Preface

How to start your own AI journey? I believe this is the first question that confuses many people (especially those who only know the concept of AI but have never actually interacted with AI technology). A few weeks ago, I was the same, having only heard of the power of AI but not having delved very deep into its practical applications.

Before this, my experience with AI was limited to OpenAI's ChatGPT. Like many people, I enjoyed the free services provided on OpenAI's official website: either through the web version or by using the official app client on Mac, iPhone, or iPad. Because my reliance on ChatGPT was not strong, the daily free quota was basically enough for me (if I exceeded it, I would just stop using it for a few hours and try again). However, this ”freeloading” approach is not a long-term solution. Since AI technology is becoming increasingly important in various fields, I also need to gain a more formal and deeper understanding of AI applications, rather than just relying on a single platform or tool.

In the beginning, I had absolutely no idea how to start learning AI. Especially when faced with different large language models, a wide variety of API providers, and questions like how to integrate AI technology into actual work and life scenarios, I was completely lost (I didn't even know what many terms meant, and I often couldn't understand what friends in the chat group were talking about~~).

In the end, I could only start with the most basic concepts, practically deploy Large Language Model UI applications (LobeChat local version, server-side database version, ChatGPT Next Web, etc.), compare the advantageous areas of various API providers (OpenAI, Azure OpenAI, Claude, etc.), and finally, basically sort out this entry-level AI knowledge and write this article.

Through this article, you can learn how to get started with AI in a simple and practical way, master the basic knowledge of Large Language Models, get to know the commonly used API providers on the market, and even build your own AI Chat application environment. Whether out of personal interest or to leverage AI in your professional field, this article can provide a clear introductory guide.

Usage of Chat-type Large Language Models

Usage of ChatGPT

As I mentioned earlier, I have also been using the official free service provided by ChatGPT, which includes the official web version (https://chatgpt.com)：

In fact, both the official web version accessed through the official URL and the local Mac app version are the ChatGPT UI interfaces I use. However, the UI backend corresponding to the free version is bound to the default GPT-4-turbo model version provided by OpenAI (this model is an optimized, lower-cost version that allows some free users to access the GPT-4o model within a certain quota). Once these quotas are used up, it will switch back to using the “GPT-3.5-turbo” model.

Note 1: The Mac version of ChatGPT can be downloaded directly (ChatGPT Mac version official download address), and there are also ChatGPT apps on iOS and iPad, but you need to switch to an overseas Apple ID to download them, and normal use requires a VPN or proxy.

Note 2: Using ChatGPT is very simple now. You only need to log in using an existing Google account, Microsoft account, or Apple account. Of course, you can also register an OpenAI account directly.

Usage of Other Chat-type Large Language Models

In addition to OpenAI's ChatGPT, we also have other similar Large Language Models to choose from. However, they may not provide an official web version or a ready-made app as an access UI like ChatGPT does; they basically only provide API access. Therefore, to use the services of Large Language Model providers other than OpenAI, you need a universal Large Language Model UI that supports various provider APIs as a tool to access the APIs.

Generally speaking, normal access to Large Language Models requires two functional components:

• Large Language Model UI

Essentially, it is a well-designed interactive interface whose main purpose is to build a bridge connecting ordinary users with powerful but complex AI technology. Imagine having a friendly control panel in front of you, and behind this panel is a massive AI system provided by various API providers. The beauty of the UI is that it hides all the technical complexity, providing you with a simple and intuitive way to operate. You don't need to know programming, nor do you need to care about how the API behind it works; you just need to input text like in a daily chat. The UI will cleverly translate your needs into instructions that the API can understand, and then convert the results returned by the API into a format that is easy for you to comprehend. It acts like a translator, passing information back and forth between you and the AI system. In addition, this interface helps you manage conversation history, save important information, and even allows you to adjust some settings to personalize your AI assistant. In short, the Large Language Model UI is a key tool that makes complex AI technology accessible, enabling everyone to easily leverage the power of AI without needing to understand the technical details behind it.

• API Providers

Behind the Large Language Model UI, API providers play a crucial role, acting as the ”brain” and ”engine” of this system. Imagine that these API providers offer a series of powerful toolboxes, each specialized in completing specific tasks. Some toolboxes excel at turning your ideas into lifelike images, as if an invisible painter is on standby at all times; others can ”understand” pictures and describe the content for you, like a careful observer interpreting visual information. Still other toolboxes can convert your voice into text, or vice versa, convert text into natural and fluent speech, as if a 24/7 secretary is helping you record and read aloud. These APIs are like a collection of various superpowers; they can understand natural language, answer complex questions, and even help you write code or compose articles. API providers constantly update and optimize these tools to make them smarter and more efficient. Through these diverse APIs, the UI can provide users with almost omnipotent services, easily handling everything from daily conversations to complex creative work and professional analytical tasks. In short, API providers are the technical teams working silently behind the scenes, providing a continuous stream of intelligent power to the UI, allowing users to obtain various amazing AI capabilities through a simple interface.

Local Large Language Model UIs

Large Language Model UIs Suitable for Personal Use

There are many choices for Large Language Model UIs on the market, but today I will only talk about the ones I think are suitable for conventional personal scenarios. Generally, there are 3 choices, which areLobeChat、ChatGPT Next Web、Chatbot-UI。

Note: RegardingChatbot-UI, I looked into it and the deployment is quite troublesome (unlike LobeChat and ChatGPT Next Web, which can be done with just a single docker run command), and it doesn't have any obvious advantages, so I won't recommend it. However, I will still keep the subsequent description, and those who are interested can research it on their own.

LobeChat

Introduction to Lobechat UI

LobeChat is a feature-rich open-source local Large Language Model user interface designed for ease of use and flexibility. The deployment difficulty is relatively low, making it suitable for users with a basic technical background. It supports Docker containerized deployment, simplifying the installation process so that even non-professional developers can quickly build their own AI assistants. In terms of extensibility, LobeChat provides a plugin system that allows users to add new features as needed. It supports multiple API providers, including OpenAI, Anthropic Claude, Azure OpenAI, etc., and is also compatible with open-source models such as llama.cpp and ChatGLM. LobeChat's interface design is clean and intuitive, supports multiple languages, and provides rich conversation management features, such as conversation export and history search. For users who value data privacy and want complete control over their AI interaction process, LobeChat is an ideal choice. Its access interface is as follows:

Lobechat Deployment Methods

LobeChat supports 2 deployment modes, divided into client-side database mode and server-side database mode based on where user data is stored.

• LobeChat Client-side Database Mode

Deploying LobeChat in this mode stores all data (such as user conversation history, model configurations, etc.) in the user's local browser cache or client-side database. This mode does not rely on a backend database server, making it suitable for individual users or small projects, with simple deployment and no need for extra server resources. User data is completely controlled locally, offering good privacy. However, because it is stored on the client side, the data is susceptible to browser cache clearing or device replacement, posing a certain risk of data loss. In addition, the client-side mode is not suitable for scenarios that require data synchronization across multiple devices. For example, if most access to LobeChat is from the same PC client, there is no need to sync data between multiple clients (no need to share LobeChat conversation data between the PC client and the mobile client).

