SLM series - InFlux Technologies: It's a question of specialisation, especially

Let’s specialise, especially

In my opinion, combining both SLMs and LLMs is more of a perspective or choice of words… I’d much rather say that using a generic model would make sense with multiple specialised models and use the generic model to reroute these to the specialised one for the task. Again, to me, small, medium and large are all very subjective and just add to the confusion; stick to its task, is it a generic or specialised model and what is it specialised in? Why add all the complications of subjectiveness to something we already have words for and use to describe it?

Working in tandem with both SLMs and LLMs, yes, I’ll start using these from here to keep it easy to follow, but keep in mind they’re specialised and generalized… I believe a selection of specialised models is better than one LLM if we’re only addressing and using it for what they are designed for; as soon as we head off track, so does the result. The LLM is the jack of all trades, master of none. Depending on the use case or the criticality/importance of the result and severity of the impact based on its output, a specialised model is what you’ll need. In contrast, a single LLM would be more than enough.

This is a good abstraction of an everyday use case we all experience. If you see your family doctor, in most cases, for the common flu, you’ll get a prescription and then you’re done and happy. But if it’s something more serious, he will likely have you visit a specialist to check it out. LLMs and SLMs should work in tandem. You get a general opinion and if you need further, you get redirected to one or more specialists.

Finding the balance

SLMs are faster to train, yes, but there’s always a point where you need to be able to bridge the information, the gap between the specialised models and the one above, slightly more generalized, or have a “larger SLM” it’s always a case where we need to find the just balance… and the whole point of this discussion, when do specialised models stop being specialised and is overarching another field, or “become an LLM.”

There should always be some overlap between multiple specialised models and restraining it too much is also something that will limit its usability. Just like when training a model, at some point, overtraining a model will just overfit it to the data and it will no longer be of any use; it’s all about finding a good balance and defining a proper scope and good feature selection without cutting out too much.

Now, whereas these should always be deployed on-premises, no it truly depends on the use case, the sensitivity of the data, urgency in terms of time-critical responses, who will use it and is a lot more inherent to the context of its application than just saying SLMs are only / should be used on premises… specialised models should and can be used everywhere and anywhere, IMO. ChatGPT even offers specialised models in their service offerings, so that says a lot already … not that it’s a good example, but it’s one of many examples.

Alex Perritaz, chief architect at InFlux Technologies.

Are SLMs environmentally sustainable due to their footprint? To what extent are we talking about it? If we’re speaking about a single instance, I’m sure. Still, then, usually, there’s not a single one that is being used that is being trained; in fact, the more specialised models there are, the less environmentally friendly it is, as we need more training, same iterations, more loads, just within a more defined scope.

In the end, we’d need multiple specialised models and if working with numerous, perhaps even an LLM to reroute it to the correct one. Then, if we extend this even further, how is the data for these models acquired, where was it stored and what type of energy was used to keep the data stored because these models needed data to train on? This chain of power usage goes on down pretty deep.

Let’s consider a carton of milk vs. a massive gallon of milk. Which is most environmentally friendly? Both produce more waste from the farm for production, packaging, transportation and consumption. Small cartons of milk will likely be way more resource-intensive than large gallon.

I believe the following questions were addressed indirectly in the above, but I’ll take it back from domain-specific LLMs vs. “SLM.” My answer is that it depends on where the SLM is to be considered to still be an SLM without being a domain-specific LLM, hence why calling it a specialised model makes everything easier.

SLM application layers

I think the question regarding which applications SLMs are best used for, well, let’s use yet a few other examples: a self-driving car, autocompletion for a mobile phone keyboard/email, video recommendations based on what you watch, all of these so-called algorithms is and has already been around for years and used as specialised models for these particular tasks, they might not be “SLM” but very much specialised. So yes, chatbots, but we’ve been interacting with many more and for a good decade at least.

Finance and retail for SLMs have been key to growth, but the entire industry should consider this valuable, too. My previous response above proves that your favourite social media app or video streaming platforms are just a few more examples of how they can efficiently push products/ads for you to consume even better. But all in all, it is an effective tool to boost creativity, efficiency and the whole industry, no matter the field and the combination of both has been doing wonders for years already and will keep doing so!”