Natural language processing(NLP) is among the most exciting subsets of machine learning.
It lets us talk to computers like theyre people and vice versa.
In todays AI landscape, smaller, targeted models trained on essential data are often better for business endeavors.
However, there are massive NLP systems capable of incredible feats of communication.
OpenAIsGPT-3, one of the most popular LLMs, is a mighty feat of engineering.
But its also prone to outputting text thats subjective, inaccurate, or nonsensical.
This makes these huge, popular models unfit for industries where accuracy is important.
This indicates a wide-open market for new startups to form alongside established industry actors, such asDataikuandArria NLG.
40% off TNW Conference!
I spoke with Emmanuel Walckenaer, the CEO of one such company,Yseop.
According to him, when it comes to building AI for these domains, theres no margin for error.
It has to be perfect, he told TNW.
While these systems are impressive, as mentioned above, they are usuallycompletely untrustworthy.
Theyre brittle, unreliable, and prone to making things up.
In laypersons terms: theyre dumb liars.
This is because of the way they are trained.
LLMs are amazing marriages of mathematics and linguistics.
But, at their most basic, theyre beholden to the data theyre trained on.
As the old saying goes, you get out what you put in.
It might give you a great recipe, it might output a random diatribe on NBA superstar Stephen Curry.
However, theres no room for that kind of uncertainty in medical, financial, or business intelligence reports.
Next, the company has to ensure its machine learning systems are free of bias, omission, andhallucination.
To overcome the problem of hallucination, Yseop relies on having humans in the loop at every stage.
The companys algorithms and neural networks are co-developed by math wizards, linguistics experts, and AI developers.
Their databases consist of data sourced directly from the researchers and businesses being served by the product.
The next problem devs need to overcome with language processing is omission.
This happens when an AI model skips over pertinent or essential parts of its database when it outputs information.
The last major hurdle that huge LLMs usually fail to pass is bias.
One of the most common forms of bias istechnical.
A prime example of technical bias would be teaching a machine to predict a persons sexuality.
But theyre entirely unsuited for standardized industries where accuracy and accountability are paramount.
Why use AI at all?
And who wants to trust a machine that hallucinates with their financial future?
Luckily for all of us, companies such as Yseop dont use open-ended datasets full of unchecked information.
But it still begs the question, why use AI at all?
Weve gotten by this far with non-automated software solutions.
Walckenaer told me there may soon be no other choice.
According to him, the human workforce cant keep up at least in the pharmaceutical industry.
And theres even good news for those who fear being displaced by machines.
He assured us that Yseops systems were meant to augment skilled human labor, not replace people.
