ARTIFICIAL INTELLIGENCE
Media coverage often portrays AI as more intelligent than it is.

AI systems seem so intelligent because they give more exposure to achievements that reflect it. Reality tells us otherwise.
Every time there’s a notable breakthrough in AI, we only hear how intelligent and skilled the systems are getting. In 2012 Hinton’s team got 63% top-1 accuracy on the ImageNet challenge. A few years later, a system topped human performance by achieving a striking +90% top-1 accuracy. The news: "AI can recognize objects better than humans." Well, no. __ When they tested this exact model on a real-world object dataset its performance dropped 40–45%.
Last year people went nuts over GPT-3. The New York Times, The Guardian, Wired, TechCrunch, DigitalTrends, and many other prestigious news magazines spread the word about its astonishing capabilities. The hype went so over the top that Sam Altman, OpenAI’s CEO, had to reduce the tone:
"[GPT-3 is] impressive […] but it still has serious weaknesses and sometimes makes very silly mistakes. AI is going to change the world, but GPT-3 is just a very early glimpse."
In contrast, almost no one talks about all the tasks at which AI is still very dumb. AI not being very intelligent doesn’t make for a catchy headline. Despite all the successes of deep learning systems, there are still crucial ways in which AI is the opposite of "superhuman." Gary Marcus, a professor of psychology at New York University and leading AI expert, mentions some examples in his book Rebooting AI. In this article, I’ll explain in which ways we still vastly outperform AI – and will still do in the foreseeable Future.
Understanding language – Pragmatics
Generative pre-trained language models are the latest trend in AI. In 2017, Google invented the transformer architecture which has become the go-to for most NLP systems. The heyday came last year when GPT-3 – a transformer-based model developed by OpenAI – revealed unmatched ability in learning and generating language. From the incredible results, it’s fair to conclude that GPT-3 mastered language. However, it only excels at syntax and semantics: the form and structure of language. It lacks understanding. It can’t link form with the underlying meaning, and it can’t access the pragmatic dimension of language.
I recently wrote an article for Towards Data Science claiming that "AI won’t master human language anytime soon." Powerful state-of-the-art models like GPT-3 can generate human-level language, but they don’t know why they’re outputting a given sentence. Although GPT-3 could say: "I ate an apple this morning," it doesn’t mean it knows what it feels like to smell, touch, or eat an apple. It lacks the subjective experience of actually doing it.
Language serves the purpose of linking reality with mental representations of that reality. AI can’t access our shared physical reality because it’s trapped in a virtual world. If instead of saying: "I ate an apple this morning," I say: "I ate an apple this morning. I went to the store, took it, ate it, and left," only humans could infer that I stole it – and the social/legal implications of doing it. Today’s AI can’t access this type of pragmatic information.
Understanding the world – The virtual trap
One of the most powerful arguments against machine learning being the right path towards AGI is that intelligence needs to be embodied. Connectionist AI – of which machine learning is the foremost paradigm – is loosely based on Cartesian duality: Mind and body are distinct, separate substances. Thus, intelligence can arise without a body. Philosopher Hubert Dreyfus was one of the first to criticize this notion. Today, there’s a whole sub-field called developmental robotics based on the opposite idea: Intelligence arises from the interaction of mind and body with the world.
Apart from embodied cognition, other arguments put connectionist AI against the ropes: Judea Pearl argues that machine learning and deep learning are techniques that can only exploit correlations in data, whereas we learn by identifying the causes and effects of events. Joshua Tenenbaum wants to imbue AI with intuitive physical notions that we have developed through evolutionary processes – and machines lack. Gary Marcus defends the idea of combining connectionist AI with older paradigms based on symbolic knowledge to instill common sense in machines.
