Great news, the GPT-4 API is fully open for use!

Source: CSDN

Arrangement: Tu Min

OpenAI, which just launched GPT-4 in March this year, only invited some developers who submitted applications to participate in the test. Seeing Greg Brockman, the co-founder of OpenAI, demonstrate the "most powerful" GPT-4 model on the spot at the time, easily generate a website through a hand-drawn sketch, and can develop a small game in 60 seconds. But it cannot be used.

And just today, the applicability of GPT-4 has been further expanded. OpenAI officially released the GPT-4 API, which is now fully open to all paid API developers!

OpenAI Roadmap: Expand GPT-4 to new developers by the end of this month

In the four months since it went online, I believe that many people have analyzed papers through technology (not too unfamiliar with GPT-4.

According to OpenAI, since the release of GPT-4 in March this year, millions of developers have requested access to the GPT-4 API, and the range of innovative products utilizing GPT-4 is growing every day.

GPT-4 differs from its predecessor, GPT-3.5, in its enhanced ability to generate text, including code, while also accepting image and text inputs.

The model performs "human-level" on various professional and academic benchmarks. Previously, OpenAI researchers also tested it, saying that if GPT-4 was a person judged by test-taking ability alone, it could get into law school, and likely many universities as well.

Like OpenAI's previous GPT model, GPT-4 was trained using publicly available data, including data from public web pages, as well as data licensed from OpenAI. From a technical perspective, GPT-4 is a Transformer-based model that can predict the next token in a document after pre-training. A central part of this project was the development of infrastructure and optimization methods. This enables OpenAI to accurately predict certain aspects of GPT-4's performance based on models trained with no more than 1/1000th the computation of GPT-4.

However, it is a bit of a pity that the **image understanding capability is not yet available to all OpenAI users. ** Still, as OpenAI announced in March this year, it is only testing with its partner Be My Eyes. So far, OpenAI has not indicated when it will open up its image understanding capabilities to a wider user base.

Today, all API developers with a successful payment record can** access GPT-4 APIs with 8K contexts, of course, this time they cannot access 32K contexts. **

At the same time, **OpenAI also plans to open access to new developers at the end of this month, and then start increasing the rate limit according to the availability of calculations. **

It’s worth noting that while GPT-4 represents an important milestone in the development of generative AI models, that doesn’t mean it’s 100% perfect. As far as GPT-4 is concerned, it is also possible to "hallucinate" and make some reasoning errors with confidence.

In the future, OpenAI says it is also working to safely enable fine-tuning for GPT-4 and GPT-3.5 Turbo, which is expected to be available later this year.

Chat Completions API

In the announcement, OpenAI also announced that the GPT-3.5 Turbo, DALL-E and Whisper APIs will be generally available to developers.

At the same time, it also shared the current status of the widely used Chat Completions API with developers. OpenAI says the Chat Completions API now accounts for 97% of its API GPT usage.

OpenAI notes that the original Completions API was launched in June 2020, providing a free-form text prompt for language models to interact with.

The Chat Completions API's structured interface (e.g., system messages, function calls) and multi-turn dialog capabilities enable developers to build conversational experiences and extensive completion tasks while reducing the risk of hint injection attacks because user-provided content can be extracted from the structure separate from the command.

OpenAI said that it is currently making up for several shortcomings of the Chat Completions API, such as the log probability of the completion token and increasing the bootability to reduce the "chatiness" of the response.

Deprecation of old models

In addition, OpenAI also released the deprecation plan of the old model. That is, starting from January 4, 2024, some old OpenAI models, especially GPT-3 and its derivative models, will no longer be available and will be replaced by the new "GPT-3 base" model, the new model Computational efficiency will be higher.

According to the announcement, the specific obsolete models include some old models in the Completions API, including the familiar davinci:

*Apps using stable model names based on GPT-3 models (ada, babbage, curie, davinci) will be automatically upgraded to the new models mentioned above on January 4, 2024. In the coming weeks, new models will also be accessible for early testing by specifying the following model names in API calls: ada-002, babbage-002, curie-002, davinci-002.

• Developers using other older completion models (such as text-davinci-003) will need to manually upgrade their integrations before January 4, 2024, specified in the "model" parameter of their API requests gpt-3.5-turbo-instruct. gpt-3.5-turbo-instruct is an InstructGPT style model, the training method is similar to text-davinci-003. This new model is a replacement in the Completions API and will be available for early testing in the coming weeks.

Meanwhile, OpenAI says that developers who wish to continue using their fine-tuned models after January 4, 2024 will need to work on the new GPT-3-based models (ada-002, babbag-002, curie-002, davinci- 002) or updated models (gpt-3.5-turbo, gpt-4) for fine-tuned replacement.

As OpenAI turns on fine-tuning later this year, they will prioritize GPT-3.5 Turbo and GPT-4 fine-tuning for users who have previously fine-tuned older models. The specific reason is that OpenAI understands that migrating from a model that has been fine-tuned on its own data is challenging, and they will provide support for "users of previously fine-tuned models to make this transition as smooth as possible."

In addition to retiring some old Completions API models, OpenAI said that users of old embedding models such as text-search-davinci-doc-001 will also need to migrate to text-embedding-ada-002 by January 4, 2024.

Finally, users using the Edits API and its associated models (such as text-davinci-edit-001 or code-davinci-edit-001) will also need to migrate to GPT-3.5 Turbo by January 4, 2024. The Edits API beta is an early exploratory API designed to enable developers to return edited versions of their prompts on command.

OpenAI wrote in the announcement, "We recognize that this is a significant change for developers using these old models. Discontinuing these models is not a decision we take lightly. We will bear the burden of users re-embedding content with these new models." Financial costs."

OpenAI said it will contact affected users in the coming weeks and they will provide more information once the new model is ready for early testing.

Preview: Next week, all ChatGPT Plus users will be able to use the code interpreter

The most anticipated thing is that OpenAI officials also announced on Twitter that the code interpreter will be open to all ChatGPT Plus users next week.

It allows ChatGPT to run code and optionally access user-uploaded files. Developers can directly ask ChatGPT to analyze data, create graphs, edit files, perform mathematical operations, and more.

I wonder if everyone is looking forward to it?

10 things you can do with GPT-4 API

Finally, with the release of the GPT-4 API this time, developers no longer have to bother to purchase Plus services. Calling the most powerful GPT-4 API in the industry so far will undoubtedly make related applications more intelligent.

So, what can we do with the GPT-4 API, and foreign media have summarized 10 common usages:

• Based on the narrative ability of GPT-4 API, novel content such as complex plots and character development can be quickly generated, completely changing the field of literary creation.
• The GPT-4 API paves the way for simulating extremely realistic conversations, reflecting the realism and precision of human interactions.
• GPT-4 API demonstrates the ability of instant language translation, effectively bridging the communication gap between various languages and cultures.
• GPT-4 API has strong capabilities in data analysis and can provide valuable insights to data analysis market players.
• The GPT-4 API enables the creation of highly realistic virtual environments that dynamically mirror the real world, enhancing immersion in areas such as gaming and virtual reality.
• The ability of the GPT-4 API to generate complex computer code makes it an indispensable ally for software developers.
• GPT-4 API can interpret and analyze medical data to help accurately diagnose and predict various health conditions.
• Leveraging its advanced language generation capabilities, the GPT-4 API ensures fast and accurate legal document generation.
• GPT-4 API demonstrates the ability to interpret consumer data and generate custom marketing content that effectively resonates with target audiences.
• The GPT-4 API has the potential to drive scientific innovation by analyzing large volumes of scientific data, discovering new insights in fields as diverse as chemistry, physics, and biology.

reference: