Language isn’t intelligence, and scaling won’t fix it
Amid record spending and grand AGI promises, a growing chorus of experts argues that large language models mimic speech but don’t think. Neuroscience evidence and new creativity research suggest hard ceilings on what probabilistic, language-only systems can achieve. The strategic takeaway: betting the future of AI on LLMs alone risks costly dead ends, while alternative approaches that model the real world may prove more fruitful.
Points clés
- Benjamin Riley, founder of Cognitive Resonance, writes in The Verge that LLMs emulate language but not the distinct cognitive processes of thinking and reasoning.
- Riley cites neuroscience showing language and reasoning activate different brain networks, including fMRI evidence and Nature commentary summarizing decades of research.
- Studies of people with impaired language abilities show thinking remains largely intact, as they can solve math problems, follow nonverbal instructions, and read emotions.
- The AI industry’s scaling obsession—more data, more GPUs, more data centers—has improved conversational fluency but not genuine reasoning, Riley argues.
- Riley warns AGI hype helps justify massive capex and environmental costs while conflating language proficiency with intelligence.
- Yann LeCun, Turing Award winner and former top AI scientist at Meta, says LLMs won’t reach general intelligence and advocates “world models” trained on physical data.
- Despite that stance, Meta CEO Mark Zuckerberg is investing billions in an LLM-driven push toward artificial “superintelligence,” a shift tied to LeCun’s recent departure.
- A Journal of Creative Behavior analysis by David H. Cropley (University of South Australia) concludes LLMs’ probabilistic nature caps creativity at average-human levels.
- Cropley says LLM outputs will remain formulaic and non-novel under current designs, making them “serviceable” but unlikely to reach expert creative standards.
- Tech leaders like Elon Musk and Sam Altman tout LLMs for breakthroughs from “new physics” to climate solutions, but Riley calls them “dead-metaphor machines” confined to training data.
À retenir
For non-experts: treat LLMs as very talkative calculators—great at sentences, not so great at thinking. Don’t outsource your next moonshot to a system that tops out at “average” creativity; pair LLMs with real-world data, human oversight, and domain expertise. And before you buy more GPUs than a small nation, ask whether you need better models, better data, or—radical idea—both; your electric bill (and sanity) will thank you.
Sources
