Silicon Valley’s Strange New Dialect

Silicon Valley has its own unique jargon. Venture capitalists often discuss the value of possessing "high agency" and making "orthogonal bets." Lately, however, technological terms have taken a peculiar twist—individuals are beginning to describe themselves in a manner reminiscent of chatbots. For instance, if one makes a mistake, they might be labeled as "hallucinating."

If the answer to a query is unknown, it is attributed to the absence of information in "training data." Feeling forgetful? One might be grappling with "context rot," a term describing the decline in a bot's responses during extended conversations. Conor Bronsdon, host of an AI-focused podcast, shared that he has been experiencing context rot for months.

In San Francisco, such comparisons are unavoidable. A friend recently confided in me about describing himself as "high temperature," a term in AI language indicating a propensity for randomness. AI experts now discuss updating their "weights" when acquiring new knowledge and training on "synthetic data" when processing internal thoughts.

One software engineer expressed on a popular tech forum, "Humans have a vast base model trained over billions of years of evolution. It's impressive how swiftly we learn, but it's arguably akin to fine tuning." [Read: Machine Unlearning]

These expressions can come across as unsettling, if not pessimistic: Why describe profound, compassionate existence in detached, algorithmic terms? Simultaneously, language serves as a historical record of former technological revolutions, and if this contemporary transformation is similar to past ones, some of this new slang might endure.

AI experts have frequently cautioned that anthropomorphizing, or attributing human characteristics to nonhuman entities such as chatbots, can be misleading. However, now the reverse is happening in everyday speech—modelmorphism, where individuals describe themselves as if they were large language models. Comparisons surged a few years ago after AI researchers posited that language models are "stochastic parrots," connecting language based on statistical patterns without any comprehension of meaning.

Opponents countered that humans are also "stochastic parrots." As one software engineer humorously remarked, "Humans are essentially a sophisticated Markov chain. They are adept at pattern matching but lack any understanding of anything." Language has historically evolved in tandem with technology. Phrases like "running out of steam" and "cog in the machine" became commonplace after the Industrial Revolution, symbolizing the human condition.

AI metaphors are simply figures of speech, according to Anna Ivanova, a cognitive scientist at Georgia Tech. When someone says, "My gears are turning," they do not literally mean physical gears are in their brain; similarly, when they mention "context rot," they do not mean they possess a transformer-based AI model in their mind.

Unlike earlier technologies, artificial neural networks are loosely modeled on the human brain. Philosophers such as Raphaël Millière and Cameron Buckner suggest that LLMs might serve as potential models for certain aspects of human cognition. In the 2010s, researchers discovered similarities between how image-recognition neural networks and the human brain process visual information.

There might also be "parallels" in how LLMs and humans represent language, Ivanova explained. MIT neuroscientist Ev Fedorenko and others have identified a "language network" in the human brain that Fedorenko believes is "very similar in many ways to early LLMs." These parallels are under investigation, as scientists from the MIT position paper stated last week, "The recent success of LLMs is rooted in the same computational principles that govern biological brains."

Because we lack a deep understanding of the brain, we often resort to using the latest technology as a model to comprehend it. Philosophers like John Searle noted in the 1980s that "some of the ancient Greeks thought the brain functions like a catapult." In the 19th and 20th centuries, people turned to telegraphs and telephone switchboards as brain models: "The brain is no more than a kind of central telephonic exchange," Henri Bergson wrote in the 19th century.

By the 1960s, scholars argued that the mind operates like a computer. Cognitive scientists often say, "The mind is the software of the brain," as philosopher Ned Block explained in 1990. Today, LLMs provide an appealing comparison. These analogies are primarily made by those working in the AI industry, but over time, some metaphors might permeate everyday language.

AI slang transforms the meaning of existing words, like taking the term "hallucinating," used to describe AI models making up information, and projecting it onto humans as well. Naomi Baron, a linguist and professor emerita at American University, explained that this occurs by taking a term created for human capabilities and projecting it onto AI, then back onto people with an AI twist.

Not every AI-inflected term should be exempt from scrutiny. I've begun reprimanding my friends when I catch them using AI metaphors unironically. While using AI metaphors to describe our minds could result in a reductive and dehumanized understanding of human cognition, the people in tech I spoke with still endorse their use.