AI firm touts fix for AI-data privacy issue: small-scale LLMs

Virtually every industry today is using generative AI solutions to boost productivity. Yet in sectors like finance, the tech’s full potential remains largely untapped.

One of the biggest obstacles for financial institutions to leverage genAI models is data privacy, according to a working paper published by the Organization for Economic Cooperation and Development (OECD).

For big players like Morgan Stanley or JPMorgan, building their own customized AI solutions has increasingly become the preferred choice to bypass this particular hurdle. 

But not all financial institutions have the resources to invest in such customized solutions. And for some, the generic US$20 per month ChatGPT Plus subscription is arguably not worth all the associated data privacy risks.

Pints.ai co-founders Partha Rao (left) and Calvin Tan / Photo credit: Pints.ai

Founded in 2021, Singapore-based AI startup Pints.ai believes it has the answer to this predicament. It’s building privacy-focused small language models (SLMs) tailored to enterprises that it says is at a fraction of the cost – up to 3x cheaper than other solutions. 

With just over US$1.3 million in funding and within a year of commercially launching its product, the company says it’s reached gross profitability and has a number of enterprise clients from around the globe. Pints.ai did not, however, disclose the financial details.

So how is this startup able to build customized genAI applications at a cheaper cost for enterprises? And how effective are these solutions for financial institutions?

Pros and cons

SLMs are a smaller, less compute-intensive version of large language models. A popular example of an SLM is Microsoft’s Phi-3, which is trained using 3.8 billion parameters.

That number is modest compared to LLMs like Meta’s Llama 3.2, which supports up to 90 billion parameters. 

In short, SLMs are trained on smaller amounts of data and fewer parameters, but that doesn’t necessarily mean they will always underperform compared to LLMs.

See also: Finding internal data is tough. Here’s how LLMs make it easier

According to Microsoft, SLMs are cost-effective and can be more easily fine-tuned to meet specific needs. They can also be useful in regulated industries and sectors, where companies typically keep their data on their own premises or a cloud data center in their country.

Pints.ai’s current offering, the 1.5 Pints, is built on about 1.5 billion parameters. Calvin Tan, the company’s co-founder and CTO, claims that despite the smaller size, 1.5 Pints can surpass the performance of models trained on much larger datasets. He credits this to his team’s focus on the quality over the quantity of data.

Growing its base

Shift from power-hungry models