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，讨论这一革命性技术的现在与未来。遗憾的是，Niki Parmar因故未能出席这一历史性的聚会。这是Transformer论文的作者团队首次在公众面前集体亮相，他们的对话充满了深度和洞察力。
在对话中，作者们表达了对Transformer模型未来发展的一些独特见解。他们指出，尽管Transformer已经在自然语言处理领域取得了显著成就，但世界仍期待着超越Transformer的新技术，以达到更高的性能水平。他们最初的研发目标是模拟Token的演化过程，希望模型能够实现非线性的文本和代码生成，而不仅仅是线性地处理信息。
讨论中，作者们还提出了一项重要议题，即大模型在解决简单问题时可能过度消耗资源，例如2+2这样的基础运算。他们认为，未来的趋势应该是自适应计算，即模型能够根据问题的复杂性智能地分配计算资源，避免浪费。
此外，他们还反思了当前模型的经济性和规模。尽管现有的大模型价格相对低廉，大约每百万个Token只需1美元，比购买一本平装书便宜100倍，但他们认为这仍存在优化空间，模型的规模和成本效率仍有待提升，以适应更广泛的应用场景。
这次论坛揭示了Transformer背后团队的深思熟虑，也预示着自然语言处理领域未来可能的变革方向。随着技术的不断演进，我们期待着新的突破，为人工智能的未来开辟新的道路。
英语如下：
News Title: “Jensen Huang in Dialogue with Transformer’s Key Authors: Exploring New Frontiers Beyond the Transformer”
Keywords: Transformer authors, GTC conference, Jensen Huang conversation
News Content: At this year’s Global Graphics Technology Conference (GTC), NVIDIA founder Jensen Huang hosted an unconventional roundtable discussion, assembling seven primary authors of the Transformer model to discuss the revolutionary technology’s present and future. Regrettably, Niki Parmar was unable to attend this historic gathering due to unforeseen circumstances. This marked the first public appearance of the Transformer paper’s author team as a collective, with their conversation brimming with depth and insight.
During the dialogue, the authors shared their unique perspectives on the future development of the Transformer model. They acknowledged that while Transformer has made remarkable strides in natural language processing, the world is eagerly awaiting新技术超越Transformer to achieve even higher performance levels. Their initial research goal was to emulate the evolution of Tokens, aiming for models that could generate text and code nonlinearly, not just process information linearly.
An essential topic of discussion was the potential overconsumption of resources by large models when tackling simple tasks, such as basic arithmetic operations like 2+2. The authors posited that the future trend should be adaptive computation, where models intelligently allocate computational resources based on the complexity of the problem, thus preventing waste.
They also reflected on the current economics and scale of the models. While existing large models are relatively inexpensive, costing approximately $1 for a million Tokens – 100 times cheaper than a paperback book – they believe there is room for optimization. The scale and cost efficiency of models need improvement to accommodate a broader range of applications.
This forum showcased the thoughtful deliberations behind the Transformer team and foreshadowed potential directions for变革 in the natural language processing domain. As technology continues to evolve, we anticipate new breakthroughs that will pave the way for the future of artificial intelligence.
【来源】https://new.qq.com/rain/a/20240321A00W5H00
Views: 1