The deployment command for LobeChat client-side database mode can refer to the following:

docker run --name lobe-chat  -d --restart=always \
  -p 3210:3210 \
  -e ACCESS_CODE=xxx \
  lobehub/lobe-chat

• LobeChat Server-side Database Mode

Deploying LobeChat in this mode stores user conversation data, configurations, etc., in a database on a remote server, which is typically achieved through a hosted database service. This mode is suitable for multi-user or team use, supporting data synchronization across multiple devices and centralized management. It enables persistent storage, ensuring long-term data security and integrity, but the deployment is relatively complex.

The reason why the deployment of LobeChat server-side database mode is complex is that, in addition to the deployment of LobeChat-database itself, it also involves components such as PostgreSQL database, MinIO COS (object storage), and Logto (authentication). Among these, MinIO COS and Logto can be replaced by other third-party services.

Since the deployment is quite complex, I won't go into detail here. Interested friends can refer to the article:Docker Series: Detailed Deployment Tutorial Based on the Open Source Large Language Model UI Framework: LobeChat。

ChatGPT Next Web

Introduction to ChatGPT Next Web UI

ChatGPT Next Web is a highly customizable large language model user interface, renowned for its powerful features and flexible deployment options. The deployment difficulty is moderate, requiring some technical knowledge, but it provides detailed documentation and community support. It supports multiple deployment methods, including Vercel one-click deployment, Docker containerization, and traditional server deployment. In terms of extensibility, ChatGPT Next Web adopts a modular design, allowing developers to easily add new features or modify existing ones. It primarily supports OpenAI's API, but through its flexible architecture, it can also integrate other API providers. Notably, it provides an API proxy function, which can solve access restriction issues in certain regions. Special features of ChatGPT Next Web include multi-user support, custom prompt libraries, exporting conversations as Markdown or images, etc. Its interface design is modern and responsive, supporting dark mode and multi-language interfaces, making it suitable for individual users and small teams who need a highly customized AI assistant. Its access interface is as follows:

ChatGPT Next Web Deployment

ChatGPT Next Web currently only supports the client-side database mode, and its deployment is relatively simple. The deployment commands can be referred to as follows:

docker run -d -p 3000:3000 \
   -e OPENAI_API_KEY=sk-xxxx \
   -e CODE=your-password \
   -e PROXY_URL=http://localhost:7890 \
   yidadaa/chatgpt-next-web

Note 1:-e OPENAI_API_KEY=sk-xxxx \-e CODE=your-password \

-e PROXY_URL=http://localhost:7890 \
yidadaa/chatgpt-next-web

can specify a proxy server (if needed).

Chatbot UI:

Note 2: Although ChatGPT Next Web does not support server-side database deployment, it provides other ways to indirectly achieve data synchronization: WebDAV and UpStash.

WebDAV:

UpStash:I haven't tested the actual performance in detail, as I usually use Lobechat.、Chatbot UI is a local large language model interface focused on providing an exceptional user experience. Its deployment difficulty is relatively low (that being said, it is not as low as the client-side database deployment of Lobechat and ChatGPT Next Web), providing detailed installation guides and multiple deployment options, including local and cloud deployment. In terms of extensibility, Chatbot UI adopts a modular architecture, allowing developers to add new features through a plugin system. It supports multiple API providers, including OpenAI, Anthropic, Cohere, etc., while also supporting the integration of open-source models through Hugging Face. A major feature of Chatbot UI is its powerful conversation management capabilities, including conversation categorization, tagging systems, and advanced search functions. It also provides detailed usage analysis tools to help users optimize their AI interactions. In terms of interface design, Chatbot UI offers various themes and layout options, allowing users to customize according to their personal preferences. In addition, it supports voice input and text-to-speech features, greatly enhancing accessibility. Chatbot UI is particularly suitable for professional users and researchers who need in-depth analysis and management of AI conversations. Its access interface is as follows:、Note 1: For the reasons mentioned earlier, the deployment section is skipped, and I do not recommend Chatbot UI for beginners.、Note 2: In addition to the above three local large language model UIs, there are actually many others, such as:、GPT4All Web UI、Oobabooga Text Generation Web UIStableLM Web UI

Comparison of API Providers Supported by Default in the 2 UIs

Vicuna Web UI

LobeChat

Langchain UI

Hugging Face Inference API UI

Note: The ability to ”customize API address and model version” is very good. Everyone should mark this, as I will mention it later.

ChatGPT Next Web

Compared to LobeChat, ChatGPT Next Web supports fewer built-in API providers by default, currently only supporting 10, though it does support commonly used ones like OpenAI, Azure OpenAI, Google, and Anthropic:

As can be seen from the images above, by default, the language models and model versions in ChatGPT Next Web are fixed, unlike LobeChat where they can be customized, resulting in less flexibility. However, you can support more API providers and model versions by customizing the API address through the ”Custom Endpoint” option:

“Custom Endpoint” option allows users to configure custom API endpoints (or model versions), meaning you can connect to other model providers or locally running models besides the 10 default supported API providers: After enabling “Custom Endpoint”, you can use the ”OpenAI Endpoint” option to connect to any third-party large language model provider's API address that complies with API call specifications. As long as the interfaces of these services are compatible with OpenAI's API, you can connect ChatGPT Next Web to these models by entering a custom URL. This option somewhat makes up for the limited number of default built-in API providers, though this support is certainly less compatible compared to LobeChat's direct support.

Note: This option is the ”customize API address and model version” feature provided by ChatGPT Next Web.

Introduction to Common Large Language Model API Providers

FBI warning: Boring Alert

This part is just for recording purposes and is relatively dry. Friends who are not interested can skip directly to the final summary section. I organized it in detail for easy reference in the future.

Prerequisites

Transformer Architecture

The Transformer architecture is a deep learning model architecture specifically designed for Natural Language Processing (NLP) tasks, proposed by Vaswani et al. in 2017. It completely revolutionized research and applications in the NLP field, becoming the foundation for many modern language models, including the GPT series, BERT, T5, etc. It includes the following core concepts.

1. Self-Attention Mechanism

The core of the Transformer is the Self-Attention mechanism, which allows the model to focus on all words in a sentence when processing it, rather than relying on fixed sequential processing (such as the sequential structure of RNNs). Each word can understand other words in the sentence through the self-attention mechanism, thereby capturing the relationships between words.

2. Encoder-Decoder Structure

The original design of the Transformer consists of two parts:

• Encoder: Processes the input sequence to generate context-dependent representations.
• Decoder: Gradually generates the output sequence based on the encoder's output and previous generation results.

However, in many language models (such as GPT), we only use a part of the encoder or decoder. For example, GPT only uses the decoder structure of the Transformer.

3. Parallel Computing

The Transformer uses a completely attention-based mechanism, abandoning the step-by-step calculation pattern found in Recurrent Neural Networks (RNNs) and Long Short-Term Memory networks (LSTMs). Its parallel computing capability allows the model to efficiently process long text inputs, significantly improving training speed.

