周一，人工智能行业龙头OpenAI的历史首届开发者大会揭幕。OpenAI首席执行官山姆·奥特曼在接近45分钟的开幕演讲中，向全球开发者和ChatGPT用户展示了AI龙头即将推出的一系列产品更新。
第一个新品是新的升级大模型：GPT-4 Turbo，特性如下：
第二个重要产品是GPTs。用户将能通过自定义指令、拓展（模型）的知识边界和下达行动命令，来构建自己的GPT，并能对外发布给全球更多的人使用。更重要的是，整个构建“自定义GPT”的过程也是通过自然语言对话形成的。 数字化转型网www.szhzxw.cn
下面是我对这两个产品的评论：
一、 GPTs
其实之前已经有很多LLM支持这样的角色定义，包括文心一言、讯飞星火等，里面都通过不同类型的“助手“的形式内置了一些特定的角色，用户也可以自行定义。
不过OpenAI此次推出的GPTs，具备一个特别之处，也带来了巨大的潜在商业价值：用户在创建角色（character）之后，可以与组织内部或者大众分享，甚至通过链接分享给特定用户（目前还没有销售功能，但是可以预见这是个必然趋势：用户创建GPT分身，上架GPT Store，OpenAI和开发者收入分成）。
这貌似是个很小的变化，但是实际上为大量的用户开辟了潜在的商业化道路。过去，为苹果的AppStore做开发者，需要一定的开发技能，才能推出APP，获取分享收入。
但是现在，用户可以只需要掌握一定的提示词（prompt）技巧，就能够通过GPT Builder，打造出一个强大的特定角色，继而为自己赚取收入。也就是从“靠代码赚钱“变成了”靠文字赚钱“，这样的生态会促使越来越多的用户（甚至完全不会编程的用户）开发不同类型的提示词技巧，挖掘GPT的潜力，形成一个正向循环。 数字化转型网www.szhzxw.cn
同时，企业也可以根据自己的业务需要，在内部打造符合特定企业管理风格、企业文化和规章制度（全部都能通过提示词提前设置）的内部GPT。
正如OpenAI的创始成员Andrej Karpathy在X中所总结的：
将会有更多的开发者和更多的GPT。GPT可以读、写、听、说、看、画、思考，使用现有计算作为工具，成为重点领域的专家，参考自定义数据，在数字世界中采取行动，以自定义方式说话或行动，以及共同协作。系好安全带。
当然，下一步需要拭目以待的，是有多少用户成为”提示词开发者“，真正创造、发布角色，有多少消费者愿意付费购买。 数字化转型网www.szhzxw.cn
如果说去年底ChatGPT的推出，是人工智能的iPhone时刻（如英伟达黄仁勋所言），那么此次的发布会，就是人工智能的AppStore时刻。
二、关于GPT-4 Turbo
三个比较大的变化：
1）上下文窗口增加到128K个Token，相当于大约300页的文本。
这意味着GPT可以在一次对话内容中消化大约一本小说长度的内容。这里面涉及到GPT工作的一个机理：在每次对话过程中，我们在持续问答时，实际上每次问题，都会把本Session中所有之前的对话内容，作为当前问题的Prompt输入，这就使得上下文变得越来越冗长。用户可能会明显感觉到，随着问题的增多，GPT的回复发生幻觉或者错误的概率就会越来越高，也就是，它开始忘记一些之前的内容，这都是上下文窗口受限导致的结果。 数字化转型网www.szhzxw.cn
这次的窗口容量增加，将让用户可以在一次对话中，提供更多背景信息和上下文（例如提供一个财报，围绕这个报告提出各种问题），同时GPT又不至于神经错乱、记忆丢失。
2）价格降低
对于开发人员来说，GPT-4 Turbo 的运行成本将降低，每 1000 个输入Token的成本降至 0.01 美元（大约为 750 个单词），而每 1000 个输出Token的成本为 0.03 美元。OpenAI 估计这个新版本比之前的版本便宜三倍。
3）版权责任
这次发布会上的另外一个重要的信息，也是很容易被科技报道所忽视的，就是鉴于很多用户考虑到在使用GPT所输出的内容（例如代码）是否存在版权问题，OpenAI此次宣布（和谷歌和微软之前的做法一致），如果GPT的用户被人投诉版权问题，OpenAI将会承担法律责任。这可以大大减轻很多用户，尤其是企业，在使用GPT方面的合规性顾虑。 数字化转型网www.szhzxw.cn
翻译：
What are the key announcements from OpenAI’s first Developer Conference
On Monday, the first developer conference in the history of artificial intelligence industry leader OpenAI kicked off. In a nearly 45-minute opening speech, OpenAI CEO Sam Altman showed developers and ChatGPT users around the world a series of upcoming product updates from the AI leader.
