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模型的七位主要作者共聚一堂，共享智慧的碰撞。值得注意的是，Transformer的贡献者之一Niki Parmar因故未能出席此次盛会。这是自Transformer诞生以来，其主要创作者们首次在公开场合集体亮相，吸引了业界的广泛关注。
在对话中，七位作者表达了对Transformer模型未来发展的一些深思熟虑的观点。他们一致认为，尽管Transformer已经在自然语言处理领域取得了显著成就，但世界仍期待超越Transformer的创新。他们希望看到一种新的技术，能够引领性能的崭新高峰，更好地模拟文本和代码的动态演化过程，而不仅仅是线性的生成。
在讨论中，作者们也触及到模型计算效率的问题。他们指出，即使是解决像“2+2”这样简单的数学问题，当前的大模型也可能需要动用万亿级别的参数资源。因此，他们认为自适应计算是未来发展的关键，以确保在特定问题上恰当地分配计算资源，避免浪费。
此外，他们还对当前模型的成本效益进行了反思。目前，模型的使用成本大约是每百万个token花费1美元，这比购买一本平装书便宜100倍。尽管如此，他们认为现有的模型规模仍然较小，性价比还有提升空间，这也预示着未来模型发展将更加注重经济性和规模的有效平衡。
这场圆桌论坛的对话，不仅揭示了Transformer模型的内在挑战，也为人工智能领域的未来发展提供了富有洞察力的思考方向。随着技术的不断迭代，我们期待看到更加智能、高效且经济的模型涌现，继续推动自然语言处理的边界。
英语如下：
**News Title:** “Huang Renxun and the Transformer Paper’s Key Authors Engage in First Public Dialogue: Exploring New Frontiers Beyond Transformers”
**Keywords:** Transformer authors, Huang Renxun forum, large model evolution
**News Content:** At this year’s Global Graphics Technology Conference (GTC), NVIDIA founder Huang Renxun hosted an innovative roundtable discussion, bringing together the seven main authors of the Transformer model for a gathering of intellectual exchange. Notably, Niki Parmar, one of the contributors to Transformer, was unable to attend the event due to unforeseen circumstances. This marked the first time since the Transformer’s inception that its primary creators appeared collectively in a public forum, attracting widespread attention from the industry.
During the dialogue, the seven authors shared their well-considered perspectives on the future development of the Transformer model. They agreed that while Transformer has achieved remarkable success in natural language processing, the world is eagerly anticipating innovations that surpass it. They expressed a desire for a new technology that could lead to unprecedented performance peaks, better simulating the dynamic evolution of text and code, going beyond linear generation.
The authors also addressed the issue of model computational efficiency. They pointed out that even solving a simple math problem like “2+2” might require trillions of parameter resources in current large models. As a result, they emphasized that adaptive computation would be crucial for future progress, ensuring the appropriate allocation of computational resources for specific tasks to prevent waste.
Moreover, they reflected on the cost-effectiveness of current models. At present, the cost of using a model amounts to roughly $1 per million tokens, which is 100 times cheaper than purchasing a paperback book. Nevertheless, they believe that existing model sizes are still relatively small, indicating room for improvement in cost-to-benefit ratios, suggesting a focus on balancing economy and scale in future model development.
This roundtable discussion not only illuminated the inherent challenges of the Transformer model but also offered insightful directions for the future development of artificial intelligence. As technology continues to evolve, we look forward to witnessing more intelligent, efficient, and economically viable models pushing the boundaries of natural language processing.
【来源】https://new.qq.com/rain/a/20240321A00W5H00
Views: 1