Is the future of AI and language models in specialization? Because anyone can build a language model. It just takes compute time and data. So is the future of software as a service, language models as a service?
what is SAS? SAS is software as a service. like
Salesforce or a CRM These are pieces of software that are given to you as a service, a monthly subscription to use this service. Language models are the foundational way that things like Claude and ChachiPT can communicate with you. And I think over the next 12 to 18 months, we're going to start to see LLMAS, large language models as a service. how are you going to specialize a language model? Fundamentally, a language model is built By collecting a very large amount of data, processing that into a mathematical formula that then predicts what the next best word. for the sentence it's writing would be. Now, if I build a language model based purely on a highly specialised piece of information,
then it's always going to be predicting the next best word based on that highly specialised information. it's worth doing because the costs of building, this general intelligence that understands. Everything is insanely huge.
if you have a specialised service, it can be smaller. And this is where you may have seen things like chat GPT is 180 billion parameters and Lama has 7 billion parameters.
It's really down to the volume of data that they have.
And one of the things that makes ChatGPT so good is it uses a cluster of experts. it has multiple specialised models within this big architecture of a big model, or so we think. So an example of a specialised LLM could be like a medical language model where it's only been trained on the world's knowledge of human medical information, and that's it. It doesn't know how to code, it doesn't know how to write poetry, but it has been trained on every single piece of medical literature that there is known to exist.
That would be a highly specialised LLM. Google is trialing a medical LLM at the moment. And the reports on the output are actually good.
So what, what are the strengths of this? it's highly, highly focused, So you can reduce that potential hallucination. It only knows this and that is the output. It's very efficient. it's a smaller model that doesn't need all of the compute training and power. it takes around 500, 000 PDFs to make a 1 billion parameter model. That's a lot of information. But as it's highly focused and highly specialised, as new information is created that can be put into this LLM and it's constantly training, only on its highly focused area of expertise. the cost of this is, you know, you've got to build it. So there's compute time and there's compute cost
with things like medical. you do need some level of multimodal ability, meaning it has to be able to look at an image, take information, and then. Process that and send it out it's going to be this payoff of like, it has to have this level of understanding and that level of understanding, but staying focused is hard. switching between specialised LLMs, do you have one login for that and one login for this and one login for that similar to a SAS or does it start to be controlled by a central processing unit? if your organization or you have all of the data needed to make an LLM, the opportunity you have to build a specialised LLM is huge. Because the data is where the value is, if you're in a market or sector where there's a lot of players who would have similar data, you could end up with this saturated market of which LLM specialised is the best the ethics of training that do you actually own all of that data?