“This is definitely an interesting topic,” says Alex Tamkin, a researcher at AI startup Anthropic. He added that this paper cleverly breaks down multi-step tasks and measures the contributions of individual components. “But that’s not all. Not all breakthroughs are illusory. This paper shows that even if we use one-step predictions or continuous indicators, My understanding is that the discontinuities were indeed found, demonstrating that it is possible to dramatically improve performance by increasing the size of the model.”
And even if we can explain the lack of emergence in current LLMs by using various measurement tools, this does not mean that future larger, more complex LLMs cannot also be emergent. “As we grow LLMs to the next level, they will inevitably absorb knowledge from other tasks and other models,” said Xia (Ben) Hu, a computer science researcher at Rice University. points out.
The issues that are changing the way we think about emergence are not abstract issues that only researchers should pay attention to. For Tamkin, this question is directly connected to efforts to predict how LLMs will function in the future. “These technologies are very broad and have unlimited applications, and I hope that the industry continues to emphasize the importance of continuing to conduct these predictive studies by viewing this discussion as a starting point.” I hope so. What should we do to avoid being surprised by the next generation model?
*This article is an original story reprinted with permission from “Quanta Magazine” operated by the Simons Foundation (edited independently from the foundation). The Foundation’s mission is to increase public understanding of science by highlighting research developments and trends in mathematics and the physical and life sciences.
(Originally published on QUANTA MAGAZINE, translated by Kei Hasegawa/LIBER, edited by Michiaki Matsushima)
*Click here for related articles on large-scale language models by “WIRED.”
Related Articles
AI models can be built without infringing copyrights – the first legally and ethically “correct” large-scale language model certified
OpenAI claims that it is “impossible” to create good AI models without using copyrighted content. However, the emergence of non-piracy-free large-scale language models (LLMs) and the increasing use of public domain datasets suggest the opposite.
Disrupting Google Search — Compatriot Sundar Pichai, whom the founder of AI search “Perplexity” sees as a rival
Goodbye chatbot. Welcome to the era of AI agents
AI agents have appeared that can not only write code, but also plan problem-solving, verify code, and implement it. The focus of development for startups and major companies is shifting from “chatbots that only answer questions” to “AI agents that can actually perform tasks.”
Magazine “WIRED” Japanese version VOL.52
“FASHION FUTURE AH!”
Fashion means clothes. It’s about knowing what cloth is made from, and thinking about the reality that found plastic bottles can be turned into thread, and the difficulties of turning old clothes into fiber. It is about seriously discussing how the next generation should be nurtured and what words we should say to them. It is about inheriting craftsmanship, gaining insight into fashion and optimism, and learning about the owners of extraordinary creativity. It means listening to the words. Of course, it also means imagining the possibilities of technology expansion, and it also means stepping on a sewing machine yourself. WIRED’s first fashion feature in about 10 years. Click here for details.
Tags: AIs emergentity illusion measurements largescale language models reveal WIRED .jp
