Five artificial intelligence failure modes shared by humans

Ian Leslie draws parallels between machine failure modes and human behavior, focusing on model collapse and overfitting. He argues that better data curation and deliberate novelty can counter both.

Ian Leslie argues that research in artificial intelligence and cognitive psychology increasingly informs each other, yielding useful concepts for both fields. Drawing on Dwarkesh Patel’s interview with Andrej Karpathy, a co-founder of OpenAI who has since departed, and the book Algorithms To Live By by Brian Christian and Tom Griffiths, he frames five shared failure modes between machines and people. He then develops two in detail: model collapse and overfitting.

Model collapse, he writes, arises when models trained on rich human data begin learning from model-generated data as the internet fills with synthetic text and images. Because synthetic outputs are more predictable and less diverse, each generation amplifies prior biases and errors while shedding nuance, creativity, and signal. The feedback loop produces a generic, repetitive monoculture. Leslie notes the human analogue: over time people overfit to their own internal models, become rigid in thought, and rely on the same small set of friends and information sources. He extends the diagnosis to culture, citing pop music that chases streaming algorithms, formulaic Hollywood scripts, and thin imitations in contemporary visual art, describing postmodernism as a kind of cultural model collapse. Mitigations include raising quality control on training data, filtering out artificial intelligence generated content, and privileging human, rare, and anomalous data while still removing clear nonsense such as QAnon-style conspiracy theories. He points to OpenAI hiring domain experts to create exclusive high-quality content. For individuals, he recommends actively curating an information diet, reading great books, seeking novelty, and finding knowledgeable contrarian voices outside familiar circles.

Overfitting, in machine learning, occurs when a model memorizes its training set rather than learning generalizable patterns. It excels on familiar examples but fails on new inputs. Engineers counter it by penalizing reliance on specific patterns or stopping training before the model becomes too tuned to a particular dataset. Leslie sees a parallel in everyday life: routines provide stability but can narrow perception and make unfamiliar situations hard to interpret, leading either to misplaced confidence outside one’s domain or to fear that shrinks one’s world. He suggests periodically breaking habits to discover better approaches, citing a study in which a Tube strike forced commuters to find more efficient routes. He highlights neuroscientist Erik Hoel’s theory that dreams function as an injection of noise that disrupts rigid neural patterns, remixing mundane memories into bizarre forms to preserve flexibility, like the Fool in King Lear keeping sense clear by inverting it.

58

Impact Score

TSMC’s Arizona fabs begin supplying advanced chips to AMD and Nvidia

TSMC’s Arizona fabs have begun producing advanced chips for AMD and Nvidia, marking a milestone for the United States’ push to reshore semiconductor manufacturing under the Chips Act. Nvidia will build Blackwell compute engines and its first United States Artificial Intelligence supercomputers there, while AMD validated its 5th gen Epyc CPUs at the site.

Rondo Energy activates 100 MWh thermal battery for enhanced oil recovery

Rondo Energy has switched on what it says is the world’s largest thermal battery, a 100 megawatt-hour system powered by off-grid solar to supply heat for enhanced oil recovery. The project showcases promising efficiency but raises questions about decarbonizing fossil fuel operations.

Aluminum as a zero-carbon fuel and what is next for energy storage

A Boston startup is scaling aluminum-powered heat and hydrogen while a record-setting thermal battery underscores the promise and pitfalls of electrified industrial heat. The newsletter also rounds up major developments in artificial intelligence, quantum computing, and tech policy.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.