对于AI行业从业者来说，刚刚可能是一夜未眠。
北京时间11月7日凌晨，美国人工智能公司OpenAI的开发者大会正式开启，创始人Sam Altman在台上和同事，只用45分钟时间，就「轰」出了团队最新的成果GPT-4 Turbo，后者不仅更快、有更长的上下文、而且更好的控制。 数字化转型网www.szhzxw.cn
同时，OpenAI下调API的价格至原来的1/3，降到了1000输入/美分，让在场开发者欢呼不已。
当然，更重要的是，OpenAI推出了「GPTs」——让人们能用自然语言构建定制化GPT，然后，你猜到了——可以把GPT上传到即将发布的「GPT Store」！
如果说GPT-4 Turbo是更好用的「iPhone」，GPT Store则可能是让OpenAI成为「苹果」一样的巨头的重要一步。
当竞争对手们依然在「AI炼丹」时，OpenAI已经开始构建起一个看起来相当宏伟的生态了。
一、GPT-4 Turbo，更快，更省钱
发布会一开始，Sam Altman就宣布了GPT-4的一次大升级，推出了GPT-4 Turbo，同步在ChatGPT和API版本推出。
Sam Altman表示团队一直在征求开发者的建议，对开发者关注的问题做了六大升级，分别是更长的上下文长度、更强的控制、模型的知识升级、多模态、模型微调定制和更高的速率限制。
其中前四条主要关于新模型的性能的提升，而后两点则主要针对企业开发者的痛点。在提升性能的同时，OpenAI还宣布了API价格的下调，可谓「加量不加价」了。
六大升级中，第一，就是上下文长度。
OpenAI原本提供的最长的上下文长度为32k，而此次，GPT-4 Turbo直接将上下文长度提升至128k，一举超过了竞争对手Anthropic的100k上下文长度。 数字化转型网www.szhzxw.cn
128k的上下文大概是什么概念？大概约等于300页标准大小的书所涵盖的文字量。除了能够容纳更长上下文外，Sam还表示，新模型还能够在更长的上下文中，保持更连贯和准确。
第二，是为开发者提供了几项更强的控制手段，以更好地进行API和函数调用。
首先，新模型提供了一个JSON Mode，可以保证模型以特定JSON方式提供回答，调用API时也更加方便。
另外，新模型还允许同时调用多个函数，同时引入了seed parameter，在需要的时候，可以确保模型能够返回固定输出。接下来几周，模型还将增加新功能，让开发者能看到log probs。
第三，则是模型内部和外部知识库的升级。
ChatGPT横空出世大概一年后，GPT的知识库终于更新到了2023年4月。Sam Altman承诺未来还将继续更新其知识库，不使其落伍。「对于GPT的知识停留在2021年，我们和你们一样，甚至比你们更恼火。」Sam Altman表示。
除了内部知识库的升级，GPT-4 Turbo也升级了外部知识库的更新方式，现在可以上传外部数据库或文件，来为GPT-4 Turbo提供外部知识库的支持。
第四，或许是最不让人意外的，多模态。
新模型支持了OpenAI的视觉模型DALL·E 3，还支持了新的文本到语音模型——开发者可以从六种预设声音中选择所需的声音。
GPT-4 Turbo现在可以以图生图了。同时，在图像问题上，目前OpenAI推出了防止滥用的安全系统。OpenAI还表示，它将为所有客户提供牵涉到的版权问题的法律费用。
在语音系统中，OpenAI表示，目前的语音模型远超市场上的同类，并宣布了开源语音识别模型Whisper V3。 数字化转型网www.szhzxw.cn
第五，模型微调与定制。
8月，OpenAI曾经发布过GPT-3.5 Turbo的微调服务。当时，有早期测试表明，经过微调的GPT-3.5 Turbo版本在某些任务中甚至可以超越GPT-4，不过定价相对较高。
而此次，Sam宣布GPT-3.5 Turbo 16k的版本目前也可以进行微调的定制了，且价格将比前一代更低。GPT-4的微调定制也在申请中了。
同时，OpenAI也开始接受单个企业的模型定制了。「包括修改模型训练过程的每一步，进行额外的特定领域的预训练，针对特定领域的后训练等等。」Sam表示。同时他表示，OpenAI没有办法做很多这样的模型定制，而且价格不会便宜。
第六，也是最后一点，是更高的速率限制。
GPT-4用户，发布会后马上可以享受到每分钟的速率限制翻倍的体验。同时，如果不够满意，还可以进一步通过API账户，申请进一步提升速率限制。
