打开国产大模型实用大门的金钥匙

(谭营,北京大学教授) 当前,大语言模型研究取得了十分显著的进展,已经成为人工智能领域的重要研究方向和大热点,恍如为人工智能,尤其是通用人工智能的发展注入了一剂强心针。国内学术和产业界都相继追赶这一当前人工智能发展的潮流,涌现出一大批优秀的大语言模型,目前各大高科技公司和研究机构都发布了他们自己的大语言模型,争当该领域发展的潮头。 同时,政府相关部门也高度重视大语言模型和通用人工智能的发展,国家网信办联合多部门公布《生成式人工智能服务管理暂行办法》,促进生成式人工智能健康发展和规范应用。在此背景下,多个国产大模型应运而生,也为国内的人工智能技术和生态打开了新的可能。事实上,相比国外而言,我国大语言模型在第一落点上是集体落后的,但我们现在有机会迎头赶上的其实是第二落点:应用。 全链路自研,打造差异化技术优势 随着大语言模型技术的发展,诸如Chat-GPT、Claude、文心一言等一系列大语言模型如雨后春笋般涌现。 技术自主是推动国产大语言模型产业发展的重要方向。国内企业必须要掌握核心技术,降低对国外技术的依赖,才能有效保护国内产业安全。同时,技术自研使得国内企业能够自主管理数据,保护用户隐私,并为模型提供更准确的本土化信息。 比如从零基础开始训练,采用终身连续学习方法的实用级大模型腾讯混元采用了全链路自研的方式,在模型的各个环节均有自主研发和控制能力,从数据采集、模型训练到应用部署,都能够实现自主闭环,保证了数据的安全性和模型的稳定性。为用户提供高效、可靠的解决方案,并且能够根据自身的需求进行灵活调整和优化。 全链路自研也使得混元具备了更强的创造力和差异化优势。面对大语言模型中普遍存在的幻觉问题,腾讯研究团队并没有完全采用传统的知识图谱和搜索外挂方式,而是创新性的提出了通过底层能力提升来增强大模型水平的方法,使得腾讯混元的幻觉相比主流开源大模型降低了30%至50%,也让大模型更加成熟可靠,具备从“实验室”走向“现实世界”的能力。这种差异化的技术优势,使得混元能够在众多大语言模型中脱颖而出,具有强大的竞争力和应用价值。 全链路自研使得腾讯混元拥有良好的创造力,使得其成为能够面向行业应用、面向产业发展,成为一个“都可以用”的“实用级”大模型。 以腾讯混元作为腾讯云MaaS(Model-as-a-Service)服务的底座,腾讯也面向社会提供大模型相关服务,在开放的平台上,混元与各行各业共同进步,以全新的方式推动产业发展,共创前进新模式。 “实用级”大模型步入全新领域 自ChatGPT问世以来,国内外掀起了大模型研究的巨大浪潮,人们看到了大模型在各个领域展露的惊人表现。大模型的涌现为我们带来了前所未有的机遇和潜力。它们为人们提供了更加智能、个性化的交互体验,推动了人与计算机之间的沟通界面的革新。 随着研究的进展,大模型已进入混战时代。全球范围内已发布的大模型有200余个,其中属于中国的大模型占将近一半。较早期的大模型,例如ChatGPT等,他们更关注于模型的通用性,能涉猎各个领域,但做不到精通每个领域。这些模型通过文本生成、聊天交互等方式,在多个容错较高的领域给出较好的答案,但是遇到专业性更强的场景,模型效果也会随之下降。随着众多大模型的登台,大模型的方向也在出现着分化。注重通用性的大模型如GPT-4,文心一言等开始采用多模态等方式扩大知识输入范围,丰富结果生成形式。MiniMax、第四范式等公司的大模型则越来越强调针对不同应用场景的实际效果。越来越多的模型开始针对性地提升在办公、金融、医疗、文娱、教育、交通等专业领域的实用效果。 相较于国外,国内在办公、文娱、交通等领域短时间内达到了不相上下的效果,但是金融、能源、医疗等领域由于数据等要素的缺失进展缓慢,如何在“实用性”上走通一条路是大模型全方位带动千行白业的一大难点。腾讯混元或能引领“实用级”大模型步入全新领域。 第一,腾讯凭借其广泛的产业布局和深入参与的行业需求,在大模型应用方面具备了独特的优势。得益于多年来在各行业中的技术积累和经验积累,腾讯能够更加明确的知道产业发展真正的需求。在丰富的业务场景的重重考验下,混元将会是一个更贴近应用实际,更能满足产业发展需要的大模型。 第二,从实践中来,也要到实践中去。腾讯在社交媒体和通信、游戏和娱乐、金融科技、教育和医疗健康等领域拥有着广泛的布局和丰富的产品。这为腾讯混元的落地应用提供了得天独厚的条件。腾讯将混元接入各领域的产品中,在行业应用中考验模型效果。以应用结果反馈模型训练与调整。信息在产品与模型间的正循环,让混元更贴合行业需要,也实际性的带动了行业应用与发展。 第三,持续学习、迭代更新能让模型长期保持“实用性”。腾讯拥有坚实的硬件基础,掌握万亿参数规模的快速训练方法。随着数据集的扩大,落地效果反馈的汇入,混元能快速迭代更新,不断学习新的知识,不断提高准确性和实用性。快速的反应能力和学习能力,也正是我国始终走在新科技革命浪潮前沿不可或缺的能力。 总之,以腾讯混元为代表的国产大模型,在全链路自研和行业应用方面展现出强大的优势,为中国企业探索中国特色的人工智能大模型之路提供了契机。以大模型的创新力量为指引,我们可以期待国产大模型全新的变革式发展以及引领力量的涌现。通过大规模应用人工智能大语言模型,中国企业将在数字化转型中取得更具竞争力的地位,为我国在全球数字经济中的发展注入新的动力,也为我们的社会和经济带来深远的影响。