4. Multi-Head Attention

The Transformer uses a multi-head attention mechanism, enabling the model to understand data from different “angles.” Through multiple independent attention heads, the model can capture relationships between words at different semantic levels, further enhancing the model's expressiveness.

5. Positional Encoding

Since the Transformer does not rely on the order of input like RNNs or LSTMs, it introduces positional encoding to represent the relative positions of words in a sentence. This allows the model to retain the sequential information of the sentence.

Applications and Advantages of the Transformer Architecture

Strong Scalability: Compared to previous RNN and LSTM models, the Transformer can better handle long sequence data, performing exceptionally well in long sentence or document comprehension.

Highly Efficient Parallelization: Due to the self-attention mechanism and sequence-independent processing, the Transformer can process the entire input sequence in parallel, greatly accelerating training speed, especially on large-scale data.

Wide Application：

• GPT Series: OpenAI developed the GPT series of models (such as GPT-3, GPT-4) based on the Transformer architecture, which are applied to tasks like text generation and dialogue systems.

• BERT: Google's BERT is also based on the Transformer, excelling in tasks that handle bidirectional contextual relationships, such as reading comprehension and text classification.

• T5: The Transformer-based Text-to-Text Transfer Transformer (T5) converts all NLP tasks into text generation problems, demonstrating the high flexibility of the Transformer architecture.

With its self-attention mechanism, parallel computing capability, and wide range of applications, the Transformer architecture has become the core architecture in the NLP field. It has driven the development of large language models and achieved great success in multiple tasks, making modern AI systems smarter and more efficient when processing language.

GPT

GPT (Generative Pre-trained Transformer) is a large language model based on the Transformer architecture, developed by OpenAI. It is capable of generating human-like text and performs exceptionally well in many natural language processing tasks. The core of GPT lies in two stages: pre-training and fine-tuning, learning from large-scale text data and optimizing through a small amount of task-specific data. Since the release of its earliest version, the GPT series of models has evolved through multiple versions, including GPT-2, GPT-3, GPT-4, etc.

The core principles of GPT are as follows:

1. Based on the Transformer Architecture

GPT is based on the decoder portion of the Transformer, which means it can understand context and generate new text through the self-attention mechanism. The Transformer architecture allows GPT to capture complex dependencies between words, performing exceptionally well when processing long texts.

2. Pre-training and Fine-tuning

• Pre-training: GPT first undergoes pre-training on massive amounts of unlabeled text data. The model learns syntactic, semantic, and other rules in language by predicting the next word.

• Fine-tuning: On specific tasks (such as text generation, translation, or question answering), GPT can be fine-tuned using a small amount of labeled data to make its performance better align with specific task requirements.

3. Autoregressive Generation

GPT is an autoregressive model that generates text step-by-step, predicting the next word based on the already generated portion until the complete text is output. This makes GPT highly suitable for generating fluent and coherent natural language text.

Main Functions of GPT:

1. Natural Language Generation

GPT excels at generating text that is highly consistent with the context and has strong coherence. It can be used for tasks such as automated writing, content creation, screenwriting, and generating product descriptions.

2. Dialogue and Chatbots

GPT can understand user input and generate appropriate responses, becoming the core technology for many dialogue systems and customer service chatbots.

3. Q&A and Information Extraction

GPT can answer questions or extract relevant information based on a given text or knowledge base, and is widely used in scenarios such as Q&A systems and search engines.

4. Zero-Shot Learning and Few-Shot Learning

GPT-3 and GPT-4 demonstrate zero-shot and few-shot learning capabilities, enabling them to complete tasks through prompts even without being specifically fine-tuned for a particular task.

5. Multi-Domain Applications：

GPT is capable of processing tasks across multiple domains, including but not limited to programming code generation, medical text understanding, legal document analysis, and translation. GPT can generate creative text and even write articles, poetry, stories, etc., demonstrating human-like creativity.

Evolution of Models and Versions

“Model” is a universal concept in the field of large language models, involving different organizations and frameworks. In addition to OpenAI's famous GPT-3 and GPT-4 models, other organizations are also developing and releasing their own models while continuously iterating and upgrading versions:

• Google's BERT, from the original BERT to subsequent improved versions (such as ALBERT, RoBERTa); Gemini from version 1.0 to the current 1.5, with 2.0 also about to be released.
• Facebook AI Research (FAIR)'s RoBERTa, which underwent larger-scale training and optimization compared to the original BERT.
• Hugging Face Maintains the Transformers library, supporting multiple models and versions, including BERT, GPT, T5, XLNet, etc.
• Microsoft The launched DeBERTa, demonstrating version updates and performance improvements.

In the article below, I list the most common large language model API providers on the market. Since there are too many, I only list the API providers I find most commonly used (excluding domestic API providers in China) and their main model versions.

OpenAI – ChatGPT

Introduction to OpenAI ChatGPT

What is OpenAI ChatGPT?

ChatGPT is a large language model product based on the GPT series launched by OpenAI, used to generate natural language dialogue. With its powerful language understanding and generation capabilities, ChatGPT can engage in multi-turn dialogues with users, answer questions, generate creative text, assist in writing code, etc., and is widely used in scenarios such as chatbots, customer service systems, and content generation.

Core features of ChatGPT:

Based on the GPT model: ChatGPT is based on the GPT (Generative Pre-trained Transformer) series of models, especially the GPT-3.5 and GPT-4 versions. GPT is a language model that uses the Transformer architecture for large-scale pre-training and fine-tuning, excelling at learning patterns from context and generating coherent text.

Dialogue capability: ChatGPT possesses powerful dialogue generation capabilities, able to understand and respond to a variety of user requests, whether simple chats or complex question answering. It can conduct multi-turn dialogues, remember context, and maintain coherence.

Generating multi-language content: ChatGPT supports multiple languages; users can communicate with ChatGPT in English, Chinese, French, Spanish, and other languages. It can not only generate text but also create poetry, stories, or write code based on prompts.

Efficient learning capability: ChatGPT uses Few-Shot Learning and Zero-Shot Learning technologies. Users only need to provide a few examples, or even no examples at all, and ChatGPT can understand the task and generate accurate text content.

Adaptability: ChatGPT can provide assistance in multiple fields, from technical support to copywriting to academic discussions, demonstrating extremely high adaptability. This allows it to be applied in various industries, including education, entertainment, business, etc.

Advantages of ChatGPT:

• Multi-tasking: ChatGPT can easily handle multiple tasks, such as answering questions, generating dialogue, writing code, language translation, etc., possessing strong versatility.

• Fast response speed: Through optimized versions like GPT-4 Turbo, ChatGPT can provide high-quality text generation in a relatively short time.

• Wide application: Whether for commercial applications or personal use, ChatGPT can adapt to various scenarios, allowing users to quickly integrate the API or deploy using ready-made applications.

Main Model Versions and Their Features

The OpenAI API provides several different versions of language models, with each model varying in performance, purpose, and cost:

GPT-4 Series Models

• GPT-4 Turbo：

This is an optimized version of GPT-4, offering faster response speeds and higher performance while reducing usage costs. GPT-4 Turbo is suitable for scenarios that require efficient generation of large amounts of text content.