六大升级以外，是API体系的全线降价。
此次新发布的GPT-4 Turbo，输入方面降至GPT-4的1/3，而输出方面降至一半，OpenAI表示，总体使用上降价大概2.75倍。 数字化转型网www.szhzxw.cn
新模型的价格是每千输入token 1美分，而每千输出token 3美分。降价的API迎来了现场开发者的欢呼。
Sam还表示，在优先解决价格之后，下一个重点解决的问题将是速度问题，很快，开发者们就会发现GPT-4 Turbo将变快很多。
二、GPT Store来了！
早在5月，OpenAI就开放了插件系统，首批上线了70个大模型相关的应用，领域包括猜词、翻译、查找股票数据等等。
当时，该功能被寄予厚望，不少媒体将其类比于苹果的App Store时刻，认为它将改变大模型应用的生态。不过虽然后期插件不断增加，但插件系统却远远没有达到苹果应用商店的影响力。
而此次发布会上，OpenAI则重新梳理了其应用商店的体系，并将其扩大到了一个全新的范畴——人人都能通过自然语言创建基于自己的知识库的AI Agent，加入OpenAI的应用商店，并获得分成。
OpenAI此次发布的应用，不再称为插件，而选择了一个相对比较奇怪的名字，GPTs。而整体的应用商店，名字叫做GPT Store，将在本月后期正式推出。 数字化转型网www.szhzxw.cn
按照Sam Altman的说法，每一个GPT像是ChatGPT的一个为了特殊目的而做出的定制版本。
为了突出新的GPT应用，ChatGPT整个页面将有小幅度的调整。左上角除了ChatGPT，下面的应用，就是此次推出的GPT应用。
在演示中，可以看到，较为复杂的插件——比如曾经在OpenAI上线插件系统时第一批进入插件系统的Zapier，仍然存在于目前的应用商店中，而且仍然可能是未来应用商店里很重要的一批应用。
演示中，OpenAI的Jessica Shay，就利用了Zapier链接了自己的日历和手机短信，通过与Zapier这个应用聊天的方式，直接安排了自己的日程，并通知了同事。
不过，Zapier的功能虽然强大，这样的应用并不是此次发布的重点。据Glassdoor数据显示，Zapier公司拥有500-1000名员工，而财富网站报道，Zapier估值已达50亿美金。指望这样的应用来填充OpenAI的羽翼未丰的应用商店，使其成为一个丰富的生态显然不太现实。
因此，此次发布中，OpenAI推出一个重磅发布：让不懂代码的人也能轻松定义一个GPT。
Sam Altman为此进行了现场展示。
「在YC工作过很多年，我总是遇到开发者向我咨询商业意见。」Sam Altman讲到，「我一直想，如果有一天有个机器人能帮我回答这些问题就好了。」 数字化转型网www.szhzxw.cn
接着，Sam Altman打开了GPT Builder，先打上一段对这个GPT的定义，类似于帮助初创公司的创始人思考他们的业务创意并获得建议，接着，在对话中，GPT Builder自己生成了这个GPT的名字、图标，并通过与Sam对话的形式，询问Sam是否要对对生成的名字和图标等进行调整。
接下来，GPT Builder主动向他询问这个应用该如何与用户交互，Sam表示可以从我的过往演讲中选择合适且有建设性的回答，然后上传了一段自己过往的演讲。
即使加上讲解，整个应用也在三分钟内就完成了。访问这个GPT的人，会收到GPT自动生成的对话开头，可以与这个GPT对话咨询创业相关的内容，而得到的，将是一个类似于Sam Altman本人的回答。
Sam表示，创建者还可以进一步为GPT增加action（动作）。
创建一个这样的GPT，本质上，用户能够定制的功能其实并不多：指令（预设的prompt），外设的知识库和动作。但是，能把三者丝滑地结合起来，让一个不懂代码的人也能更简单地创建应用，确实是此次Open AI的创举。
GPT发布后，应用可以选择私有，专属企业拥有和公开所有三种方式。而Open AI表示，将为受欢迎的应用提供利润分享。
很明显，OpenAI在这里的发布，并没有希望普通用户能够通过自然语言创建出多么复杂的应用，其中的想象空间，更重要的在于个人和企业能够将自己的知识库上传到OpenAI，一键构建专属应用。
