Voice AI: Build or buy? How to choose the right solution for your contact center Learn more

Play Podcast

Size Matters: The rise of small LMs

In this episode of ‘Deep Learning with PolyAI,’ Damien and Nikola discuss the evolution and significance of small language models (SLMs) amid the ever-growing landscape of large language models (LLMs). They dissect the predicted rise of SLMs, their applications in privacy-focused and low-power devices, and the implications for enterprises balancing cost, efficiency, and modernization. They also explore the enduring relevance of SLMs, the pursuit of artificial general intelligence (AGI), and what the future holds for localized and on-premises machine learning solutions.

There's a growing trend towards using smaller language models (SLMs) alongside large language models (LLMs), reflecting a shift from the "bigger is better" approach to more practical, specialized models.

Smaller models, like the original language models behind PolyAI’s product, were common due to limitations in compute power and data access, but modern advancements have expanded the capabilities of both small and large models.

SLMs offer advantages such as faster processing, lower energy consumption, and the ability to operate offline, which can enhance privacy and security.

Nikola Mrkšić

Co-Founder and CEO at PolyAI

Damien Smith

Senior Communications Manager at PolyAI

Never miss an episode.

 

Subscribe


"There's nothing new about small language models. The initial language models were a lot smaller than these. These are still monster models compared to, you know, what an average person about 10 years ago would have encountered ."

Co-Founder and CEO at PolyAI

Traditionally in deep learning, you would always kind of like pre train on a lot of data. And then you would specialize these models and probably also distill, make them smaller so that you can put them on small devices so that you can, um, potentially, you know, put them, you know, like, um, you know, a phone or like a fridge, microwave, whatever, right?
Um, at the very kind of internet of things, right? Where it would be able to run on. Like without internet, right? Locally that was always the appeal potentially on devices that don’t use up a lot of energy at all. Right. So I think people used to be a lot more excited about the internet of things and like the non cloud version of things where it’s not like connected to things online.

"Traditionally in deep learning, you would always pre train on a lot of data. And then you would specialize these models and probably also distill, make them smaller so that you can put them on small devices."

Co-Founder and CEO at PolyAI

"Today's LLM is tomorrow's SLM. If you want to use this notation, so, you know, give it like two or three years and something like GPT 3.5 will have no trouble running on a personal device."

Co-Founder and CEO at PolyAI

"In summary, basically, you're probably already dealing with SLMs. Most machine learning applications have are based on SLMs right now, but there is the pursuit to drive bigger is better. That being said, what we're looking at more localized devices."

Senior Communications Manager at PolyAI