• GPT-4 (gpt-4o)：

This is OpenAI's latest and most powerful model. It has outstanding language understanding and generation capabilities, able to handle complex text tasks: whether generating long articles or handling contextual dialogue, GPT-4 performs exceptionally well, making it suitable for tasks like high-precision content generation, complex dialogue systems, and technical document writing. It can understand long-term context and generate answers highly relevant to previous dialogue content, suitable for use cases such as news media, academic research, technical writing, and customer support.

• GPT-4-mini (gpt-4o-mini)：

This is a streamlined version of GPT-4. Although its performance is slightly lower than the standard GPT-4, it still possesses powerful text processing capabilities. Compared to GPT-4, its computational requirements and usage costs are lower, making it suitable for scenarios that need to balance performance and cost. It is applicable to tasks of medium complexity, such as chatbots and automated customer service, where it has greater advantages in speed and cost. Suitable for customer service in small and medium-sized enterprises, content moderation, daily text generation, etc.

GPT-3.5 Series Models

• GPT-3.5 (davinci)：

GPT-3.5 is the predecessor to GPT-4. Although slightly inferior to GPT-4 in performance, it is still capable of handling most text generation tasks. It achieves a good balance between the quality and speed of generated content. It is suitable for a variety of tasks, such as content generation, dialogue systems, text translation, and code generation. Its ability to handle complex tasks is second only to GPT-4. It is suitable for companies requiring automated generation or interaction with customers, such as news platforms, online education, and technical support.

• GPT-3.5-mini (curie)：

This is a lightweight version capable of handling most daily text tasks. Compared to davinci, it falls slightly short when processing complex content, but is more suitable for scenarios that require efficient processing of large volumes of simple tasks. It is highly suitable for tasks such as basic chatbots, text summarization, and simple content creation. It can quickly generate short documents or simple answers. Suitable for e-commerce, content recommendation systems, automated email replies, etc.

OpenAI o1 Series Models

OpenAI O1 is a relatively optimized version in the GPT model family, specifically designed to provide more efficient computation and more powerful generation capabilities in different scenarios. While retaining the core capabilities of large language models, the O1 model provides high efficiency and accuracy through optimized architecture and computational performance, making it suitable for a variety of tasks such as dialogue generation, text analysis, and content creation.

Main features of the O1 model:

Optimized performance: O1 is optimized for higher performance compared to traditional GPT models. It can handle complex language generation tasks, such as writing code, long-form text generation, and complex Q&A tasks, while maintaining efficient execution speeds.

Wide Range of Application Fields: The O1 model can be applied to multiple fields, including business services, technical support, content generation, natural language processing tasks, etc. It can adapt to more complex context requirements and provide accurate and coherent outputs in dialogue and generation tasks.

High-Quality Generation: Compared to more lightweight models, O1 provides higher generation quality, especially in tasks requiring context understanding and complex reasoning, where it can output text that better aligns with human language logic.

Advantages of the O1 model:

• Balanced Computational Cost and Performance: The O1 model finds a good balance between computational resource usage and generation quality, providing sufficient reasoning capabilities without consuming excessive computational resources like larger models (such as GPT-4).

• Suitable for Multitasking: O1 is capable of handling tasks ranging from simple text generation to more complex dialogues and code writing. Whether generating creative copy, processing technical documents, or responding to real-time dialogue needs, O1 performs excellently.

• Optimized Response Speed: Through architectural optimization, O1 significantly reduces response latency while providing high-quality output, making it particularly suitable for application scenarios that require fast generation and feedback.

Differences between O1 and O1-mini:

• Model Scale: Compared to O1-mini, O1 has more parameters, and therefore performs better in handling complex tasks and providing precise generation.

• Generation Capability: The text generation quality of O1 is generally higher than that of O1-mini, especially in context understanding and complex task processing. O1-mini is more suitable for lightweight tasks, while O1 is better suited for tasks requiring higher accuracy and more complex reasoning.

• Computational Cost: Although O1 consumes more computational resources than O1-mini, its performance in high-performance tasks is also superior, making it suitable for users with higher requirements for generation quality.

Overall, OpenAI O1 is an efficient and powerful model version, suitable for application scenarios that seek to maintain balanced, precise, and fast generation across a variety of natural language tasks.

Additional Knowledge: Knowledge Cutoff (Training Cutoff Date)

On OpenAI's official pricing page, you can see that when different models are introduced, they all have a ”knowledge cutoff”, which is currently October 2023:

1. Limitations of Model Training Data：

Models like GPT do not receive information in real-time; instead, they generate based on a large amount of pre-collected and processed training data. Every once in a while, OpenAI trains the model with a new dataset (a single training run takes several months). However, once the model goes online, it does not actively update until the next version finishes training. Therefore, at any given point in time, the model's knowledge only includes data from its last training. This is because updating a model is not as simple as just adding new data; after a model update, a significant amount of debugging, optimization, and validation is required to ensure the accuracy, rationality, and controllability of its generated content. If new data is introduced too frequently, the model's quality may become unstable. Therefore, model updates often need to be performed periodically rather than in real-time.

2. Maintaining Transparency：

To let users understand the model's knowledge scope, OpenAI will specify the model's knowledge cutoff(knowledge cutoff) in the documentation and instructions. This way, users can clearly know that when the model answers questions, it will not include the latest information after October 2023,such as new technologies, news events, policy changes, etc. Another reason is that current GPT models are trained offline, and they cannot obtain or learn new information from the internet in real-time. These models need to acquire knowledge from static pre-trained datasets, so they cannot access and process the latest news, research, or dynamic information like a search engine.

3. Avoiding Misleading Users：

This prompt is also to prevent users from misunderstanding and thinking that the model can provide real-time information or respond to the latest situations. For example, when a user asks about the latest events in 2024, the model may not be able to answer accurately, and OpenAI reminds users of this through this prompt.

4. **Model Version Stability:

When providing APIs and commercial services to the public, OpenAI typically uses stable versions that have undergone rigorous validation. After a model is released, its knowledge may become outdated, but prioritizing stable performance across multiple tasks and ensuring safety and reliability take precedence over the ability to update data in real-time.

Therefore, although the knowledge cutoff date may seem “outdated,” this is a very common practice in the current field of large language models.

Companies like OpenAI may shorten this lag time in the future through technological advancements (such as enhanced real-time update mechanisms), but currently, this “lag” remains a relatively common phenomenon.

Note: There is also a model naming convention related to time, such as ”gpt-4o-2024-08-06”. This is only the model release or update date, which is different from the ”knowledge cutoff". Everyone should pay attention to this.

Pricing Unit: Token

In OpenAI API pricing,Token is the basic unit for calculating usage costs. Understanding Tokens is very important for correctly estimating API call fees.

What is a Token?

• Basic Concept of Tokens：

A Token is the smallest unit used by a language model when processing text. It can be a word, part of a word, or even a punctuation mark. For example, the English word “chat” is one Token, while a long word like “incredible” might be split into multiple Tokens. Punctuation marks and spaces may also count as Tokens.