The first new product is the new upgraded large model: GPT-4 Turbo, with the following features:
The second important product is GPTs. Users will be able to build their own GPT by customizing instructions, extending the boundaries of knowledge and issuing action commands, and will be able to publish it to more people around the world. What’s more, the entire process of building a “custom GPT” is also formed through natural language dialogue.
Here are my reviews of both products:
1. GPTs
In fact, there have been many LLM to support such role definition, including Wenxin Word, Xingfei spark, etc., which has some specific roles built in through different types of “assistants”, and users can also define themselves. 数字化转型网www.szhzxw.cn
However, the GPTs launched by OpenAI has a special feature that also brings huge potential business value: After users create characters, they can share them with the organization or the public, or even share them to specific users through links (there is no sales function at present, but it can be predicted that this is an inevitable trend: Users create GPT doppelgons, put them on the GPT Store, OpenAI and share developer revenue).
This may seem like a small change, but it actually opens up a potential commercialization path for a large number of users. In the past, being a developer for Apple’s APP Store required a certain amount of development skills in order to launch apps and earn shared revenue.
Now, however, users can use GPT Builder to create a powerful, specific character that will earn them an income by simply mastering prompt skills. That is, from “making money by code” to “making money by words”, such an ecosystem will prompt more and more users (even users who can’t program at all) to develop different types of prompt word skills, tap the potential of GPT, and form a positive cycle.
At the same time, enterprises can also build internal GPT according to their own business needs, which conforms to the specific corporate management style, corporate culture and rules and regulations (all can be set in advance by the prompt word). 数字化转型网www.szhzxw.cn
As Andrej Karpathy, a founding member of OpenAI, sums up in X:
There will be more developers and more GPT. GPT can read, write, listen, speak, see, draw, think, use existing computing as a tool, become experts in focus areas, refer to custom data, act in the digital world, speak or act in custom ways, and collaborate together. Fasten your seat belt.
The next step, of course, is to see how many users become “prompt word developers” who actually create and publish characters, and how many consumers are willing to pay for them.
If the launch of ChatGPT at the end of last year was the iPhone moment of artificial intelligence (as Nvidia Jen-Hsun Huang said), then this event is the AppStore moment of artificial intelligence.
2. About GPT-4 Turbo
Three big changes:
1) The context window is increased to 128K tokens, equivalent to about 300 pages of text.
This means that GPT can digest about the length of a novel in a single conversation. This involves a mechanism of how GPT works: During each conversation, when we continue to ask and answer questions, in fact each question, we will take all the previous conversations in this Session as Prompt input for the current question, which makes the context more and more verbose. The user may clearly feel that as the number of questions increases, the probability of GPT’s replies experiencing hallucinations or errors increases, that is, it starts to forget some of the previous content, which is the result of the limited context window. 数字化转型网www.szhzxw.cn
This increased window capacity will allow users to provide more context and context in a single conversation (for example, provide an earnings report and ask questions around the report) without GPT losing its nerve and memory.
2) Lower prices
For developers, GPT-4 Turbo will cost less to run, down to $0.01 per 1,000 input tokens (roughly 750 words) and $0.03 per 1,000 output tokens. OpenAI estimates that this new version is three times cheaper than the previous version.
3) Copyright liability
Another important message from the conference, which is easily overlooked in the tech press, is that since many users are concerned about copyright issues in the output of GPT (such as code), OpenAI announced (in line with Google and Microsoft’s previous practice) that if GPT users are subject to copyright complaints, they will not be able to use it. OpenAI will be held legally responsible. This can greatly alleviate the compliance concerns of many users, especially enterprises, when it comes to using GPT. 数字化转型网www.szhzxw.cn
本文由数字化转型网（www.szhzxw.cn）转载而成，来源于崔博效率手册；编辑/翻译：数字化转型网宁檬树。
免责声明: 本网站(https://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
