Are tech companies on the verge of creating thinking machines with their tremendous AI models, as top executives claim they are? Not according to one expert.
We humans tend to associate language with intelligence. We tend to be compelled by those with greater linguistic skills as orators or writers.
But the latest research suggests that language isnโt the same as intelligence, says Benjamin Riley, founder of the venture Cognitive Resonance, in a
essay for The Verge. And thatโs bad news for the AI industry, which is predicating its hopes and dreams of creating an all-knowing artificial general intelligence, or AGI, on the large language model architecture itโs already using.
โThe problem is that according to current neuroscience, human thinking is largely independent of human language โ and we have little reason to believe ever more sophisticated modeling of language will create a form of intelligence that meets or surpasses our own,โ Riley wrote. โWe use language to think, but that does not make language the same as thought. Understanding this distinction is the key to separating scientific fact from the speculative science fiction of AI-exuberant CEOs.โ
AGI, to elaborate, would be an all-knowing AI system that equals or exceeds human cognition in a wide variety of tasks. But in practice, itโs often envisioned as helping solve all the biggest problems humankind canโt, from cancer to climate change. And by saying theyโre creating one, AI leaders can justify the industryโs
exorbitant spending and
catastrophic environmental impact.
Part of the reason why AI capex has been so out of control is the obsession with scaling: by furnishing the AI models with more data and powering them with ever growing-numbers of GPUs, AI companies have made their models better problem solvers and more humanlike in their ability to hold a conversation.
But โLLMs are simply tools that emulate the communicative function of language, not the separate and distinct cognitive process of thinking and reasoning, no matter how many data centers we build,โ Riley wrote.
If language were essential to thinking, then taking it away should take away our ability to think. But this doesnโt happen, Riley points out, citing decades of research
Original Postsychology/linguistics/2024-fedorenko.pdf" shape="rect">summarized in a commentary published in
Nature last year.
For one, functional magnetic resonance imaging (fMRI) of human brains has shown that distinct parts of the brain are activated during different cognitive activities, Riley notes. Weโre not recruiting the same region of neurons when pondering a math problem versus a language one. Meanwhile, studies of people who lost their language abilities showed that their ability to think was largely unimpaired, since they could still solve math problems, follow nonverbal instructions, and understand other peoplesโ emotions.
Even some leading AI figures are skeptical of LLMs. Most famous of all is the Turing Award winner and โgodfatherโ of modern AI Yann LeCun, who until recently was Metaโs top AI scientist. LeCun has long argued that LLMs will never reach general intelligence, and instead believes in pursuing so-called โworldโ models that are designed to understand the three dimensional world by training them on a variety of physical data, rather than just language. Itโs likely that this view led to his recent departure; despite LeCunโs position, Meta CEO Mark Zuckerberg has pivoted to
pouring billions of dollars into a new AI division for creating an artificial โsuperintelligenceโ using LLM technology.
Other research adds to the idea that LLMs have a hard ceiling. In a
new analysis published in the
Journal of Creative Behavior, a researcher used a mathematical formula for determining the limits of AI โcreativity,โ with damning results. Because LLMs are a probabilistic system, they reach a point where they are no longer capable of generating novel and unique outputs that arenโt nonsensical. As a result, the study concluded that even the best AI systems will never be anything more than serviceable artists that write you a nice wordy email.
โWhile AI can mimic creative behavior โ quite convincingly at times โ its actual creative capacity is capped at the level of an average human and can never reach professional or expert standards under current design principles,โ study author David H Cropley, a professor of engineering innovation at the University of South Australia, said in a
statement about the work.
โA skilled writer, artist or designer can occasionally produce something truly original and effective,โ Cropley added. โAn LLM never will. It will always produce something average, and if industries rely too heavily on it, they will end up with formulaic, repetitive work.โ
That isnโt a promising portent if LLM-powered AI is supposed think up new innovations and push the envelope of our understanding of the world. How will it invent โnew physics,โ as Elon Musk says it will, or solve the climate crisis, as OpenAI CEO Sam Altman
has suggested, if the tech struggles to string together new sentences that arenโt based on preexisting writing?
โYes, an AI system might remix and recycle our knowledge in interesting ways,โ Riley writes. โBut thatโs all it will be able to do. It will be forever trapped in the vocabulary weโve encoded in our data and trained it upon โ a dead-metaphor machine.โ
