MIT study: generative AI fails to develop a coherent understanding of the world

Today's big models have a wide range of remarkable abilities, from reciting poetry to generating efficient computer code, giving the illusion that they seem to "vaguely understand" some of the fundamental laws of the real world. However, a new study suggests that this may not be the case. The researchers found that a popularGenerative AI The model was able to be in New York CityProvides near-perfect step-by-step navigation guidance, but did not result in a truly accurate map.

MIT study: generative AI fails to develop a coherent understanding of the world

According to theMassachusetts Institute of TechnologyTechnology Review reported on November 5, local time that Ashesh Rambachan, principal investigator at MIT's Laboratory for Information and Decision Systems (LIDS), said, "We're hopeful that the big models' excellence in language might allow them to be used in theOther areas of scienceThe big picture. However, if one wants to use these technologies to explore new discoveries, it is crucial to judge whether they form a coherent worldview."

Researchers have found that a popular generative AI model is able to deliver New York City'sNear-perfect step-by-step navigation guidewhileDoesn't really form an accurate map of the city.

Although the model demonstrated excellent navigational capabilities, when the researchersCertain streets were closed and detours were put in placeWhen the model behaves butslump.

Further analysis shows that the model implicitly generates a map of New YorkContains a large number of non-existent streets, these streets twist and connect between grids across widely separated intersections.

This could have important implications for generative AI models in real-world applications - a model that performs well in a given context may not be able to cope when the environment or task changes slightly.

The researchers focused on a type of generative AI model known as a "transformer," which forms the core of LLMs such as GPT-4. transformers are trained with large amounts of language-based data to predict the next token in a sequence, such as the the next word in a sentence.

The researchers demonstrated the impact of this result by adding detour paths to a map of New York City, which caused all navigation models to fail to function properly. "What surprised me was how quickly the model's performance deteriorated once we added detours. If only 1% of streets are turned off, the accuracy drops immediately from nearly 100% to only 67%."

And when researchers restore the model-generated city maps, they find that they look like a "Fictional New York": Hundreds of streets intersect and overlap on top of the grid. The map often features random bridges across streets, or streets that cross at incredible angles.

The relevant papers are attached below:

statement:The content is collected from various media platforms such as public websites. If the included content infringes on your rights, please contact us by email and we will deal with it as soon as possible.
Information

Optimizing Short Play Generation Capabilities, KunlunWanwei SkyReels AI Short Play Platform Launches in the US on December 10th

2024-11-11 2:23:39

Information

AI Impacts Construction: U.S. Private Data Center Construction Spending Surges to Nearly $30 Billion/Year, More Than Twice as Much as End of 2022

2024-11-11 9:40:08

Search