• Difference Between Tokens and Characters：

Tokens do not simply correspond one-to-one with characters or bytes. The model splits and encodes the input text to generate multiple Tokens. Therefore, the longer the input text, the more Tokens are used.

Specific Example: For example, the sentence “OpenAI is awesome!” consists of 4 Tokens: ”OpenAI”, ”is”, “awesome”, “!”. A sentence in Chinese, however, might be broken down into more Tokens. For some complex texts, the number of Tokens may increase significantly.

The Role of Tokens in OpenAI Pricing

In the OpenAI API,each API call is billed based on the number of Tokens. When calling the API to generate text or perform other operations, the system calculates the total number of input and output Tokens and charges according to the Token count.

• Input Tokens: The text you send to the model.

• Output Token: Model-generated response.

Total Token count per API call = Input Token count + Output Token count。

Token Billing Example: Suppose you input a text containing 100 Tokens and request to generate a response of 50 Tokens, the system will bill you for 150 Tokens. Depending on the specific model used (such as GPT-4 or GPT-3.5), the price per Token varies.

Current OpenAI Token Prices

Currently, the prices for GPT-4o on the official website are as follows:

OpenAI has other model versions and features, which I won't introduce one by one. For more OpenAI model versions and prices, please refer to the official website:https://openai.com/api/pricing/。

Azure OpenAI

What is the relationship between Azure OpenAI and the OpenAI mentioned in the previous section?

The relationship between Azure OpenAI and OpenAI can be described as ”Microsoft Azure platform provides OpenAI's models through cloud services,” allowing developers to access and use OpenAI's advanced AI models within the Azure environment. Specifically:

• OpenAI is an independent AI research company that developed models like GPT-3, GPT-4, Codex and other large language models, which are widely used for tasks such as natural language processing, text generation, and code generation.

• Azure OpenAI is Microsoft Azure Cloud Platform a service on which provides OpenAI's model interfaces, enabling users to easily integrate and deploy OpenAI's models within the Azure environment.

The relationship between the two can be summarized as:

1. Technical Cooperation: Microsoft and OpenAI have established a deep strategic partnership, where Microsoft provides the cloud infrastructure and OpenAI focuses on developing advanced AI models.
2. API Provision: Through the Azure OpenAI service, Microsoft's cloud customers can directly access OpenAI's GPT models, Codex, and other cutting-edge AI technologies via APIs. Users can leverage the integration advantages provided by Azure (such as security, data management, etc.) to build and deploy AI applications.
3. Differentiated Platform: Although OpenAI has its own platform providing APIs, the Azure OpenAI service is part of the Azure ecosystem, helping users seamlessly combine OpenAI's technology with Microsoft's other cloud products (such as Azure Machine Learning, storage, databases, etc.).

Actually, the most critical point is: OpenAI itself restricts access from mainland China IPs, so under normal circumstances, you must use VPNs or proxies to use OpenAI; and if you directly access the official OpenAI API, you often need to use third-party proxies to initiate access, which is not only unstable but also a hassle. However, the API provided by Azure OpenAI can be accessed directly without using VPNs or proxies, which is extremely rare (currently, Azure OpenAI seems to be the only foreign API provider that can be used directly in China without VPNs or proxies).

If users normally use a local large language model UI to access OpenAI, there are many options. The key lies in some scenarios where the official OpenAI API address is hardcoded (such as WordPress AI plugins; for specific setup steps, please refer to the article:Home Data Center Series: Exploring the Functions and Built-in Tools of the WordPress Chatbot Plugin ”AI Engine”), when third-party API providers' API addresses cannot be used normally, having an OpenAI API address that can be used directly without VPNs or proxies is very important.

At the same time, another advantage of Azure OpenAI compared to other API providers is that its API billing method is not the official OpenAI monthly subscription model, but pay-as-you-go. For example, if you only use 1,000 Tokens this month, you only need to pay for those 1,000 Tokens. This is especially suitable for friends who usually use other API providers but just need a backup solution.

Note: Azure provides newly registered users with a ”free trial credit worth $200” to try out their various cloud services within the first 30 days. This offer applies to first-time Azure registrants, allowing them to experience and test Azure's various products and services without immediate payment (which means you can use OpenAI freely for the first month!).

The ”OpenAI” I mention in the subsequent content of the article includes both ”OpenAI” and "Azure OpenAI" without distinction, because they are essentially the same product sold through different channels. However, both have their own advantages.

Official OpenAI:

• Disadvantages: The Plus plan has a $20 monthly subscription, which is very unfriendly to light users (like me) (can't afford it~).
• Advantages: There is a free daily quota of GPT-4o to use for free (and it feels like the trained knowledge base is newer, though I don't know if it's just my imagination). After the quota is used up, it switches to GPT-3.5-turbo, which is enough for light users if used sparingly.

Azure OpenAI:

• Disadvantages: No free daily quota, very unfriendly to free riders.
• Advantages: No monthly subscription fee, pay-as-you-go, can be used as a backup option.

Google Cloud AI – Gemini

Introduction to Google Gemini

Google Gemini is the latest AI language model series released by Google, representing Google's top-tier level in the fields of generative AI and natural language processing. Gemini is Google's flagship large language model, designed to process and generate natural language, supporting various tasks from text analysis to complex conversations. Through Gemini, Google will significantly enhance its capabilities in the field of generative AI.

Google Gemini Features:

1. Multimodal Support：

The Gemini model not only supports text generation and processing but can also understand and generate content in various forms such as images, videos, and code. This multimodal capability makes Gemini perform excellently in applications such as automatic image captioning and multimedia content understanding.

2. Powerful Language Understanding and Generation：

Similar to OpenAI's GPT-4, Gemini possesses extremely high natural language understanding and generation capabilities. It can handle complex contexts and generate high-quality, contextually appropriate text, suitable for tasks such as chatbots, automatic document generation, and code autocompletion.

3. Wide Range of Application Fields：

Gemini can be applied to multiple industries and fields, including healthcare, law, finance, customer service, etc. Whether it is the need to automatically generate professional reports or conduct large-scale data analysis, Gemini can quickly adapt and provide intelligent solutions.

4. Efficient Knowledge Integration：

Google leverages its vast data resources and knowledge graph to ensure Gemini has a strong knowledge base. This makes it more accurate and able to incorporate the latest facts and knowledge when generating answers, writing articles, and performing other tasks.

Market Share and Broad Applicability：

• Google's Powerful Ecosystem：

The Gemini model can be seamlessly integrated with Google's other products and services (such as Google Cloud, Google Workspace), making it easy for enterprises to quickly deploy AI solutions.

• Multimodal Capabilities：

Compared to its competitors, one of Gemini's greatest advantages is its multimodal processing capabilities. This allows it to not only process text but also understand and generate content in multiple forms, making it suitable for more scenarios and applications.

• Scalability and Security：

Relying on Google Cloud's powerful infrastructure, Gemini performs stably when processing large-scale data and provides enterprise-grade security protection, making it suitable for the needs of small to large enterprises.

