Factuality — the golden goose for Large Language Models
Is it?
Large Language Models (LLM) are swear-words in some industries.
“Did you hear what happened when XYZ tried to integrate ChatGPT into their search engine?”, was the question across the board room, one evening.
It may appear as a naysayer’s attempt at maintaining the status quo when Large Language Models were acing school essays.
School essays. That’s a great example indeed. One can really sound credible and believable in a school essay when there are no quantifiable ways to measure precision or accuracy.
There are open, unmet demands in the market for LLMs though, where one can verify how accurate an answer is; and where precision is important.
Capital-intensive industries — where reversibility is expensive and time-consuming — are great examples. There is no room for fluffy answers when there is a choice to be made to improve the safety standard in oil rigs or road intersections.
Potentially, that is where the value is.
How can LLMs have more factuality? Can these models be created at reasonable price points without literally boiling the ocean?
What is factuality?
In my view, factuality improves an LLMs ability to be more accurate, less verbose, and more precise in its responses.
This may mean that the models are trained on deep, domain-specific, spatial, and/or supply-chain-specific datasets that are generic and scientific in nature.
Would there be a willingness to pay?
Absolutely. The value that LLMs provide for generic use cases (essays, copy, summaries) has been proven. The moment that value crosses over to precision land, where domain expert meets geek, there is much pain to be relieved and much value to be created.
Considering that domain experts’ time is expensive, there would be a high willingness to pay.
I do not have any data points on costing models of specialist LLMs, at the time of writing this.
