Aside from the northward advance of killer bees in the 1980s, Artificial Intelligence few things have instilled as much dread in headline writers as the rise of artificial intelligence. The concern began to take root after Deep Blue defeated chess champion Garry Kasparov in 1997, challenging the notion of human supremacy over machines. Initially, AI faced limitations across various domains, from medical diagnosis to speech transcription.
Artificial Intelligence
Around a decade ago, a transformative shift occurred with the emergence of deep learning, bolstering computer brains, or neural networks, to rival human capabilities in image recognition, sign reading, photo enhancement, and speech-to-text conversion. However, these achievements had constraints, with deep learning excelling in specific tasks but struggling to adapt its expertise to diverse areas.
The landscape changed further with the advent of generative AI in the 2020s, marking a departure from the deep learning revolution. These systems, powered by large language models (LLMs) like ChatGPT, began to replicate human creativity. They could answer questions eloquently, compose poems, articles, legal briefs, create high-quality artwork, and even generate custom videos.
LLMs owe their prowess to extensive training on vast datasets, including digitized content from the internet and countless printed books. By deciphering patterns in linguistic elements, they predict word arrangements, often likened to “autocorrect on steroids.” Despite skepticism, LLMs demonstrated remarkable abilities such as emulating authors’ styles, solving riddles, and discerning contextual meanings.
This development, however, stirred concerns within the tech community. Some experts warned that unchecked LLMs could lead to widespread unemployment, societal upheaval, and the displacement of professions like magazine columnists. On the contrary, skeptics argued that these fears might be exaggerated, at least for the present.
The crux of the debate revolves around whether LLMs genuinely comprehend their actions or merely simulate understanding. While some propose that LLMs can reason and potentially achieve a form of consciousness, others, including computer scientist Melanie Mitchell, assert that current LLMs lack real-world understanding akin to human cognition.
Mitchell and co-author Martha Lewis recently published a paper demonstrating that LLMs struggle to adapt skills to new scenarios, a fundamental aspect of human understanding. The study showcased how LLMs, when faced with variations in alphabets or symbols, failed to generalize concepts, unlike humans who exhibited high performance in both original and counterfactual scenarios.
Mitchell emphasized that true understanding involves reliability and correct decision-making in novel situations, based on abstract concepts. Human cognition relies on these mental models to infer cause and effect, predict outcomes, and apply knowledge to unforeseen circumstances.
While Mitchell doesn’t rule out the possibility of AI reaching human-like understanding in the future, she questions whether LLMs, which learn language before abstracting concepts, are on the right path. The contrasting process in human development, where concepts precede language acquisition, raises doubts about the efficacy of reading the internet as the optimal strategy for artificial or human intelligence.
