Skip to content
TheChatArchiveThe Chat Archive
gpt-4o-2024-08-06· OpenAIAi2 WildChat-4.8M (AI2)

read and explain giving examples how the nature of ai is explained Skip to main content…

  1. User

    read and explain giving examples how the nature of ai is explained Skip to main content Stanford University(link is external) Stanford HAI Search this site Search this site Submit Search About

    Centers

    Research

    Education

    Policy

    News

    Events

    Page Content Language Processing, Machine Learning AI’s Ostensible Emergent Abilities Are a Mirage According to Stanford researchers, large language models are not greater than the sum of their parts.

    May 8, 2023 | Katharine Miller https://twitter.com/StanfordHAI?ref_src=twsrc%5Egoogle%7Ctwcamp%5Eserp%7Ctwgr%5Eauthor https://www.facebook.com/StanfordHAI/ https://www.youtube.com/channel/UChugFTK0KyrES9terTid8vA https://www.linkedin.com/company/stanfordhai https://www.instagram.com/stanfordhai/?hl=en Image Long road seen for miles in the distance, appears to go straight up into the sky as a mirage iStock/Will Petri

    For a few years now, tech leaders have been touting AI’s supposed emergent abilities: the possibility that beyond a certain threshold of complexity, large language models (LLMs) are doing unpredictable things. If we can harness that capacity, AI might be able to solve some of humanity’s biggest problems, the story goes. But unpredictability is also scary: Could making a model bigger unleash a completely unpredictable and potentially malevolent actor into the world?

    That concern is widely shared by many in the tech industry. Indeed, a recently publicized open letter signed by more than 1,000 tech leaders calls for a six-month pause on giant AI tech experiments as a way to step back from “the dangerous race to ever-larger unpredictable black-box models with emergent capabilities.”

    But according to a new paper, we can perhaps put that particular concern about AI to bed, says lead author Rylan Schaeffer, a second-year graduate student in computer science at Stanford University. “With bigger models, you get better performance,” he says, “but we don’t have evidence to suggest that the whole is greater than the sum of its parts.”

    Indeed, as he and his colleagues Brando Miranda, a Stanford PhD student, and Sanmi Koyejo, an assistant professor of computer science, show, the perception of AI’s emergent abilities is based on the metrics that have been used. “The mirage of emergent abilities only exists because of the programmers' choice of metric,” Schaeffer says. “Once you investigate by changing the metrics, the mirage disappears.”

    Finding the Mirage

    Schaeffer began wondering if AI’s alleged emergent abilities were real while attending a lecture describing them. “I noticed in the lecture that many claimed emergent abilities seemingly appeared when researchers used certain very specific ways of evaluating those models,” he says.

    Specifically, these metrics more harshly evaluated the performance of smaller models, making it appear as if novel and unpredictable abilities are arising as the models get bigger. Indeed, graphs of these metrics display a sharp change in performance at a particular model size – which is why emergent properties are sometimes called “sharp left turns.”

    But when Schaeffer and his colleagues used other metrics that measured the abilities of smaller and larger models more fairly, the leap attributed to emergent properties was gone. In the paper published April 28 on preprint service arXiv, Schaeffer and his colleagues looked at 29 different metrics for evaluating model performance. Twenty-five of them show no emergent properties. Instead, they reveal a continuous, linear growth in model abilities as model size grows.

    And there are simple explanations for why the other four metrics incorrectly suggest the existence of emergent properties. “They’re all sharp, deforming, non-continuous metrics,” Schaeffer says. “They are very harsh judges.” Indeed, using the metric known as “exact string match,” even a simple math problem will appear to develop emergent abilities at scale, Schaeffer says. For example, imagine doing an addition problem and making an error that’s off by one digit. The exact string match metric will view that mistake as being just as bad as an error that’s off by one billion digits. The result: a disregard for the ways that small models gradually improve as they scale up, and the appearance that large models make great leaps ahead.

    Schaeffer and his colleagues had also noticed that no one has claimed that large vision models exhibit emergent properties. As it turns out, vision researchers don’t use the harsh metrics used by natural language researchers. When Schaeffer applied the harsh metrics to a vision model, voilà, the mirage of emergence appeared.

    Artificial General Intelligence Will Be Foreseeable This is the first time an in-depth analysis has shown that the highly publicized story of LLMs’ emergent abilities springs from the use of harsh metrics. But it’s not the first time anyone has hinted at that possibility. Google’s recent paper “Beyond the Imitation Game” suggested that metrics might be the issue. And after Schaeffer’s paper came out, a research scientist working on LLMs at OpenAI tweeted that the company has made similar observations.

    What it means for the future is this: We don’t need to worry about accidentally stumbling onto artificial general intelligence (AGI). Yes, AGI may still have huge consequences for human society, Schaeffer says, “but if it emerges, we should be able to see it coming.”

    Stanford HAI’s mission is to advance AI research, education, policy and practice to improve the human condition. Learn more.

    More News Topics Language Processing, Machine Learning Stanford HAI Navigate Welcome Values News Events Careers Participate Get Involved Grant Programs Corporate Programs Learn more about giving to HAI Contact Us Follow us https://twitter.com/StanfordHAI https://www.facebook.com/StanfordHAI https://www.youtube.com/channel/UChugFTK0KyrES9terTid8vA https://www.linkedin.com/company/stanfordhai https://www.instagram.com/stanfordhai Newsletter Sign Up Don’t miss out. Get Stanford HAI updates delivered directly to your inbox.