比如作为一个拥有货运价格表的代理，可以将文件上传到OpenAI后，一键部署出自己的询价助理，这样简洁丝滑的应用部署，在之前还并不存在。而如果发布最终能够得到用户认可的话，类似的应用也将能够填充OpenAI的应用商店，使其成为各种信息的宝库。 数字化转型网www.szhzxw.cn
三、零代码创建AI Agent
如果你觉得上述的0代码的GPT很酷，此次OpenAI也推出了让开发者更容易使用OpenAI API的开发方式——Assistants API。
Sam Altman表示，市面上基于API构建agent的体验很棒。比如，Shopify的Sidekick可以让用户在平台上采取行动，Discord的Clyde可以让管理员帮忙创建自定义人物，Snap的My AI是一个自定义聊天机器人，可以添加到群聊中并提出建议。
但问题是，这些agent很难建立。有时需要几个月的时间，由数十名工程师组成的团队，处理很多事情才能使这种定制助手体验。这些事情包括状态管理（state management）、提示和上下文管理（prompt and context management）、扩展功能（extend capabilities）和检索（retrievel）。
在OpenAI开发者大会上，这些事情被API化——OpenAI推出Assistants API，让开发人员在他们的应用程序中构建「助手」。 数字化转型网www.szhzxw.cn
使用Assistants API，OpenAI客户可以构建一个具有特定指令、利用外部知识并可以调用OpenAI生成式AI模型和工具来执行任务的「助手」。像这样的案例范围包含，从基于自然语言的数据分析应用程序到编码助手，甚至是人工智能驱动的假期规划器。
Assistants API封装的能力包括：
持久的线程（persistent threads），人们不必弄清楚如何处理长的对话历史；
内置的检索（Retrieval），利用来自OpenAI模型外部的知识（例如公司员工提供的产品信息或文档）来增强开发人员创建的助手；提供新的Stateful API管理上下文；
内置的代码解释器（Code Interpreter），可在沙盒执行环境中编写和运行Python代码。这一功能于3月份针对ChatGPT推出，可以生成图形和图表并处理文件，让使用Assistants API创建的助手迭代运行代码来解决代码和数学问题； 数字化转型网www.szhzxw.cn
改进的函数调用，使助手能够调用开发人员定义的编程函数并将响应合并到他们的消息中。
Assistants API处于测试阶段，从今天开始可供所有开发人员使用。开发者可以前往Assistants Playground来尝试Assistants API测试版，而无需编写任何代码。
Assistants API被OpenAI视为帮助开发者在其应用程序中构建「类agent体验」的第一步。有了Assistants API，构建agent应用将变得更容易。OpenAI表示，随着时间的推移，将会持续提高它们的能力。并且，未来计划允许客户提供自己的copilot工具，以补充其平台上的Code Interpreter、检索组件和函数调用。
OpenAI开发者大会上的产品升级，再次告诉人们，距离每个人都能有一个甚至多个专属私人助理、使用自然语言就能开发软件、还能像浏览应用一样，付费/免费购买流行的私人助理，这样的一个未来，正在加速向人们走来。
从GPT-4到GPT-4 Turbo和GPT Store，OpenAI只用了半年多一点的时间。而在6个月之内，全球的科技和AI行业，已经是天上人间。
当多模态、长文本输入、更便宜、个性化……这些旨在和OpenAI大模型错位竞争的特点，都被OpenAI抢先拿来自我革命；当OpenAI的产品在B端和C端都体现出强大的吸引力、并且还将利用GPT Store率先笼络住全球AI开发者时，不知道全球的AI对手们的心情如何。 数字化转型网www.szhzxw.cn
但OpenAI的技术进步是令人兴奋的，而团队所采用的商业策略，又有着超出一般创业公司的成熟——我们目睹着一个行业的潮起，也可能正在见证一个巨头的诞生。
翻译：
In a short 45-minute conference, how did OpenAI once again keep the AI circle awake all night
For those in the AI industry, it may have been a sleepless night.