Main Model Versions and Their Features

Google has released multiple versions of the Gemini model, each tailored to different use cases and needs. Below are the current main versions and their corresponding features:

Gemini 1

Gemini 1 is the earliest model version, supporting powerful language generation capabilities, suitable for basic chatbots, text summarization, and content generation tasks. It supports natural language processing and generation, making it suitable for simple dialogue systems, article generation, and other applications.

Gemini 1.5

Gemini 1.5 is an upgraded version of the first-generation model, improving generation quality and reasoning capabilities, and optimizing multimodal processing to handle complex tasks including images and code. It not only supports text processing but also understands and generates images and videos, making it suitable for multimodal application scenarios such as automated content creation and code generation.

Gemini 2

An upcoming version, Google states that this version will further enhance multimodal AI capabilities, allowing for more precise integration of text, images, audio, and other data to provide smarter solutions for enterprises and users. It is expected to significantly improve model reasoning efficiency and provide support in more complex industry scenarios, such as medical diagnosis and legal analysis.

Gemini Pro and Gemini 1.5 Pro

• Gemini Pro: This is a high-end version tailored for enterprise-level users, offering faster processing speeds and stronger generation quality. It is suitable for large enterprises that require high accuracy and high concurrency processing.
• Gemini 1.5 Pro: An enhanced version based on Gemini 1.5, focusing on efficient processing and security in enterprise application scenarios, particularly suitable for scenarios requiring high-performance reasoning and multinational data analysis.

Advantages and Comparison of Gemini Models

1. Gemini vs. GPT-4 ：

Gemini's multimodal support and integration with the Google Knowledge Graph give it an advantage in handling complex tasks. In contrast, GPT-4 focuses on language generation, performing exceptionally well in text generation quality and the refined processing of language models.

2. Differences Between Gemini and Other Models：

Gemini emphasizes cross-domain capabilities; it can not only generate text but also understand and generate multimodal content such as images and videos, whereas most language models focus primarily on text generation.

Note: In addition to the Gemini large language model, Google has another very powerful natural language processing tool: Google Cloud Natural Language API. Both have their own strengths; the former focuses on generative AI, while the latter focuses on text analysis and understanding. However, the Google Cloud Natural Language API is mainly targeted at enterprises and developers, and generally has little to do with individual AI users (unless called via plugins, etc., see the next section).

Additional Knowledge: Google Cloud Natural Language API

Google Cloud Natural Language API is a cloud-based service designed to help developers and enterprises understand and process text information. It can automatically analyze articles, documents, or conversations to extract key topics, sentiments, and entities (such as people, places, organizations, etc.). In short, Google's Natural Language API gives computers the ability to understand human language, helping businesses extract useful information from massive amounts of text.

Product Features of Google Cloud Natural Language API:

1. Sentiment Analysis: It can automatically determine the emotional inclination of an article or a piece of text, whether it is positive, negative, or neutral. This is highly helpful for customer feedback and social media monitoring.
2. Entity Recognition: The API can extract important information from text, such as mentioned people, places, organizations, products, etc. This is very useful for content classification, information retrieval, and structured text processing.
3. Syntactic Analysis: It can parse sentence structures and analyze the vocabulary and grammatical relationships within them. This has great application value for tasks such as text understanding, automatic translation, and language learning.
4. Content Classification: The API can automatically classify text into different topic categories (such as sports, technology, entertainment, etc.), which is highly practical in scenarios like automated content management and news classification.
5. Multi-language Support: Google Cloud Natural Language API supports not only English but also the analysis of multiple languages, including Chinese, Spanish, French, etc., making it suitable for global enterprises and users. Some translation software can complete multi-language translation by calling this API, such as the translation plugins ”GTranslate” and ”TranslatePress” on WordPress. Friends who are interested in using these two plugins can refer to my other article:Home Data Center Series: Implementing Multi-language Automatic Translation and Multi-language SEO Best Practices on WordPress Sites (GTranslate & TranslatePress)。

Market Share and Broad Applicability：

Google Cloud Natural Language API holds a leading position in the global market, especially in processing unstructured text data. It is widely used in fields such as data analysis, customer service systems, financial services, and content recommendation. Relying on Google's powerful cloud computing infrastructure, enterprises can easily scale their API usage, whether analyzing millions of user reviews or processing complex social media data.

Greatest Competitive Advantages：

• Deep Integration with the Google Ecosystem: Google Cloud Natural Language API seamlessly integrates with Google's other cloud services (such as BigQuery, Cloud Storage), facilitating data storage, analysis, and visualization.
• Accurate Language Understanding: Due to Google's long-accumulated natural language processing technology, its API has high accuracy in sentiment analysis, entity recognition, and content classification.
• Scalability: Google's cloud infrastructure ensures that the API can handle various needs from small-scale text to large-scale datasets, making it suitable for use by small to large enterprises.

Anthropic Claude API

Introduction to Anthropic Claude API

Anthropic Claude API is an artificial intelligence language model platform launched by Anthropic, aimed at providing developers with powerful and secure natural language processing capabilities. Through the Claude API, users can use these models in various applications for tasks such as text generation, dialogue, Q&A, content writing, and data analysis. Named after Claude Shannon, the founder of information theory, Claude's design philosophy places special emphasis on safety, controllability, and reliability to ensure that the output content poses no potential risks to users and society.

Claude API emphasizesSecurity和ethical standards, and is designed to minimize the risk of generating inappropriate, misleading, or harmful content as much as possible. This makes it highly suitable for use in fields with high output quality requirements, such as healthcare, finance, and education.

Another major advantage of the Claude API lies in itsease of use和scalability, allowing developers to very easily integrate it into existing systems and access powerful natural language processing capabilities through simple API calls. In addition, the Claude model features a high degree ofStability和adjustability, capable of adjusting the style and complexity of the generated content according to needs.

Main Model Versions and Their Features

Anthropic provides several different Claude versions to adapt to different computing needs and task types. Each version differs in performance, response time, and feature set, ensuring coverage of diverse applications from lightweight tasks to complex text generation.

Claude 3.5 Sonnet

Claude 3.5 Sonnet is a high-performance variant of the Claude 3.5 version, specifically optimized for performance in complex text processing and long article generation. It can generate longer, more coherent text, making it suitable for tasks that require processing large amounts of data or long-form content, such as report generation and novel writing.

Main Functions：

• Provides high-quality, coherent long text generation.

• Capable of maintaining contextual consistency in complex dialogues and multi-turn Q&A.

• Optimized for generation tasks with longer processing times.

Claude 3.5 Lite

Claude 3.5 Lite is a lightweight version specifically designed to handle relatively simple text generation tasks. It has faster response times and is suitable for applications with high real-time performance requirements, such as intelligent customer service and online Q&A scenarios.

Main Functions：

• Provides fast-response text generation.

• Consumes fewer computing resources, suitable for lightweight scenarios.

• Suitable for tasks with high speed requirements, such as real-time dialogue and user interaction.