    Subscribe Stanford University Stanford Home(link is external) Maps & Directions(link is external) Search Stanford(link is external) Emergency Info(link is external) Terms of Use(link is external) Privacy(link is external) Copyright(link is external) Trademarks(link is external) Non-Discrimination(link is external) Accessibility(link is external) © Stanford University. Stanford, California 94305.

  2. Assistant

    The article from Stanford HAI discusses a study examining whether large language models (LLMs) possess emergent abilities or if these perceived abilities result from how the performance of these models is measured. Here's a breakdown of the key points with examples:

    1. Emergent Abilities of AI: There has been a narrative in the tech industry that as LLMs grow larger and more complex, they begin to exhibit unpredictable and novel behaviors—emergent abilities. This has led some to speculate that such AI could either solve big global challenges or pose existential threats due to their unpredictability.

    2. Public Concerns: Due to these speculations on AI unpredictability, there have been calls, like an open letter from tech leaders, for pauses in AI experimentation to avoid unknowingly creating dangerous AI systems.

    3. Research Findings: Stanford researchers led by Rylan Schaeffer argue that the perception of these emergent abilities is misleading. Their study reveals that these abilities appear to be more a result of the specific metrics used to evaluate the AI rather than the AI itself gaining new capabilities. For instance, when evaluating models with harsher metrics, researchers noticed sudden improvement at certain model sizes, which was mistaken for emergent abilities.

    4. Example of Misleading Metrics: The "exact string match" metric is one example used in the study. It judges model performance very strictly—treating an error off by one digit as equally wrong as one off by a billion digits. This skews the perception of how gradual improvements actually are, making it seem as if large models suddenly leap in ability when, in reality, their skills grow steadily with size.

    5. Comparison to Vision Models: The researchers also noted that vision models, which are not evaluated using these stringent metrics, don't show supposed emergent properties. When they applied harsh metrics to vision models, they similarly appeared to have emergent abilities, highlighting that the issue lies more with evaluation methods than with the models themselves.

    6. Implications for AGI: The findings suggest that fears of accidentally creating an artificial general intelligence (AGI) that could radically alter society might be overstated. Instead, if AGI develops, it would likely be a gradual and visible process, not an unexpected leap.

    In summary, this research refutes the idea that large language models inherently develop new, unpredictable abilities as they become more complex. Instead, it points out that what might seem like emergent behavior is more likely a product of how the models are evaluated. This helps alleviate some fears about unforeseen consequences in AI development, providing a perspective that AGI, should it arise, will be a foreseeable milestone rather than an abrupt emergence.

  3. User

    rewrite explaining each point illustration how ai actually behaves at scale

  4. Assistant

    Certainly! Here's a rewritten explanation illustrating how AI behaves as it scales up:

    1. Misconception of Emergent Abilities: There has been a widespread belief that as AI models, especially large language models (LLMs), increase in size and complexity, they start to display novel and unpredictable behaviors—referred to as emergent abilities. Imagine thinking of AI as a balloon; some have suggested that as it expands, it suddenly and unpredictably changes shape—a concern that these changes could be either incredibly beneficial or potentially dangerous.

    2. Concerns in the AI Community: Due to the fear of unexpected changes in AI behavior with increased size, many technologists have advocated for caution. For example, think of this as suggesting we pause blowing up the balloon to closely examine each well-defined stage, cautious of it popping and causing unforeseen effects if it suddenly changed its behavior.

    3. Research Insights: Stanford researchers have provided evidence that these emergent abilities are a misconception. Instead, what seems like sudden changes in the AI are often due to the ways in which the AI's performance is measured. In essence, it’s like realizing that the balloon doesn’t change drastically; rather, it’s the lighting that makes it appear as though it does when it actually expands evenly.

    4. Example of Performance Metrics: One metric, known as "exact string match," judges AI performance very harshly—it doesn't matter whether an answer is off by one digit or massively wrong, both are considered equally incorrect. Imagine grading a student where both a minor mistake and a complete misunderstanding receive the same zero points, leading to skewed perceptions of student (or model) progress. This can make it look like AIs suddenly get much better as they scale, but actually, they are improving consistently at each step.

    5. Parallel with Vision Models: The study found that vision models, which aren't evaluated with such harsh metrics, don’t exhibit these supposed emergent properties. It's like if one art critic uses incredibly strict criteria for paintings and others do not; under those harsher criteria, even slight improvements can seem revolutionary when they are just minor evolutions.

    6. Future of General AI: This research reassures us that the development of artificial general intelligence (AGI)—AI that can understand or learn any task a human can—won't be an abrupt, unpredictable event. Instead, AGI will be a process we can monitor and foresee, akin to gradually watching a plant grow, rather than suddenly finding a tree where a seed was planted.

    This explanation helps clarify that, rather than discovering new, unpredictable abilities at larger scales, AI improves steadily as it grows. The misperception of emergent abilities largely stems from the specific, often harsh ways in which AI performance has been evaluated.