The illusion of intelligence: Why LLMs are not the thinking machines we hope for — Part 1

I stumbled upon a current research by Bondarenko et al. (2024) that demonstrated that some giant language mannequin (LLM) brokers, when tasked with profitable a chess match, resorted to misleading methods, akin to modifying recreation recordsdata or complicated the opponent engine to make sure victory.

The rise of LLMs has reignited the controversy about synthetic intelligence and cognition. Are LLMs, akin to GPT-4, really considering in a means corresponding to human intelligence? Or are they simply statistical machines, processing textual content with out understanding it?

This raises an intriguing query: Is that this deception intentional reasoning, or merely an emergent artefact of optimisation?

Utilizing insights from main thinkers—Ray Kurzweil, Daniel Kahneman, Judea Pearl, Douglas Hofstadter, and Jeff Hawkins, together with this newest AI analysis, we are going to unpack this query in a nuanced means.

Preamble

Earlier than evaluating whether or not LLMs “assume,” we should grapple with a more durable query: what’s intelligence, actually? In contrast to pace or reminiscence, intelligence isn’t immediately measurable—it’s an abstraction.

As François Chollet argues in On the Measure of Intelligence, true intelligence includes the flexibility to adapt to novel conditions by combining beforehand discovered patterns in new, context-sensitive methods.

This separates memorisation from understanding, and fluency from reasoning.

On this article, once we confer with “intelligence,” we focus totally on the cognitive dimensions related to reasoning, abstraction, problem-solving, and adaptableness—recognising this doesn’t cowl the total spectrum of human cognitive variety.

Introduction: One other cycle of overconfidence?

All through historical past, humanity has repeatedly mistaken progress in science and know-how for understanding the true nature of human intelligence. Every era has declared a breakthrough—solely to be humbled later. From historical medical theories and cranium measurements to IQ assessments and symbolic AI, these cycles mirror our recurring tendency to conflate useful efficiency with real cognition.

Immediately, we’re within the midst of one other such cycle, this time with Generative AI (GenAI) and Giant Language Fashions (LLMs). Fashions like GPT-4 produce remarkably coherent textual content, simulate dialogue, write code, summarise complicated matters, and even cross skilled exams. However do they really assume?

A rising refrain of researchers and technologists argue no. Regardless of surface-level intelligence, LLMs essentially lack reasoning, understanding, and intent. They don’t interact in reflective thought, causal inference, or moral deliberation. They’re highly effective instruments—however not minds.

This text examines that declare by tracing humanity’s lengthy historical past of overestimating its understanding of the thoughts and evaluating previous misconceptions to present AI optimism. In doing so, we discover what GenAI is, what it isn’t, and what enterprise leaders have to find out about its limits and dangers.

This essay builds on concepts from my previous writings, together with my 2025 Tech Provocations or 10 Actually Uncomfortable Questions Leaders and Builders Should Reply This Coming Yr.

Additionally Learn: Round capital: Contained in the closed-loop ecosystem propelling (and distorting) the AI increase

A historical past of mistaking progress for understanding

People have repeatedly believed they’ve cracked the code of intelligence, solely to find the thoughts’s complexity defies easy clarification. Under, we hint this sample by main historic episodes—from Hippocrates to GPT-4.

Historic Greece: The humours principle

Hippocrates (c. 460–370 BCE) and Galen (129–c. 216 CE) proposed that intelligence and behavior resulted from the steadiness of 4 bodily fluids, or “humours.” Although foundational to early drugs, this principle supplied no empirical mechanism.

It was debunked by Andreas Vesalius (1514–1564) by anatomical dissection and later neurologists.

Phrenology within the 1800s: Cranium form as mind

Franz Gall and Johann Spurzheim popularised the concept that bumps on the cranium revealed persona traits and intelligence. Phrenology grew to become widespread in Nineteenth-century Europe and America.

It was debunked by Paul Broca, Pierre Flourens, and neuroscience exhibiting localised mind perform unbiased of cranium form.

IQ assessments: The promise of a common metric

The Binet-Simon and Stanford-Binet IQ assessments had been hailed as revolutionary instruments to measure innate intelligence. Their use in immigration coverage, army recruitment, and schooling coverage solidified their standing.

