GPT-3: a maths genius doesn’t necessarily have a knack for languages
“GPT-3, aka the strength and weakness of brute force,” Belgian daily newspaper De Standaard writes. The article zooms in on a topic frequently discussed at Nalantis: GPT-3, the majestic Artificial Intelligence of OpenAI. Processing enormous quantities of data enables GPT-3 to generate texts with actual content entirely independently. AI could, for example, compose a poem, jot down a tweet or compile an essay. Perhaps it has even written this blog! Who knows?
But is GPT-3 really the AI of the future? To answer this question, we first have to examine how GPT-3 works. Just like many other AIs, GPT-3 is based on big data, deep learning and machine learning. This means that it processes data – millions of online pages in this case – to become smarter. That’s not all: GPT-3 is a veritable transformer model that processes data super fast and keeps on learning, without any need for human intervention. It is entirely autonomous, and therefore capable of what we call unsupervised learning.
© Adobe Stock
Cathy O’Neil - Weapons of Math Destruction
Human, I don’t understand
Sounds pro, doesn’t it? Well, it is. But as Cathy O’Neil (Weapons of Math Destruction) so aptly puts it: “Big data processes codify the past; they do not invent the future.” While GPT-3’s greatest strength lies in its ability to scan and process the gigantic internet, this is also its weakness.
Because it is particularly difficult to fit an ‘infinite object’ like language into a statistical or mathematical model. You can compare it to your classmates when you were in school: some were good at languages, but had no aptitude for maths. Conversely, anyone who excelled at maths generally wasn’t interested in improving their vocabulary or grammar. Models like GPT-3 are practically perfect at recognising and reproducing patterns (maths), but don’t yet have the ability to truly understand the meaning of words (linguistics).
It is therefore impossible to control a product that is almost entirely based on language, such as a virtual assistant or a chatbot, exclusively through mathematical functions. This is basically the difference between OpenAI’s GPT-3 and the linguistics technology of Nalantis: our approach is based on something that is of crucial importance: ‘understanding’. We focus on interaction between humans and machines, a concept that depends on the ability of machines to effectively understand us.
“Deep learning alone will never outperform natural language understanding (NLU). For all the progress that has been made in AI, NLU is the one hard problem that has remained fundamentally unsolved. Understanding language is the holy grail of AI.”
(No) Black Box AI
A typical facet of unsupervised deep learning models is what we refer to as the Black Box AI. This impenetrable black box refers to systems in which the input and activities are not visible to users or other interested parties. Not only does this make the model difficult to fathom for programmers, it can also take a long time for errors in the system to be detected.
Take AI discrimination, for example. It is no secret that when deep learning models scan the internet, they process a great deal of wrong, racist and sexist data. This is – hopefully – not the intention of the makers, but it does lead to errors and mistakes. Due to invisible errors in the algorithm, a recruitment AI could prefer men’s CVs to those of women when filling a technical job, for example.
At Nalantis, we consciously opt for ‘No Black Box AI’. Transparency is one of our core values and our open approach helps us avoid poor performance and deviating algorithms. Mind you: No Black Box AI is not easy, particularly now that machine learning is becoming increasingly complex. But we strongly believe that it is more sustainable for companies and governments to understand and control how their data is used.
Let’s go hybrid
Should we consider GPT-3 a threat? No, we should see it as an opportunity. Think of it like putting together two powerful engines to obtain a single, mega-powerful one. Right? We believe the solution lies in a hybrid technology where linguistics converses with mathematics. Compare it to a stellar group project at school, where you have a maths genius collaborating with someone with a knack for languages. Thanks to this combination, we can achieve a perfect level of understanding, correction and self-learning that can be applied in a variety of disciplines. Preferably with a transparent AI that has thrown the Black Box overboard.
So ... OpenAI, shall we get together to discuss this?
Written by Frank Aernout, CEO of Nalantis
Would you like to read more and stay updated on new developments?
Connect with Frank on LinkedIn.
