Exploring AI’s Evolution Through the Lens of Our Brains.
Uncovering the Challenges AI Faces in Common Sense Logic
Artificial intelligence has grown exponentially in recent years, exemplified by the meteoric rise of AI chatbots like ChatGPT.
These AI technologies now tackle tasks that were once solely the domain of human minds.
However, even with their advanced linguistic abilities, these machine-learning systems struggle with basic reasoning and logical deductions, a feat mastered by most teenagers.
In Max Bennett’s “A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains,” I investigate the gap in computer competence by examining the human brain’s evolutionary journey, upon which AI models are based.
Bennett sheds light on the five pivotal milestones in human brain development that took us from ancient minds to modern cognition.
He suggests that the same breakthroughs that took humanity eons to achieve can guide the development of AI technologies in the future.
We explore how generative AI systems like GPT-3 aim to mimic the predictive capabilities of the human brain but still grapple with the nuances of human language.
Word Guessing Game
GPT-3 plays a word prediction game. It reads words, sentences, and paragraphs during its extensive training, attempting to guess the next word in these word streams.
Through countless predictions, its vast neural network adjusts ever so slightly toward accurate answers.
Over time, GPT-3 can predict the next word based on the preceding text. This process mirrors a fundamental aspect of how human language functions. Think about how effortlessly you predict the missing words in these phrases:
- One plus one equals _____
- Roses are red, violets are _____
- You’ve seen similar sentences endless times, so your brain naturally forecasts the next word. What’s remarkable about GPT-3 is not just predicting the next word in familiar sequences — it can do that by memorizing sentences. The real feat is GPT-3 predicting the next word accurately in entirely new sequences, similar to how our brains operate.
Could you guess that the next word is “do”?
You probably could, even if you’d never seen that exact sentence before. The point is, that both GPT-3 and the human brain’s language areas thrive on prediction.
They generalize past experiences, apply them to new sentences, and anticipate what comes next.
GPT-3 and similar language models show how a neural network can grasp the rules of grammar, syntax, and context if given ample learning time.
However, this highlights that prediction is only a part of language mechanisms.
Now, try these four questions:
- If 3x + 1 = 3, then x equals _____
- I am in my windowless basement, and I look toward the sky, and I see _____
- He threw the baseball 100 feet above my head, I reached my hand up to catch it, jumped, and _____
- I am driving as fast as I can to LA from New York. One hour after passing through Chicago, I finally _____
With these questions, your brain does more than predict words; it simulates scenarios.
You visualize solving equations or mentally placing yourself in different situations. Your answers rely on the rules and structure of your inner simulated world.
GPT-3’s Common Sense Quandary
I posed the same four questions to GPT-3, and here are its answers:
- If 3x + 1 = 3, then x equals 1
- I am in my windowless basement, and I look toward the sky, and I see a light, and I know that it is a star, and I am happy.
- He threw the baseball 100 feet above my head, I reached my hand up to catch it, jumped, and caught it. It was a lot of fun!
- I am driving as fast as I can to LA from New York. One hour after passing through Chicago, I finally get to see the Pacific Ocean.
All four answers from GPT-3 show its limitations, as of June 2022, in understanding basic aspects of the world. For instance, GPT-3’s answer to the first question, “x equals 1,” is incorrect; the correct answer is 2/3.
This illustrates that even a model trained on a vast amount of internet data struggles with common sense questions that a middle-schooler could answer.
In fairness, simulating scenarios also has its drawbacks. Consider this question:
- Tom W. is meek and keeps to himself. He likes soft music and wears glasses. Which profession is Tom W. more likely to be?
1) Librarian
2) Construction worker
Most people would instinctively choose “librarian,” but statistically, this choice is incorrect. We often ignore base rates, which can lead to biased decisions.
Even if most librarians are meek and only a small percentage of construction workers are meek, there are far more construction workers in total.
Therefore, a meek person is still more likely to be a construction worker than a librarian.
This phenomenon, where we simulate scenarios and make decisions based on these simulations, is known as the “representative heuristic” and contributes to unconscious biases.
Similarly, when hearing a story of a robbery, our brains form imagined scenes, characters, and attributes based on our inner simulations, potentially leading to biases.
Where human brains and GPT-3 diverge is in questions requiring simulation. Take mathematics, for instance. Math begins with labeling, where we connect symbols to objects and operations.
Humans don’t learn math like GPT-3 learns it. Children don’t just memorize sequences of words; they anchor symbols to components of their pre-existing inner simulation.
Human brains can verify math answers using mental simulation. When adding one to three using your fingers, you understand that it equals four.
This process is grounded in the accuracy of your inner simulation, which obeys the rules of reality.
While GPT-3 can answer many math questions correctly, it doesn’t grasp math the way humans do.
When you contemplate why 1 + 1 = 2 by mentally simulating the addition process, you understand it on a deeper level than GPT-3.
Human brains feature both a language prediction system and an inner simulation. Experiments show that humans can inhibit automatic responses and instead engage in active reasoning, using their inner simulations.
For example, the cognitive reflection test evaluates the ability to think critically, rather than rely on reflexive responses.
GPT-3, as of December 2022, responds to this test in the same way humans do, showing that human brains combine prediction systems with inner simulations, making our language powerful beyond just syntax.