In the early morning of November 7, Beijing time, the developer conference of the American artificial intelligence company OpenAI was officially opened, and founder Sam Altman and his colleagues on the stage in only 45 minutes, “bang” out of the team’s latest achievement GPT-4 Turbo, which is not only faster, has a longer context, but also better control.
At the same time, OpenAI lowered the price of the API to one-third of the original, down to 1000 inputs/cents, to the cheers of developers present.
Of course, more importantly, OpenAI introduced “GPTs” – allowing people to build custom GPT in natural language, and then, you guessed it – upload GPT to the upcoming “GPT Store”!
If GPT-4 Turbo is the “iPhone” that makes it easier to use, GPT Store could be an important step toward making OpenAI a giant like Apple.
While competitors are still “AI refining,” OpenAI has begun to build what looks like a pretty grand ecology. 数字化转型网www.szhzxw.cn
First, GPT-4 Turbo, faster, more economical
At the start of the event, Sam Altman announced a big upgrade to GPT-4, introducing the GPT-4 Turbo, which was simultaneously rolled out in ChatGPT and API versions.
Sam Altman said the team has been asking for suggestions from developers and has made six major upgrades to developer concerns: longer context length, stronger controls, model knowledge upgrades, multimodality, model fine-tuning customization, and higher rate limits.
The first four focus on the performance improvements of the new model, and the second two focus on the pain points of enterprise developers. At the same time of improving performance, OpenAI also announced a reduction in API prices, which can be described as “increasing volume without increasing price”.
Of the six major upgrades, the first is context length.
OpenAI originally provided a maximum context length of 32k, and this time, GPT-4 Turbo directly increased the context length to 128k, surpassing the 100k context length of competitor Anthropic.
What is the general concept of 128k context? It is approximately equal to the amount of text contained in a 300-page standard-sized book. In addition to being able to accommodate longer contexts, Sam says the new model is also able to be more coherent and accurate in longer contexts.
The second is to provide developers with several stronger controls to make better API and function calls.
First, the new model provides a JSON Mode, which ensures that the model provides answers in a specific JSON manner and makes it easier to call the API. 数字化转型网www.szhzxw.cn
In addition, the new model allows multiple functions to be called at the same time, and introduces seed parameter to ensure that the model can return a fixed output when needed. In the coming weeks, new features will be added to the model that will allow developers to see log probs.
Third, upgrade the knowledge base inside and outside the model.
About a year after ChatGPT came out, the GPT knowledge base was finally updated to April 2023. Sam Altman promises to keep updating its knowledge base in the future. “We are as annoyed as you are, if not more, that the knowledge of GPT is stuck in 2021.” Sam Altman says.
In addition to the upgrade of the internal knowledge base, GPT-4 Turbo has also upgraded the update mode of the external knowledge base. You can now upload external databases or files to support the external knowledge base of GPT-4 Turbo.
Fourth, and perhaps least surprising, is multimodality.
The new model supports OpenAI’s visual model, DALL·E 3, as well as a new text-to-speech model – developers can choose from six preset sounds.
The GPT-4 Turbo can now make pictures from pictures. At the same time, on the issue of images, OpenAI has introduced a security system to prevent abuse. OpenAI also said it would cover the legal costs of any copyright issues involved for all of its customers. 数字化转型网www.szhzxw.cn
In speech systems, OpenAI said that the current speech model is far from the market, and announced the open source voice recognition model Whisper V3.
Fifth, model fine-tuning and customization.
In August, OpenAI released a fine-tuning service for GPT-3.5 Turbo. At the time, early tests showed that the fine-tuned version of the GPT-3.5 Turbo could even outperform the GPT-4 in some tasks, though at a relatively high price.
