GPT 4 出来之后，图片处理，逻辑和理解能力，以及对于上下文的处理能力，让业界为之震撼，GPT-4发布之后，OpenAI 的开发者日之后，几天的时间中，微信里面铺天盖地看到的都是 GPT 和 OpenAI 的文章。
特别是给出一张网站页面的设计草图，GPT 4 直接给出了网站产品的实现，这简直是不可思议。 最早 GPT-4 出来的时候，这个功能是没有开放的。
如今随着 11月6日 GPT-4-Turbo 的发布，给出页面设计操作，直接给出网站的前端 HTML 代码的功能已经可以使用。另外的例子是，给出图片， ChatGPT 可以帮助我们理解图片，给出实时指导。比如拍一张山地自行车的照片，跟 ChatGPT 询问如何调整自行车的车座高低，ChatGPT 能够根据车的具体情况，给出调整的方法。 用户可以上传一张肺部的CT照片，ChatGPT 分析出来得出里面是否有肿瘤的结论。
GPT 4 和 GPT 3.5 的直观差别
GPT 4 跟 GPT 3.5 差异有哪些？因为图片（包括其他多媒体比如语音、视频的能力）没有开放，所以一开始给我的感觉是，GPT 4 带来的变化只是一次对话（发送消息+返回消息）比 GPT 3.5 能处理的文字长度多了1倍， 从 4096 个 Token 变成了 8192 个，相当于中文的字符数从 2000 个变成了4000，而英文能处理的单词数大约 从 3000 变成了 6000。Token 的计算规则，一个汉字大约2个 token，一个英文单词大约 0.75 token。
另外一个是 GPT 4 的价格高了很多倍(30-60倍)，而 GPT-4-turbo 因为支持了 128k 的上下文，费用更高，每千 token 费用降低一半，但 token 数变成了 16 倍，所以 GPT-4-turbo 的最高费用是 GPT-4 的8倍。OpenAPI 的调用， 按照汇率7折算，
GPT 3.5 的费用是 0.002 $ 1k Token，最大支持4k Token，一次对话的费用：
0.002 * 7 * 4 ￥ = 0.056 ￥ 元
GPT 4 的费用是 0.06 $ 1k Token，最大支持8K Token，一次对话的费用：
0.06 * 7 * 8 ￥ = 3.36 ￥ 元
GPT 4 的价格让人惊讶 OpenAI 在抢钱，3.5 只需6分钱一次，4 需要 3.3 元，一次对话最大相差 60 倍的价格。（当然，如果你是 ChatGPT Plus 的20$ 一个月的付费订阅用户，Web 界面上使用 GPT 4，除了 20$ 一个月的订阅费用，是不用另外付费每次对话的。这里表格中的是 API 调用的费用。）
|
模型
|
输出1000 Token费用
|
最大Token数 |
一次对话最多费用
|
|
GPT 3.5 Turbo
|
0.002 $
|
4096
|
0.008 $
|
|
GPT 4.0
|
0.060 $
|
8192
|
0.480 $
|
|
GPT-4-Turbo
|
0.030 $
|
128k
|
3.84 $
|
随着 GPT-4-turbo （目前是 gpt-4-1106-preview 的模型）的发布，gpt能够支持 128k 的长文本，可以一次喂给 GPT 一整本史记的内容。如果输出是大头，gpt-4-turbo 的一次最大费用是 gpt-3.5-turbo 的 480 倍。如果输入占大头，那么输入 128 k 的长文本，整体费用大约是 1.28 $，大约是 GPT 3.5 对话的 160 倍。
这里的费用按照输出的费用来算，通常的情况下输入的内容比较短，输出的比较长。
对于另外输入是大段的内容比如输入整本书的内容的情况下，输入的费用要低，比如对于 GPT-4-Turbo 输入的 Token费用是 0.01$/1k token，只有输出的 1/3。
这里是官网的价格说明： https://openai.com/pricing
GPT 4 和 GPT 3.5 使用过程中的实际差别
最近在使用不同的提示语做对话，发现部分大家推荐的提示语（提示词）在 GPT 3.5 的时候，没有得到很好的答复。于是对比了一下 GPT 4，结果发现大量场景下， GPT 4 和 GPT 3.5 的差别很大。 在 GPT 3.5 回答不好的时候， GPT 4 往往能够给出很好的回答。下面来看一些例子。
1.广告投放
提示语： 我想让你做广告投放的营销经理。你是一个出色的营销经理。即将就一个在线招聘产品进行广告投放，你将在哪些渠道，如何投放广告？预计成本，投放周期如何？请详细分析
同样的关于广告投放的问题， GPT 4 的回答比 GPT 3.5 的回答更加详细具体。
2.市场分析 - 合作商场品牌分析
商业合作的商场品牌分析，GPT 3.5 整体是拒绝的， GPT 4 给出了很好的回复。
3.表格分析（对比 GPT 3.5 和 GPT 4 的差异）
我把 OpenAI 官网关于 ChatGPT 4 的介绍粘贴到对话的内容里面， 分别让 GPT 3.5, GPT 4 做出差异分析的表格。下面是对话的发出内容。
GPT 4 的说明：GPT-4 vs GPT-3: A Brief HistoryBefore we dive into the exciting possibilities of GPT-4, let's take a quick trip down memory lane and see how it compares to its predecessor, GPT-3.Model performance on various tasks | GPT-1 paper - GPT-4 vs. GPT-3Model performance on various tasks | GPT-2 paper - GPT-4 vs. GPT-3GPT-1 and GPT-2 were impressive language models in their own right, but they had limitations regarding their dataset and capabilities. With GPT-3, the game changed by having a massive dataset and the ability to generate human-like responses that were often difficult to distinguish from those of a human.Now, the AI world is buzzing with anticipation for GPT-4. While the specifics of the model will be shared in the article, to give you the gist, it's more capable than GPT-3. It has extensive datasets to generate more detailed and personalized responses than before.The excitement around GPT-4 is electrifying, with many eagerly anticipating its release. It's poised to be the next big breakthrough in AI language models, with immense implications for businesses and industries.Along with GPT-4, ChatGPT has launched ChatGPT Plugins that can overcome the limitations of