在今年的全球图形技术大会（GTC）上，英伟达创始人黄仁勋举办了一场别开生面的圆桌论坛，邀请了Transformer模型的七大原创作者共聚一堂，进行深度对话。值得注意的是，其中的Niki Parmar因故未能出席这一历史性的聚会。这七位引领自然语言处理领域革命的科学家们首次集体在公众视野中亮相，他们的见解引发了广泛关注。
在论坛中，作者们表达了对Transformer模型未来发展的一些独特看法。他们指出，尽管Transformer在当前取得了显著成就，但世界的需求远不止于此，期望能有更优秀的技术取而代之，推动性能达到新的高峰。他们坦诚，Transformer的原始目标是模拟Token的演化，而不仅仅是线性的文本生成，这一目标尚未完全实现。
讨论中，作者们还触及了大模型的计算资源问题。他们举例说，即便是简单的算术问题，如2+2，也可能需要动用到万亿参数的模型。因此，他们认为未来的趋势将是自适应计算，即根据问题的复杂性动态分配计算资源。
此外，他们认为当前的AI模型在经济性和规模上仍有提升空间。以1美元处理百万个Token的价格计算，这个成本比购买一本平装书还要便宜100倍，显示出AI模型在效率和性价比上的巨大潜力。这一观点为未来的AI发展方向提供了新的思考角度。
来源：腾讯科技
英语如下：
News Title: “Transformer Pioneers Gather at GTC: Pursuing Innovation and Envisioning the Future of Computing Revolution”
Keywords: Transformer creators, GTC conference, Huang Renxun dialogue
News Content: At this year’s Global Graphics Technology Conference (GTC), NVIDIA founder Huang Renxun hosted a groundbreaking roundtable discussion, assembling the seven original authors of the Transformer model for an in-depth conversation. Notably, Niki Parmar was absent from this historic gathering due to unforeseen circumstances. These seven scientists, who have revolutionized the field of natural language processing, made their first collective public appearance, drawing significant attention.
During the forum, the authors shared unique perspectives on the future development of the Transformer model. They acknowledged its remarkable achievements but emphasized the world’s appetite for even better technology to push performance to new heights. They disclosed that the original aim of Transformer was to emulate the evolution of tokens, beyond linear text generation, a goal that has yet to be fully realized.
The authors also delved into the issue of computational resources for large models. They illustrated that even simple arithmetic, like 2+2, might require models with trillions of parameters. Thus, they believe the future trend will be adaptive computing, dynamically allocating resources based on the complexity of the problem.
Furthermore, they contended that current AI models still have room for improvement in terms of efficiency and scale. At a cost of processing a million tokens for just a dollar, this is 100 times cheaper than purchasing a paperback book, highlighting the tremendous potential of AI models in both efficiency and cost-effectiveness. This perspective offers a new angle for considering the future direction of AI development.
Source: Tencent Technology
【来源】https://new.qq.com/rain/a/20240321A00W5H00
Views: 3