Claude 3.5 Chat

Claude 3.5 Chat is specifically optimized for dialogue systems, capable of maintaining efficient and consistent performance in dialogue generation and multi-turn interactions. Its application scenarios include intelligent customer service, virtual assistants, and other occasions requiring continuous dialogue with users.

Main Functions：

• Efficiently handles multi-turn dialogues.
• Maintains the coherence of dialogue context.
• Optimized for answering complex questions and dialogue generation.

Claude 3.0

Claude 3.0 is the base model version of this series, suitable for various general natural language processing tasks. It can handle a wide range of tasks from text generation and sentiment analysis to simple Q&A systems.

Main Functions：

• General-purpose natural language generation and understanding.
• Supports multi-turn dialogue and context maintenance.
• Provides efficient and accurate text generation in most cases.

Differences between model versions:

• Claude 3.5 Sonnet is a model targeted at high-quality, long-form text generation, suitable for scenarios requiring long-term contextual consistency, such as complex report writing and long-form content generation.

• Claude 3.5 Lite is a lightweight version with fast response times, suitable for applications requiring rapid generation of results, such as customer service systems or dialogue systems.

• Claude 3.5 Chat is specifically optimized for multi-turn dialogues, ensuring smooth and natural interactions in complex conversations.

Generally, we can just directly use the latest Claude 3.5 version.

Additional Knowledge: Other Official API Providers

In addition to OpenAI, Azure OpenAI, Google Gemini, and Anthropic Claude, there are several other commonly used large language model API providers on the market. Below is a brief introduction to these providers, their characteristics, and the model versions they offer, to give everyone an overview.

Cohere

Cohere is a company that provides large-scale natural language processing APIs, focusing on natural language understanding and generation, suitable for various tasks such as text analysis, document classification, and sentiment analysis. Cohere's models target developers and enterprises, providing flexible and efficient language model services.

Features and Advantages:

• Text Generation and Understanding: Cohere's API provides powerful text generation and understanding capabilities, suitable for tasks such as generating complex text, summarization, and translation.
• Custom Models: Cohere supports users in performingFine-tuningto adapt to specific industry needs.
• Open Architecture: Allows users to customize model outputs, supporting different languages and styles, suitable for multilingual applications.
• Security and Compliance: Cohere emphasizes model security and the protection of user data privacy.

Main Models:

• Command: Cohere's flagship model, focusing on instruction-based text generation tasks, suitable for text writing, answering questions, and generating dialogues.
• Rerank: Used to improve the accuracy of search results, suitable for information retrieval and ranking tasks.
• Embed: Focuses on text embedding, suitable for tasks such as text classification and similarity analysis.

Hugging Face (Transformers API)

Hugging Face is a leader in the open-source AI community, providing hundreds of natural language processing models. Its Transformers API allows developers to call various pre-trained large language models, including the GPT series, BERT, RoBERTa, and other models, to meet the needs of various natural language processing tasks.

Features and Advantages:

• Rich Variety of Models: Hugging Face provides a massive library of open-source models, allowing developers to choose from various open-source models and fine-tune existing models to meet specific needs.
• Community Support: The Hugging Face community is highly active, allowing developers to quickly learn and apply models through community resources.
• Multi-task Support: Supports multiple tasks such as text generation, translation, sentiment analysis, information extraction, and dialogue systems.
• Low Cost and Open Source: Most models are free to use, and free fine-tuning and hosting services are provided.

Main Models:

• GPT-2/3: Hugging Face provides open-source versions of the GPT series, supporting tasks such as text generation and dialogue.

• BERT: Suitable for tasks such as text classification, sentiment analysis, and Q&A systems.

• RoBERTa: An improved version of BERT, excelling in text understanding and context analysis.

Mistral

Mistral is an emerging large language model provider focusing on high-performance open-source language models. Its released models perform exceptionally well in accuracy and computational efficiency, making them suitable for tasks requiring high-quality generation and text processing. Mistral focuses on providing flexible open-source models, suitable for customization by developers and enterprises.

Features and Advantages:

• Open Source Model: Mistral provides open-source large language models, allowing users to customize applications locally or in the cloud.
• High-Performance Models: Its models can significantly improve processing speed without sacrificing generation quality, making them suitable for large-scale concurrent tasks.
• Focus on Text Generation: Mistral's models are particularly good at generating natural and fluent text, suitable for scenarios such as content creation, code generation, and Q&A systems.

Main Models:

• Mistral-7B: Mistral's flagship model with 7 billion parameters, suitable for complex text generation tasks. Compared to other similar models, Mistral-7B offers higher cost-performance in terms of performance and accuracy.

Meta (LLaMA API)

Meta (Facebook) has launched its own large language model LLaMA, which is mainly used for research and commercial applications. LLaMA provides different versions suitable for various natural language processing tasks, including text generation, context understanding, and Q&A.

Features and Advantages:

• Lightweight: Compared with other large language models, LLaMA significantly reduces the consumption of computing resources while maintaining high performance.
• Open Source Models: LLaMA's models are open-source, allowing developers to perform secondary development and fine-tuning based on LLaMA.
• Research-Driven: LLaMA is mainly oriented towards academia and research institutions, suitable for advanced research tasks.

Main Models:

• LLaMA 2: LLaMA 2 is the next-generation language model launched by Meta, featuring enhanced context understanding and text generation capabilities, suitable for all types of natural language processing tasks.
• LLaMA 13B/65B: With parameter scales of 13 billion and 65 billion respectively, specifically designed to handle complex natural language tasks.

Third-Party API Provider: OhMyGPT

Introduction to OhMyGPT API

OhMyGPT API is a third-party platform specializing in providing API integration services for multiple large language models (official website:https://www.ohmygpt.com/)

. Its main function is to help users access different large language model APIs more conveniently and flexibly, without the need to separately configure and manage multiple accounts or platforms. Therefore, users can call language models from multiple providers such as OpenAI, Anthropic, and Google in a one-stop manner through OhMyGPT, thereby reducing administrative and technical burdens.：

• Product FeaturesUnified Interface Management
• : OhMyGPT provides users with a platform that integrates APIs of multiple language models, calling different models through a unified interface, simplifying the development process.Transparent Pricing
• : Although OhMyGPT itself does not provide models, it integrates API services on the market and helps users save costs through packaged pricing or flexible billing.Seamless Provider Switching
• Developer-friendly: OhMyGPT provides concise documentation and rich examples, making it easy for developers with different technical backgrounds to get started.

Features and Market Competitive Advantages：

• Simplified Integration: Developers do not need to register separate accounts for multiple large language model APIs. OhMyGPT provides a unified API integration service, simplifying the integration process.
• Cost Optimization: By bundling services from different API providers, OhMyGPT provides users with more flexible payment options, suitable for budget-conscious developers and small to medium-sized enterprises.
• Support for Multiple Providers: Whether it is OpenAI or Anthropic Claude, OhMyGPT allows users to quickly integrate and flexibly choose the most suitable model based on their needs.