Nowadays, AI isn’t embodied, doesn’t understand causality, physics, or language, and doesn’t show common-sense reasoning. Yet, we use these abilities to navigate the world every day. So long as machines live and develop within a computer, enclosed within the limits of a virtual reality, they won’t be able to understand the world. In the words of professor Ragnar Fjelland: "As long as computers do not grow up, belong to a culture, and act in the world, they will never acquire human-like intelligence."
Adapting to new circumstances – Extrapolation
Andre Ye wrote a fantastic article diving into the theoretical foundations of neural networks. In his article – which Steven Pinker and Gary Marcus praised on Twitter – he explains the difference between generalization (interpolation) and extrapolation. He argues that neural networks are nothing more than "great approximators." The goal of a neural network, he says, is to generalize within the range of training, but not outside of it. A neural network is, in essence, a highly-skilled multidimensional interpolator.
But humans can extrapolate. We can make "reasonable predictions outside a given trained range (emphasis in original)." Extrapolation consists of translating conclusions from a problem to another of different – usually more complex – nature. The inability of machine learning systems to carry out this process is largely due to the "independent and identically distributed (iid)" assumption. It states that real-world data and training data have the same distribution, which is false. The three leading pioneers of the deep learning revolution, Geoff Hinton, Yoshua Bengio, and Yann LeCun, explained in a recent paper why this is a critical flaw:
"Humans can generalize in a way that is different and more powerful than ordinary iid generalization: we can correctly interpret novel combinations of existing concepts, even if those combinations are extremely unlikely under our training distribution."
Nothing assures that machine learning systems will ever be able to extrapolate as we do. That’s also why Elon Musk, who has repeatedly promised that fully self-driving cars would be ready early in 2021, had to take back his words in a Tweet earlier this week:
"Generalized self-driving is a hard problem, as it requires solving a large part of real-world AI. Didn’t expect it to be so hard, but the difficulty is obvious in retrospect. Nothing has more degrees of freedom than reality."
Learning new things quickly – Evolutionary fitness
Machine learning systems – deep learning ones especially – need huge amounts of data (labeled or unlabeled) to learn and tons of computational power to train. The three main training frameworks – supervised, reinforcement, and self-supervised learning – are subject to these limitations. AlphaZero played 44 million games against itself before beating Stockfish. GPT-3 was trained with most of the internet’s text data. Expert chess players and world-class writers need nowhere near that amount of practice.
In general, human children learn rather quickly in comparison. At 1 month old, prelinguistic infants can distinguish phonemes. Before 1 year of age, they can infer cause-effect relationships. And kids learn to speak fluently at 5 y.o from sparse data. Adults are no different. We can learn the basics of driving with a few hours of practice. Self-driving cars, as we saw earlier, are still way below our level after years of simulations.
As Hinton, Bengio, and LeCun note in their paper, "[h]umans […] seem to be able to learn massive amounts of background knowledge about the world, largely by observation, in a task-independent manner." We’re simply unmatched in our ability to learn fast from experience and observation. A key piece of this puzzle is evolution. We’ve developed fine tools to navigate the world, whereas AI is designed directly by us. AI has to compensate for the lack of evolutionary advantages with learning. We come to the world mostly prepared to deal with it. AIs have to learn from scratch.
Summarizing
AI is great at solving some problems. Machine learning systems are starting to reach excellence in perceptual tasks, also called system 1 tasks – borrowing the concept from psychologist Daniel Kahneman. But they can’t seem to dominate high-level cognitive abilities.
Humans are also great at solving some problems. There’s some overlap with AI, but we remain unchallenged at higher cognitive tasks. Reasoning, planning, decision-making, understanding causality, language, or intuitive physics…all these brain processes still lie outside the limits of AI. And it doesn’t look like it’ll change anytime soon.
Travel to the future with me for more content on AI, philosophy, and the cognitive sciences! Also, feel free to ask in the comments or reach out on LinkedIn or Twitter! 🙂
Recommended reading
Unpopular Opinion: We’ll Abandon Machine Learning as Main AI Paradigm