Thesis.

Intelligence that will eat the world.

Our bet

Shunya Research is an independent research lab based in Bangalore, India. We're working on embodied intelligence, physical AI and neuromorphic systems. We believe true generalization requires more than just text and images, it requires the agent to explore and experience the physical realm to build context based on pragmatics and not just semantics.
While scaling has delivered impressive results, the returns from LLM-scaling have not justified the scale of investments and data poured in. Terabytes of VRAM, billions in capital and almost all of humanity's knowledge and yet the AI that solves cancer is nowhere in sight. All of these point to something fundamental missing in our approach towards artificial general intelligence.
Our bet is on exploiting neuroscience based world models that learn through experience. We also believe by imitating biological architectures of the brain and sensori-motor systems we can make AI systems more sample and energy efficient. The field of neuromorphic computing and neuroscience is underexplored in this sense.

Current shortcomings

Academia used to be where unconventional ideas got a home. However, that isn't the case anymore. Publication counts and committee-driven funding made most institutions risk-averse, they'd rather fund something safe than back a question that takes a decade to answer. This obviously, isn't a shortcoming of individual researchers but a failure of the system as a whole.
Non-LLM approaches to scaling intelligence have been met with scrutiny and almost zealotry levels of skepticism, treated less as a research direction and more as a distraction. However, pioneers of the field such as Yann LeCun, Richard Sutton and Jurgen Schmidhuber have consistently shared the view that text based scaling of current architectures will not generalise.

A few problems we currently face are:

• No continual learning.
• Sample inefficient learning.
• Agents cannot reason about the consequences of their actions in physical spaces, i.e. no causal understanding.
• Energy inefficiency - The Human brain runs on about 20 watts.

Our solution

Our approach rests on three principles:

1. World Modelling paired with RL: Agents that learn through exploring their environments. When paired with RL, the agents learn reward driven behaviour and not just statistical patterns in the world/environment they're exploring.
2. Neuromorphic systems: Spiking neural networks and neuromorphic chips are known to be extremely energy efficient while being on-par with ANNs on similar tasks, their sparsity contributes to their low energy usage. We are currently exploring ways to exploit non backprop based solutions to enable on-chip learning of SNN models (check research tab).
3. Copy nature: The smartest beings in the universe (that we know of) are humans. Our minds are incredibly sample efficient, can continually learn and our reasoning outclasses AI performance in terms energy efficiency. Our goal is to study biological designs such as that of the Drosophila melanogaster and C. Elegans to understand bio neural circuits and quite literally copy them.

Goals

Research is a tedious and rigorous process that takes a lot of time, despite that we aim to achieve the following by next year:

• Publish our first technical results on event-based vision and spiking neural networks.
• Open-source tooling built for our own research workflow, starting with formal mathematics.
• Test and report our attempts at on-chip learning for SNNs using evolution strategies.
• Explore sustainable continual learning methods for world models.

• Build embodied agents that are sample efficient and can continually learn with actual industry deployment.
• Establish Bangalore/India as a hub for frontier AI research.
• Demonstrate neuromorphic hardware-software co-design at a scale competitive with ANN baselines on real robotic tasks.

Conclusion

We're in an age where technology is touching every aspect of our lives, AI has proven to be useful in various fields, be it novel protein synthesis, efficient aerodynamic designs or superhuman levels of medical diagnosis. AI has the potential to usher in an actual utopia for the human race. However, we don't think today's dominant paradigm gets us there. Closing that gap is what Shunya is built to do. We're inspired by the likes of Bell Labs, Google Deepmind and Sakana AI in terms of our research approach. If this resonates, we want to hear from you, whether that's mentorship, funding, or compute.
Ultimately, it's about building intelligence that will eat the world.