Main API Providers Supported by OhMyGPT and Their Features

OhMyGPT does not have its own model; its role is to provide users with a platform that integrates services from major large language model API providers. Below are the main language model API providers supported by OhMyGPT and their features:

1. OpenAI API: Supports GPT-4 and GPT-3.5 series models, excelling in text generation, dialogue systems, and text analysis tasks.
2. Anthropic Claude API: Supports the Claude series models, excelling in highly secure and robust dialogue and text generation tasks.
3. Google Gemini API: Provides powerful multimodal processing capabilities, suitable for the comprehensive processing of different data types such as images and text.
4. Azure OpenAI API: Provided by Microsoft, it integrates Azure's cloud computing advantages in addition to OpenAI model services, making it suitable for enterprise-level application scenarios.

All supported API providers and models are as follows:

Advantages of OhMyGPT

Using services from other API providers through OhMyGPT offers the following advantages:

• Cross-platform Use: Through OhMyGPT, users only need to pay once to seamlessly switch between multiple API providers, adapting to different usage scenarios (requires use with a local large language model UI that supports quick switching, such as LobeChat (ChatGPT Next Web should also work, though I haven't tried it)):

• One-stop Experience: Simplifies the management of multi-platform APIs and reduces the difficulty of use, making it especially suitable for developers who need to frequently switch between different models. In other words, multiple API providers can be accessed through a single API address (same as the point above, requires app or UI support). At the same time, you can clearly see the consumption details of each API:

• Multiple API Addresses to Choose From: The official site provides multiple API routes, suitable for use in different domestic and international network environments.

• High Cost-Performance Ratio: OhMyGPT adopts a prepaid points system and provides daily free points, making it more suitable for friends with small-scale, infrequent use (like me~). Most importantly, the basic usage fee is only 20 RMB (even I can barely afford it):

Note: Remember the ”custom API address and model version” feature supported by the local large language model UI marked earlier? As long as it can be customized, in theory, it can be used with OhMyGPT to access multiple API providers and model versions. However, there is a gap between ideal and reality; theory is theory, and in practice, not all of them will succeed (due to API format compatibility issues), so all I can say is you can try it out first.

Summary

The previous text mentioned different language models and versions from multiple API providers, and these language models have different areas of expertise and pricing, which might make your head spin when choosing. When choosing different API providers and specific model versions in daily life, it depends on your considerations of the model's ”functional requirements, speed requirements, budget, and application scenarios.”

Below is a comparison of some key aspects of commonGPT-4o-mini、GPT-4o、OpenAI o1、Google Gemini 1.5, andAnthropic Claude 3.5 Sonnet for your reference.

GPT-4o-mini

This is an OpenAIoptimized lightweight model, which is faster and more economical than the standard GPT-4o, while retaining the powerful reasoning capabilities of the GPT-4 series.

Applicable Scenarios:
• Suitable fordaily conversations, lightweight tasksor applications requiring fast responses, such as simple customer service systems, real-time chat, etc.

Advantages:
• Lower cost, suitable for users with limited budgets.
• Fast response speed, suitable for real-time application scenarios.

Disadvantages:
• Compared to the full version of GPT-4o, it performs weaker on more complex tasks.
• Relatively lower language understanding and reasoning capabilities.

GPT-4o

A standard optimized version of GPT-4, with powerful language understanding and reasoning capabilities, performing excellently in handling complex problems.

Applicable Scenarios:
• Suitable for applications requiringhigh precision and complex language processing, such as content creation, technical Q&A, complex customer support, etc.

Advantages:
• Powerful text generation and reasoning capabilities, suitable for a wide range of complex tasks.
• Performs excellently in multitasking.

Disadvantages:
• Relatively high usage cost.
• Response speed may not be as fast as lightweight versions.

OpenAI o1

Another important model in the OpenAI series, emphasizingmulti-domain performance, and excels at handling various complex NLP tasks.

Applicable Scenarios:
• Suitable for applications requiring multi-domain knowledge and higher precision, such as content creation and technical analysis in vertical fields like law, medicine, and education.

Advantages:
• Broad domain coverage, excelling at handling highly difficult problems.
• Performs outstandingly when generating text with complex context.

Disadvantages:
• Relatively high price.
• May seem overly powerful and wasteful of resources for simple tasks.

Google Gemini 1.5

Google's large language model series, featuring powerfulMultimodalprocessing capabilities (text, images, etc.), with the Gemini series emphasizing integration with real-world knowledge.

Applicable Scenarios:
• Suitable for application scenarios requiring cross-modal processing, such as generating content combining text and images, mixed vision and language tasks, and knowledge reasoning.

Advantages:
• Multimodal support, suitable for tasks requiring image and text processing.
• Deeply integrated with Google Knowledge Graph, making information more precise and reliable.

Disadvantages:
• Higher cost, especially for small and medium-sized applications.
• May not have a clear advantage for pure text processing tasks.

Anthropic Claude 3.5 Sonnet

The Claude series of models is known forsafety and stability, focusing on the optimization of controllability and dialogue generation, emphasizing safe handling of sensitive topics and ethical issues.

Applicable Scenarios:
• Suitable for applications requiringDialogue systems, customer supportand other applications, especially in situations with high requirements for model safety and sensitive content management, such as healthcare, psychological counseling, and other fields.

Advantages:
• Strong dialogue generation capability, focusing on safety, suitable for scenarios requiring high ethical standards.
• Outstanding context understanding capability in long dialogues.

Disadvantages:
• Cost may be high, especially in cases of widespread application.
• Compared to other models, it may have slight limitations in data breadth.

How to Choose?

Limited budget, pursuing fast response：
• Choose GPT-4o-mini. Its performance is sufficient to handle most daily tasks, with lower cost and faster response speed.

Requiring high precision and complex problem solutions：
• Choose GPT-4o or OpenAI o1 (expensive). These two models can handle complex tasks, such as content creation, in-depth technical Q&A, or multi-domain analysis.

Cross-modal processing (text + images)：
• Choose Google Gemini 1.5. If your application needs to handle combined text and image content, this model is more suitable.

Emphasizing safety and stability in dialogue generation：
• Choose Claude 3.5 Sonnet. If your application has high requirements for ethical safety, such as in healthcare, mental health, and other fields, the advantages of this model will be more obvious.

Finally, no matter which of the above models you choose, you can try using them through OhMyGPT.

Note 1: It is said that from the perspective of writing code, OpenAI o1 or OpenAI o1-preview is the best, but too expensive, while Claude 3.5 Sonnet is not much different from OpenAI o1 but much cheaper. If any friends have the need to write code using AI, you can verify if this is accurate.

Note 2: In addition to these API providers and models that I consider commonly used, there are many other API providers and models that have advantages in certain specific vertical fields. However, due to my limited knowledge, I haven't heard of them and naturally cannot list them all. If you have unconventional needs, do not limit yourself to the conventional API providers in this article. You can look at other API providers with advantages in specific fields, weigh the model's capabilities, cost, and application scenario fit based on specific needs, and ultimately confirm the most suitable API provider.

Note 3: It is said that some domestic API providers are also good to use, interested friends can give them a try.