This time, Sam announced that the GPT-3.5 Turbo 16k version can now be fine-tuned and customized, and the price will be lower than the previous generation. Fine-tuning customization of GPT-4 is also being requested.
At the same time, OpenAI is beginning to accept model customization for individual enterprises. “This includes modifying each step of the model training process, doing additional domain-specific pre-training, domain-specific post-training, and so on.” “Says Sam. At the same time, he said, OpenAI can’t do a lot of such model customization, and the price will not be cheap.
The sixth and final point is higher speed limits. 数字化转型网www.szhzxw.cn
GPT-4 users can enjoy a doubling of the speed limit per minute immediately after the launch. At the same time, if you are not satisfied, you can further apply for a further increase in the rate limit through the API account.
In addition to the six major upgrades, the price of the API system is reduced across the board.
With the newly released GPT-4 Turbo, the input is reduced to one-third of the GPT-4, while the output is reduced to half, and OpenAI says the overall usage price is reduced by about 2.75 times.
The price of the new model is 1 cent per thousand input tokens and 3 cents per thousand output tokens. The reduced price of the API welcomed the cheers of the developers in the field.
Sam also said that after prioritizing price, the next priority will be speed, and soon, developers will find that the GPT-4 Turbo will be much faster. 数字化转型网www.szhzxw.cn
Second, GPT Store is coming!
Back in May, OpenAI opened its plug-in system and launched the first 70 big model-related applications, including guessing words, translating, finding stock data, and more.
At the time, the feature was highly anticipated, and many media likened it to Apple’s App Store moment, thinking that it would change the ecology of large model apps. However, although the late plug-ins continue to increase, the plug-in system is far from reaching the influence of Apple’s App Store.
At this press conference, OpenAI has re-combed its app store system and expanded it to a new category – everyone can create AI agents based on their own knowledge base through natural language, join OpenAI’s app store, and get a share.
OpenAI’s application is no longer called a plug-in, but has chosen a relatively strange name, GPTs. The full app Store, called the GPT Store, will launch later this month.
According to Sam Altman, each GPT is like a customized version of ChatGPT for a special purpose.
In order to highlight the new GPT application, the entire ChatGPT page will be slightly adjusted. In addition to ChatGPT in the upper left corner, the following application is the GPT application launched this time. 数字化转型网www.szhzxw.cn
In the demo, you can see that more complex plugins, such as Zapier, which was one of the first to enter the plugin system when OpenAI launched the plugin system, still exist in the current app store, and may still be an important batch of apps in the future app store.
In the demo, OpenAI’s Jessica Shay used Zapier to link her calendar to her phone’s text message, chat with the Zapier app, and directly arrange her schedule and notify her colleagues.
However, while Zapier is powerful, such applications are not the focus of this release. According to Glassdoor, Zapier has 500 to 1,000 employees, and Fortune reports that Zapier is valued at $5 billion. Expecting such apps to populate OpenAI’s fledgling app store and make it a rich ecosystem is obviously not realistic.
Therefore, in this release, OpenAI has launched a blockbuster release: to make it easy for people who don’t know the code to define a GPT.
Sam Altman showed it live. 数字化转型网www.szhzxw.cn
“Over the years at YC, I’ve always had developers ask me for business advice.” Sam Altman said, “I always thought it would be nice if one day there was a robot that could answer these questions for me.”
Next, Sam Altman opened GPT Builder, typed a definition of the GPT, similar to helping startup founders think about their business ideas and get suggestions, and then, in the conversation, GPT Builder generated the name and icon of the GPT itself, and through the form of a conversation with Sam, Ask Sam if he wants to make adjustments to the generated names, ICONS, etc.
Next, GPT Builder asked him how the app should interact with users, and Sam said he could select appropriate and constructive responses from my past talks, and then uploaded a clip of his own.
Even with the explanation, the whole app was finished in three minutes. Visitors to this GPT will receive a GPT auto-generated conversation starter, can ask about entrepreneurship with this GPT conversation, and will get an answer similar to that of Sam Altman himself.
