Laya: Opening Up System 1 AI

In a world where autoregressive LLMs are increasingly being used for everything, could we be overlooking a different approach to AI?

Long before the current LLM boom, machine learning had already explored ways to make fast, specialised predictions without generating text token by token. Now, a new wave of interest in these approaches is challenging our assumptions about what AI models should look like, how they should run, and what they can make possible.

Laya: Opening Up System 1 AI explores the story of Laya, an open-source AI model developed by independent developer Nandakishore M. from Kerala. Gaining attention on Hugging Face, Laya brings an alternative approach to the forefront: non-autoregressive, structured decision-making models designed for high-throughput inference.

This talk will explore how System 1 models differ fundamentally from conventional autoregressive LLMs, the kinds of problems they are suited to solve, and what their approach means for developers building AI-powered applications. Beyond the technology, we will look at the story behind Laya, the ideas and engineering that brought it to life, and the role open source plays in making alternative approaches to AI accessible to a wider community.

Through Laya's journey, we will examine a broader question: in an ecosystem increasingly centred on generative AI, what opportunities emerge when we rethink the model itself?

• Understanding System 1 AI: Learn about non-autoregressive decision models, how they differ from conventional autoregressive LLMs, and when each approach is appropriate.

• The story of Nandakishore M.: Discover the journey of an independent developer from Kerala building an open-source AI model that gained attention on Hugging Face.

• Beyond generative AI: Explore practical use cases for fast, structured predictions and high-throughput inference.

• The role of open-source AI: Understand how open development enables experimentation, accessibility, and alternative approaches to building AI systems.

• Rethinking AI architecture: Learn why different problems can call for fundamentally different model architectures, rather than defaulting to an LLM for every task.