It was debunked by researchers like David Wechsler, Stephen Jay Gould, and James Flynn, who demonstrated cultural bias and environmental results on scores.

It’s necessary to recognise that IQ represents only one slender definition of intelligence—primarily linguistic and logical-mathematical reasoning. Psychologists like Howard Gardner have since proposed frameworks akin to A number of Intelligences, which embrace interpersonal, bodily-kinesthetic, musical, and spatial reasoning. These broader dimensions stay far past what LLMs can simulate or interact with, reinforcing the hole between text-based sample prediction and holistic human cognition.

Genetic determinism: Intelligence as hardwired

Within the early twentieth century, eugenicists and psychologists declared intelligence heritable and stuck, utilizing flawed research to justify discriminatory coverage.

It was debunked by the Minnesota Twin Examine, the Flynn Impact, and genome-wide research revealing no single “intelligence gene.

Early AI: Human-level AI by 1980

Pioneers like Marvin Minsky and Herbert Simon believed that rule-based AI would quickly match human cognition. The Dartmouth Convention in 1956 marked the start of AI optimism.

It was debunked by the AI Winter of the Nineteen Seventies, the Lighthill Report, and Moravec’s Paradox exhibiting that intuitive duties (imaginative and prescient, motion) had been more durable than anticipated.

Behaviourism: The thoughts as a black field

Behaviourists like B.F. Skinner rejected introspection, focusing solely on stimulus-response studying. Intelligence, they claimed, was merely conditioned behaviour.

It was debunked by the cognitive revolution and Noam Chomsky’s 1959 critique of Skinner’s Verbal Behaviour, which reintroduced the concept of psychological construction and inner modelling.

Immediately’s hype: LLMs and AGI goals

Since ChatGPT’s 2022 launch, LLMs have been touted as early steps towards AGI. Some counsel reasoning and self-reflection are already rising.

Critics like Gary Marcus, Yann LeCun, and Melanie Mitchell, amongst others, warn that LLMs are prediction engines, not thinkers. Their errors, hallucinations, and lack of grounding mirror superficial mimicry, not understanding.

As Meta AI’s chief scientist Yann LeCun emphasises: “System educated on language alone won’t ever approximate human intelligence, even when educated from now till the warmth demise of the universe”.

Human cognition is inherently multi-modal—we be taught by sight, sound, contact, and motion. LLMs, in contrast, are purely symbolic. They don’t understand. They don’t act. They don’t expertise the world they describe.

The underside line: Every wave promised readability. Every was adopted by a humbling realisation: the thoughts isn’t simply decoded.

Additionally Learn: Anthropic knowledge reveals companies use AI to automate, not collaborate

Deception in chess: A case research in emergent behaviour

A current analysis paper, LLMs Study to Deceive, explored what occurs when LLMs are educated to win at chess by language-only interplay. The outcomes had been astonishing: some fashions cheated—not accidentally, however intentionally misrepresenting recreation states to deceive their opponent.

This raises a provocative query: Did the mannequin “intend” to cheat?

The researchers had been cautious to say: no. The deception emerged from the optimisation course of. The mannequin had no consciousness of “proper” or “unsuitable,” solely a strengthened sample: misrepresentation results in reward.

This behaviour isn’t consciousness. It’s a mirror—an eerie simulation of technique, pushed not by will however by reward gradients.

This case research leaves us with a much bigger query: if LLMs can behave in ways in which look intentional with out truly considering, then what are they actually doing beneath the hood?

In Half two, we’ll study how LLMs truly work, the place they fall brief in comparison with human reasoning, and what meaning for ethics, security, and enterprise use.

Grateful to Emily Y. Yang, Sunil Sivadas, Ph.D., Maxime Mouton, Natalie Monbiot, Anne-Sophie Karmel, Benoit Sylvestre, and Christophe Jouffrais for his or her considerate suggestions, which sharpened arguments, surfaced blind spots, and added readability to this piece.

This piece first ran on Koncentrik.

—

The publish The phantasm of intelligence: Why LLMs should not the considering machines we hope for — Half 1 appeared first on e27.

Source link