI-powered software development platform Builder.ai that entered insolvency last year amid allegations of financial misconduct and “AI washing”, now wants to rethink how intelligence itself is constructed.
His next venture is at the research and development stage, working with a small number of early customers. While the company remains in stealth, Duggal says the nine-membered startup, founded in September last year, will offer a public preview in early 2026.
This is his fourth entrepreneurial endeavour, after Builder.ai—backed by Microsoft and Qatari Investment Fund, among others—and shoto and Nivio.
The work, he says, sits at the intersection of intelligence, knowledge, and neuroscience—and represents a deliberate shift away from purely probabilistic AI systems toward architectures that can reason, explain, and adapt. That’s where atomisation—using AI to break down complex content—becomes important.
Atomisation: From Tokens to Concepts
Duggal argues that current AI systems are built on the wrong primitives. “What’s the atomic unit of software? Its features. What’s the atomic unit of a conversation? What’s the atomic unit of a document? Concepts.”
Instead of treating words or tokens as the foundation of intelligence, neurosymbolic systems go deeper and operate at the level of concepts. This distinction matters because modern language models implicitly assume that language is how knowledge is stored. Duggal challenges this directly. “We often think about words as how we hold information. But words are how we transmit information.”
He illustrates this with mathematics. One can express addition using symbols or words, yet the concept of addition exists independently of either representation. Meaning is stable; language is merely a transmission mechanism.
Neurosymbolic AI is designed to work at this conceptual layer. By atomising information correctly, systems can retrieve meaning more accurately, understand context more deeply, and convert scattered information into verified knowledge rather than surface-level pattern matching.
Intelligence as Meaning and Reasoning
For Duggal, this architectural shift defines the next era of AI. “We’re going to get to a world where we’re really going to start mapping out what intelligence means, which is meaning and reasoning,.” he says.
This reframes intelligence away from fluency and scale toward understanding and explanation. Neurosymbolic systems are built not just to answer questions, but to reason through them in a way that can be inspected and adapted.
One implication of this approach is the personaliszation of understanding. Duggal questions the assumption that everyone should receive the same version of information. “We’re not all in the same headspace. We’re not all at in the same level of understanding. So what if my first two paragraphs were different to your first two paragraphs? Because it knew what you knew.”
In learning, education, and skill development, neurosymbolic AI capability could fundamentally change how knowledge is delivered, shifting from one-size-fits-all explanations to systems that adapt to the learner’s conceptual state.
‘Bias is Good’, Culture, and Multiple Truths
Duggal is also explicit that bias is not inherently harmful. “People think bias is a bad thing. Bias is a really good thing. Bias is makeing sure that you look at the world the same way every day.” Bias, in this sense, provides cognitive stability.
He extends this idea to culture and history. Explaining Indian independence to a child in Bengaluru, London, or New York inevitably produces different versions of the same truth. These differences are not distortions but reflections of cultural context and historical perspective.
“The problem in language models is there’s a single version of that truth,” Duggal argues. “So which version is it?” Purely probabilistic systems collapse plurality into a single statistical narrative. Neurosymbolic systems, by contrast, can support multiple contextual interpretations while maintaining structured reasoning beneath them.
The ability to accommodate cultural nuance without sacrificing coherence is one of the reasons Duggal believes neurosymbolic AI will matter deeply as AI becomes embedded in society.
From Documents to Recipes
Duggal believes neurosymbolic architectures will also transform how work and knowledge are organised. “Most file systems are receipts and documents. I think the world is going to go to recipes.”
Receipts record outcomes. Recipes encode process, logic, and causality. They can be reused, recomposed, and improved. Modern AI systems often produce answers without explaining how they arrived at them. “Answers without sources are just opinions. Governance without verification is just theatre,” Duggal says.
He draws an analogy from education, where students were often rewarded more for showing their working than for the final answer. “It’s funny, the AI world—there’s no working out. Suddenly, we’re going to have long division visible, so that experts can actually see, correct, and reuse.”
Neurosymbolic AI makes this possible by keeping reasoning explicit and inspectable.
Why Neurosymbolic AI Matters Now
As AI agents become more prevalent, Duggal warns against recreating human inefficiencies at machine scale. “Why do we need to treat them like humans, the way we’re doing right now?” he argues.
Neurosymbolic systems allow agents to share structured reasoning directly, reducing the need for constant synchronisation and redundant communication.
Neurosymbolic AI is not a rejection of deep learning, nor a return to brittle rule engines. It is an architectural correction driven by real-world demands. As AI becomes central to how people learn, work, and govern, fluency alone is no longer enough.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.