GPT-4 like generating outdated responses.But before we rush to the model specifics, let's look at the critical differences between GPT-4 and its predecessor — GPT-3.GPT-4 vs GPT-3: Key DifferencesLet's dive into the exciting world of GPT-4 and its critical differences compared to GPT-3! It's like comparing a superhero to their sidekick - both are impressive, but one always steals the spotlight.With the release of GPT-4, businesses anticipate the impact of this advanced language model on their operations and here's why GPT-4 is making such big headlines.GPT-4 Has 100 Trillion Parameters: Compared to its predecessor, GPT-3, GPT-4 boasts a substantially larger-sized database, faster speed, and improved performance across a range of benchmarks. GPT-4 has 100 trillion parameters, compared to GPT -3's 175 billion parameters, making it substantially more prominent and influential.This increased parameter size will enable businesses to generate more accurate and coherent text. GPT-4 can generate more accurate financial reports, forecasts, and investment strategies based on vast amounts of market data.Improved Factual Performance: GPT-4 improves GPT -3's factual correctness, scoring 40% higher on OpenAI's internal factual performance benchmark and producing fewer "hallucinations" or factual errors. This improved solid performance is critical for businesses that need to generate accurate and reliable content.GPT-4 - GPT-4 vs. GPT-3GPT-4By using GPT-4, businesses can avoid costly mistakes that could damage their reputation. For example, GPT-4 can generate more accurate and efficient product descriptions, news articles, or legal documents.For example, a law firm can use GPT-4 to analyze and summarise legal cases with greater accuracy and efficiency, reducing the risk of errors and improving the quality of legal advice. In fact, Legal AI company Casetext has the world's first GPT-4-powered AI legal assistant, which claims to be the first AI to pass the bar exam.Enhanced Steerability: GPT-4 is more "steerable," allowing users to command it to write in different styles, tones, or voices. This enhanced steerability significantly improved over GPT-3, which could generate text in different types but required substantial retraining.Businesses can use GPT-4 to customize the language model to their specific needs.Let's say you own a marketing agency, you can use GPT-4 to create customized marketing campaigns that resonate with customer segments based on their preferences, interests, and behaviors.Image Inputs Capability: GPT-4 can use image inputs, enabling users to specify any vision or language task by entering text and images. This will benefit businesses dealing with visual or multimedia content, such as advertising, entertainment, and e-commerce.GPT-4 - GPT-4 vs. GPT-3GPT-4An online retailer can use GPT-4 to generate product descriptions and reviews based on images and videos, making it easier for customers to understand and compare products.Human-Level Performance on Various Benchmarks: GPT-4 achieved human-level performance on various professional and academic benchmarks, including the Uniform Bar Examination, LSAT, and SAT.GPT-4 - GPT-4 vs. GPT-3GPT-4This will benefit businesses that require high levels of language skills and critical thinking, such as education, legal, and consulting industries. In