According to Sam, creators can further add actions to GPT.
To create such a GPT, there is essentially not much that the user can customize: instructions (default prompt), peripheral knowledge base, and actions. However, the ability to silkily combine the three, so that a person who does not know the code can more easily create applications, is indeed the innovation of Open AI. 数字化转型网www.szhzxw.cn
After the release of GPT, applications can choose between private, exclusive enterprise owned, and public all three ways. Open AI says it will offer profit sharing for popular apps.
Obviously, OpenAI’s release here does not expect ordinary users to be able to create complex applications through natural language, where the imagination space, more importantly, is that individuals and enterprises can upload their knowledge base to OpenAI, and build exclusive applications with one click.
For example, as an agent with a freight price list, you can upload files to OpenAI and deploy your own inquiry assistant with one click, which is a simple and silky application deployment that did not exist before. If the release is eventually accepted by users, similar apps will be able to fill the OpenAI App Store, making it a treasure trove of information.
Third, zero code to create AI Agent
If you think the 0 code GPT mentioned above is cool, OpenAI is also launching Assistants apis to make it easier for developers to use the OpenAI API. 数字化转型网www.szhzxw.cn
Sam Altman says that the experience of building agents based on apis is great. For example, Shopify’s Sidekick lets users take action on the platform, Discord’s Clyde lets administrators help create custom personas, and Snap’s My AI is a custom chatbot that can be added to group chats and make suggestions.
But the problem is, these agents are hard to build. Sometimes it takes months, with teams of dozens of engineers, to handle a lot of things to make this custom assistant experience. These things include state management, prompt and context management, extend capabilities, and retrievel.
At the OpenAI Developer Conference, these things were API-enabled – OpenAI launched Assistants apis to let developers build “assistants” into their applications.
Using Assistants apis, OpenAI customers can build an “assistant” that has specific instructions, draws on external knowledge, and can invoke OpenAI generative AI models and tools to perform tasks. Examples like this range from natural language-based data analysis applications to coding assistants and even AI-powered vacation planners. 数字化转型网www.szhzxw.cn
Assistants API encapsulation capabilities include:
persistent threads, where people don’t have to figure out how to handle long conversation histories;
Built-in Retrieval, leveraging knowledge from outside the OpenAI model (such as product information or documentation provided by company employees) to enhance developer-created assistants; New Stateful API management context;
The built-in Code Interpreter for writing and running Python code in a sandbox execution environment. The feature, launched for ChatGPT in March, generates graphs and diagrams and processes files, allowing Assistants created using Assistants apis to iterate through code to solve code and math problems;
Improved function calls that enable assistants to call developer-defined programming functions and incorporate responses into their messages. 数字化转型网www.szhzxw.cn
Assistants API is in beta and will be available to all developers starting today. Developers can go to the Assistants Playground to try out the Assistants API beta without having to write any code.
Assistants apis are seen by OpenAI as the first step in helping developers build “agent-like experiences” into their applications. With the Assistants API, building agent applications will be easier. OpenAI says it will continue to improve their capabilities over time. And, there are future plans to allow customers to provide their own copilot tools to complement Code Interpreter, retrieval components, and function calls on their platform.
The product upgrade at the OpenAI Developer conference once again told people that a future where everyone can have one or even multiple dedicated personal assistants, develop software using natural language, and buy popular personal assistants for free or pay just like browsing apps is accelerating towards people. 数字化转型网www.szhzxw.cn
From GPT-4 to GPT-4 Turbo and GPT Store, OpenAI took just over half a year. And within six months, the global technology and AI industry has been heaven and earth.
When multimodal, long text input, cheaper, personalized… These features, which are designed to compete with OpenAI’s large model dislocation, have been preempted by OpenAI to revolutionize itself; When OpenAI’s products show a strong attraction at both the B and C ends, and will also use the GPT Store to take the lead in attracting global AI developers, I do not know how the mood of global AI rivals will be.
But OpenAI’s technological advances are exciting, and the business strategy the team has adopted is more mature than the average startup – we are witnessing the rise of an industry, and we may be witnessing the birth of a giant. 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，来源于极客公园；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。
