TAIPEI (Taiwan News) — The production of semiconductors and AI technology are intimately linked, and Taiwan’s world-leading position in the production of chips places it at the forefront of current developments in machine learning.
Startup Island TAIWAN Podcast invited Dylan Patel, chief analyst at industry research firm SemiAnalysis, to discuss some of the possible avenues for the future of semiconductor development. He would also air his thoughts about the way Taiwan's industry can expand the capabilities of AI through semiconductor technology.
Patel said that semiconductor hardware is particularly important for large language learning models.
ChatGPT has made waves around the world, and is a language model that uses billions of inputs. Patel said some developers are now focusing on language models with reduced inputs to focus on specific tasks.
He said that even given five years of additional technological progress, it will still be physically impossible for large language models like ChatGPT to run on a smartphone, but smaller language models could. Patel said he envisages a future where he can give a voice command for a meeting on his phone in ordinary spoken word, and an AI will be able to infer from context who he wants to meet, where, when, and how.
He said this concept can be expanded to include models specifically designed for laptops, PCs, discreet GPUs, and other computing systems.
Listen to the full conversation on Startup Island TAIWAN Podcast.
Patel said Taiwanese developers will also be thinking about how hardware that allows powerful chips running AI applications to function. There is a huge space for innovation in cooling things cost effectively and efficiently, he said.
Continuing on the subject of inputs, Patel said that AI developers are working on systems that can produce outputs using multiple pieces of information at the same time, in addition to simple text or image-based inputs. He said it comes down to how much data can effectively be input into an AI model at one time.
“What if you could stick in thousands and thousands of words? What if you could paste in three different research papers and then start asking it questions about (them)?” he said.
Turning to TSMC’s proposed German production facility, Patel said there is strong potential for the company to bring its strengths to the country in a way that is mutually beneficial to itself and local partners. He said that companies like Bosch have highly advanced processes in some areas, more so than TSMC, but TSMC still has the advantage when it comes to making chips.
“TSMC will absolutely still continue to do what they've done at three nanometer and two nanometer and even seven nanometer and so many other technologies,” Patel said. “But I think this is maybe a way they can continue to grow in spaces where they haven't been able to traditionally.”
