Millions of developers have been waiting for GPT-4 access since March. All those developers who have successfully paid will start receiving GPT-4 Model access by the end of this month. It will boost performance and accuracy by almost 40% compared to the last version. Recently, OpenAI has seen a decline in ChatGPT usage, around 9.7% compared to June.
Recently, ChatGPT has also disabled the Browse with Bing feature because it was generating wrong messages and error codes. This feature allows users internet access, but the plugin remains intact. The company may release the Browse with Bing feature with future updates.
GPT-4 Capabilities
This is a significant upgrade. GPT-4 now accepts images and has improved text accuracy regarding prompts. With the previous GPT-3 model, it was only capable of processing text, but now it can also process images. As for performance, the company has seen significant improvements in academic benchmarks, with the ability to score in the top 10%.
- Create long-form content, extend conversations, and conduct document analysis and search tasks.
- Can write more realistic dialogue.
- Has more authenticity and precision in human interactions.
- Better language translation.
- Advanced data analytics and better predictions of the subsequent text.
- Can generate, debug, and write more robust computer code.
- Has better accuracy in diagnosis and prediction of various health conditions.
- Can adjust inappropriate requests.
- Provides fact-based responses.
Not all features will be instantly available to everyone. Some features may still have restrictions, like image processing capabilities. The company is already testing this functionality with Be My Eye for image processing capabilities. GPT-4 can also remember more recent conversations, and since it is more accurate, it is improbable for GPT-4 to go off-topic.
Existing API developers can access GPT-4 with 8K content and a successful payment history. This translates to about 13 pages of text. New models will be responsible for 99.9% of embedding API use, and the company is aware that switching may cost developers with the GPT-4 re-embedding.
New GPT-4 Based Model
OpenAI has introduced five new AI models based on GPT-4, which are successors to the existing GPT-3 based models. Developers can test these new models by deploying them on their local machines and services or by trying them directly from the OpenAI Playground.
The five new models are:
- ada-002
- babbage-002
- curie-002
- davinci-002
- gpt-3.5-turbo-instruct
- text-embedding-ada-002
Developers can train these AI models with their own datasets and personalize them to respond more specifically to their apps and services. Additionally, the company is extending the character limit from 9,000 to 100,000 tokens, which translates to an impressive 32,000 tokens at the upper end. Among the competition, AGL models like Antrohic are expanding their text generation modules from 9,000 tokens to 100,000.
Regarding the Chat Completions API, which was introduced in March, it now accounts for 97% of the OpenAI API and is getting better with the new models.
How to Use GPT-4 in OpenAI Playground
Using GPT-4 in OpenAI Playground is quite similar to how you used GPT-3. All you need to do is select a GPT-4-based model. Before proceeding, we would like to mention that you need access to GPT-4 to use it.
- Visit the OpenAI Platform in your browser and log in with your OpenAI account.
- After logging in, you will be redirected to the Playground. If not, click on Playground from the top navigation bar.
- You will find some settings to choose from or adjust on the right sidebar.
- From the drop-down menu of Models, choose any of the GPT-4-based models. All these models have their capabilities and strengths in performing tasks.
GPT-4 is already available on ChatGPT exclusively to ChatGPT Plus members, which costs $20/month and is also available as a free service on some third-party websites. Additionally, some companies like Snapchat, Opera, Duolingo, Stripe, and more use the GPT-4 model in their apps and services.
GPT-3.5 Turbo, DALL·E, and Whisper APIs
The company also promises to allow developers to fine-tune GPT-4 and GPT-3.5 Turbo and train their models with their own data by the end of this year. On the other hand, the company has also announced making Whisper APIs, a speech-to-text model, and DALL-E 2 for image-generation mode, generally available to everyone.
GPT-3 Will Shut Down in 2024
OpenAI has also announced that they will discontinue the GPT-3 model on January 4th, 2024. By that time, we will have much better AGL models that will be more efficient and benefit companies and users with their capabilities.