fact, the popular learning platform, Khan Academy, is already working to integrate it.For example, a consulting firm can use GPT-4 to analyze and solve complex business problems, such as market segmentation, pricing strategy, and risk management.Outperformance on Multiple Benchmarks: GPT-4 outperformed existing large language models on traditional machine learning benchmarks, as well as in multiple-choice questions in 57 subjects, commonsense reasoning, and grade-school science questions.GPT-4 - GPT-4 vs. GPT-3GPT-4This will benefit businesses that require high levels of accuracy and efficiency in language processing and understanding, such as customer service, chatbot, and virtual assistant industries.For example, a customer service center can use GPT-4 to provide faster and more accurate responses to customer inquiries based on natural language understanding and context awareness. Popular payment infrastructure platform, Stripe, is already working on it. And guess what? Intercom is working on it too.Multilingual Capability: Brands like Duolingo have already adapted GPT-4 to teach multiple languages. Additionally, GPT-4 outperformed GPT-3.5 and other large language models in 24 out of 26 languages tested, based on the MMLU benchmark. This will benefit businesses that operate in multilingual environments or international target markets, such as tourism, hospitality, and e-commerce.For example, a hotel chain can use GPT-4 to provide multilingual customer service and support based on accurate and context-sensitive translations of customer inquiries and requests.As you can see, the differences between GPT-4 and GPT-3 are substantial, and the implications for businesses and industries are enormous. The advancements in speed, adaptability, and language skills of GPT-4 will enable enterprises to improve their efficiency, productivity, and customer experience.Therefore, it's no surprise that companies are eagerly anticipating the release of GPT-4 and strategizing how they can integrate it into their business operations.Let's look at how GPT-4 and its advanced capabilities can transform businesses' operations.GPT 4 跟 GPT 3.5 有什么差别，请用表格列出主要差别。另外，我需要用 GPT 4 吗？
4.图片的理解
GPT-4-vision 可以理解图片。 如果使用的 ChatGPT 官网的 GPT-4 服务，里面集成了 gpt-4-vision，可以自动理解输入的图片。
5.画图，Code Interpreter 和联网功能
在官网的 Web 网站上， GPT-4 中集成有 Code Interpreter , 浏览网页的功能，以及 DALL-E。
可以要求 GPT 从网络上获取最新信息。可以让 GPT-4 调用 Dall-e 3 画图，Dall-e 3 目前是两个画图最好的 AI产品之一，另外一个是 Midjourney。可以上传数据让 GPT 来做分析，或者上传代码让 GPT 来分析，这个时候 GPT 可以使用 Code Interpreter 来做数据分析，代码分析。
GPT-3.5 和 GPT-4 有什么区别？
主要差别包括：
- 模型大小：GPT4 比 GPT3.5 有更多的参数，所以 GPT4 的模型规模更大，这意味着 GPT4 可以处理更多的数据，生成更长、更复杂、更连贯、更准确、更多样化和更有创造力的文本；
- 处理复杂问题的能力：在处理更复杂的问题时，GPT-4 比 GPT-3.5 更可靠，更有创造性，能够处理更细微的指令，例如 GPT-4 在完成的各种考试时能够取得更好的结果（SAT、LSAT、GRE 等）；
- 模型输入：GPT4 是一个多模态（multimodal）模型，支持接受图像和文本输入，输出文本。
但是对于大多数用户来说，正常使用过程中很难区分 GPT3.5 和 GPT4，并且 GPT4 的反应速度很慢，还有使用限制（3 小时 50 条），小初有时候更喜欢用 GPT3.5。
怎么区分？
虽然 GPT4 已经出来一段时间了,但真正能够使用 GPT4 API 的平台并不多,因为 Anthropic 目前只向部分客户开放了 GPT4 API。因此,当前市面上大多数自称支持 GPT4 的对话平台,实际上都是使用 GPT3.5 进行回复。
那么如何区分 GPT3.5 和 GPT4 呢?主要从以下几个方面:
- GPT4 的回复更加逻辑清晰,内容更加具体和贴近问题。它能够进行更复杂的逻辑推理和数学运算。
- GPT4 更好地处理长段对话的上下文和语义联系,不会轻易失去语义连贯性。
- GPT4 对常见知识的理解更加广泛和准确,不会出现明显的知识错误。
- GPT4 的语言表达能力更加丰富,词汇量更大,语法结构更复杂。
- GPT4 更加谨慎,对不确定的问题会表示无法回答或需进一步确认。
- GPT4 可以承认自己的局限,不会强行胡说八道。
- GPT4 更加注重道德和价值判断,不会轻易生成有害或不当内容。
或者如果你的plus会员可以从这里直观的看出