Congratulation on the successful completion of “The Thirtheenth International Conference on Swarm Intelligence (ICSI’2022)” held from July 15th to 19th!

On July 17, The 13th International Conference on Swarm Intelligence (ICSI’2022) was Successfully Concluded. Due to the rapid overspreading of the Omicron variant of COVID-19, the ICSI 2022 organizing committee has made the decision that our ICSI 2022 will continue as scheduled on July 15-19, by being converted to a Read more…

Congratulations to Professor Tan Ying for being selected as the 2021 Chinese Highly Cited Scholar

On April 14, Elsevier officially released the 2021 “Highly Cited Chinese Researchers” list. The 2021 Elsevier “Highly Cited Scholars in China” list is based on Scopus, the world’s most authoritative citation and index database, as the statistical source of scientific research achievements of Chinese scholars. This list Adopts the selection Read more…

Congratulation for the successful completion of “The Twelfth International Conference on Swarm Intelligence (ICSI’2021)” held during July 17th and 21st!

The Twelfth International Conference on Swarm Intelligence (ICSI’2021) has been successfully held on July 17th to 21st 2021 in Qingdao, China. ICSI’2021 invited leading researchers in the field of swarm intelligence and delivered high-quality speeches and talks after the opening ceremony. Prof. Qirong Tang presented his speech on “Swarm Intelligence Read more…

Congratulation for the successful completion of “2020 Peking University Online Graduate Academic Conference on Frontier of Swarm Robotics Research” held during November 7th and 8th!

The 2020 Peking University Online Graduate Academic Conference on Frontier of Swarm Robotics Research was successfully held on November 7-8, 2020. This conference was sponsored by Graduate School of Peking University and International Association of Swarm and Evolutionary Intelligence, undertaken by Computational Intelligence Laboratory of Peking University, and funded by Read more…

Congratulations to Professor Yuhui Shi of Southern University of Science and Technology for being the cover author of IEEE Xplore!

Congratulations to Professor Yuhui Shi of Southern University of Science and Technology for being the cover author of the IEEE Xplore Digital Library. Chair Professor of Computer Science and Engineering Department of Southern University of Science and Technology, expert in the field of international computational intelligence, founder of brainstorming optimization Read more…

Prof. Ying Tan’s new book “Handbook of Research on Fireworks Algorithms and Swarm Intelligence” has been published by IGI Global in December, 2019.

Prof. Ying Tan’s new book “Handbook of Research on Fireworks Algorithms and Swarm Intelligence” has been published by IGI Global in December, 2019. Professor Tan is a leading expert in the filed of swarm intelligence. He leads the computational intelligence laboratory in Peking University and has contributed valuable work to Read more…

International Association of Swarm and Evolutionary Intelligence Established

The International Association of Swarm and Evolutionary Intelligence (IASEI) was officially launched on 27 July 2019 during the tenth International Conference on Swarm Intelligence (ICSI’2019) and the fourth International Conference on Data Mining and Big Data (DMBD’2019) at Chiang Mai, Thailand.  All ICSI conference participants are automatically enrolled as the Read more…