Organizations are finally coming close with the reality of how to responsibly use Large Language Models(LLMs) for the business use cases that are not merely experimental but provide real business value.
While Large Language Models (LLMs) have been great with generic language understanding; they are costly to customize for a specific domain or simply for the organizational context. SLMs could be filling this gap, making domain specialization economically feasible.
So, how small are the SLMs?
Well, let’s understand how large are LLM’s first.
GPT-1, introduced in 2018, had approximately 117 million parameters.
GPT-2, released in 2019, significantly increased the parameter count to 1.5 billion.
GPT-3, unveiled in 2020, made a massive leap to 175 billion parameters. This represented a 100x increase over GPT-2.
GPT-4, while OpenAI has not officially disclosed the parameter count for GPT-4, estimates suggest it has around 1.5 to 1.8 trillion parameters. That’s almost 10 times larger than GPT-3.
In contrast to this SLMs typically have 10 billions or less parameters. This makes them ideal for specializing specific domains and tasks, improving the accuracy, robustness and reliability compared to their bigger siblings. You can deploy them on a device, on-premise, on your private cloud or put them to public cloud of your choice.
While SLMs will require initial investment and organizational maturity, if carefully developed these might be cheaper to run and better suited for your domain in a long run.
Thinking Space 