Examples of Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Big Bangs and Big Bangs


The original version about This article appeared inside Quanta Magazine.

Two years ago, in a project called Beyond the Imitation Game benchmarkor BIG-bench, 450 researchers compiled a list of 204 tasks designed to test the power of Major languages, which offers chatbots like ChatGPT. For most applications, performance improves exponentially and steadily as models go up – the larger the sample size, the better. But with some jobs, the jump in skills was not smooth. The activity remained close to zero for a while, then the activity jumped. Other studies have found similar jumps in extinction.

The authors described this as “success” behavior; Some researchers have compared it to a phase change in science, such as when liquid water freezes into ice. In paper published in August 2022, researchers noted that these behaviors are not surprising but not predictable, and that they should inform what is happening around AI. security, probability, and risk. They called it art”discoveries“a term that describes collective behavior that appears only when a system reaches critical mass.

But things will not be easy. New paper and three researchers at Stanford University say that the sudden appearance of this skill is just the result of how researchers measure LLM work. They argue that this skill cannot be known or spontaneous. “The change is more obvious than people say,” he said Sanmi Koyeo, a computer scientist at Stanford and lead author of the paper. “The power of discovery claims has as much to do with how we choose to measure as it does with what the samples are doing.”

We are only seeing and reading this because of the growth of these species. Major languages ​​are taught by major analysis data sets-words from the Internet including books, web searches, and Wikipedia-and finding links between words that often appear together. Its size is measured in terms of parameters, approximately equal to all the ways in which words can be connected. The more, the more links LLM gets. GPT-2 had 1.5 billion units, while GPT-3.5, the LLM that powers ChatGPT, uses 350 billion. GPT-4, which started in March 2023 and is now down Microsoft Copilotthey say they use 1.75 trillion.

This rapid growth has led to an incredible increase in functionality and technology, and no one can deny that large enough LLMs can complete tasks that smaller models cannot, including those that are not trained. The three at Stanford who seemed like a “miracle” realize that LLMs become more useful the older they get; Then, additional problems of larger types should make it better for complex and varied problems. But they argue that whether the change appears smooth and predictable or the results are skewed and sharp from the choice of metric – or even the scarcity of test samples – rather than the use within the sample.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *