OpeaAI recently launched their latest and most powerful chatbot model to date and it is chatGPT-4 Turbo(gpt-4–1106-preview). It is faster, more accurate and has knowledge base of up to April 2023. Let’s delve into some more details.
Introducing GPT-4 Turbo: A Major Upgrade
I have been using GPT-4 Turbo in the OpenAI playground for quite some time now, and I can surely tell that it represents a substantial enhancement to the pre-existing GPT-4 large language model. This evolution introduces a slew of improvements, such as a significantly expanded context window and access to more recent knowledge. OpenAI has been steadily augmenting the capabilities of GPT-4 within ChatGPT, with the integration of customized instructions, ChatGPT plugins, DALL-E 3, and Advanced Data Analytics. This latest iteration introduces a plethora of thrilling new features.
GPT-4 Turbo with vision is a specialized version of GPT-4 Turbo that incorporates optical character recognition (OCR) capabilities. In essence, it has the ability to analyze images and extract text contained within them. For instance, if you provide it with an image of a menu, it can decipher the menu items displayed in the image. Similarly, it can process an image of an invoice and automatically extract the vendor’s name and item details.
The “with vision” functionality will come as a default feature in ChatGPT and will be available to developers through the selection of the “gpt-4-vision” model in the OpenAI API.
Key Features of GPT-4 Turbo
GPT-4 Turbo boasts a range of enhancements compared to its predecessors, elevating its capabilities to new heights. Here are some standout features:
1. Enhanced Knowledge Cutoff
Previously, GPT-3.5 and GPT-4 had a knowledge cutoff set at September 2021, rendering them incapable of addressing real-world events beyond that date unless provided with external data sources. GPT-4 extends this knowledge cutoff by nineteen months, enabling it to access information and events up to April 2023. This significantly enhances its reliability as a source of up-to-date information. OpenAI’s CEO, Sam Altman, has also pledged to maintain this current knowledge base, ensuring it remains relevant.
2. 128K Context Window
The context window of a large language model defines the duration of its conversational memory. A larger context window allows for more coherent and accurate responses in extended conversations or when dealing with lengthy documents.GPT-4 now offers a maximum context length of 32,000 tokens, whereas GPT-4 Turbo pushes this boundary to 128,000 tokens, equivalent to around 240 pages with 400 words per page. This aligns it with Nous Research’s YARN-MISTRAL-7b-128k model, exceeding the 100,000 token limit of Anthropic’s Claude 2 model.
It’s worth noting that the effectiveness of a longer context window throughout the entire conversation is yet to be determined, as existing research indicates that long context models excel primarily in accessing information near the beginning or end of a document. Moreover, efforts are underway to achieve the aspiration of “infinite context” with early-stage research from Microsoft and Xi’an Jiaotong University aiming to scale language models to one billion tokens of context.
3. Function Calling
Function calling is a feature designed to streamline interactions with the model for developers embedding generative AI into their applications. It allows them to describe their app’s functions or external APIs to GPT-4 Turbo. This feature supports the calling of multiple functions within a single message, eliminating the need for extensive back-and-forth exchanges with the model.
The cost reductions extend to GPT-3.5 Turbo models as well, offering input tokens at 0.1 US cents per 1000 tokens (3x cheaper) and output tokens at 0.2 US cents per 1000 tokens (2x cheaper). Furthermore, fine-tuned GPT-3.5 Turbo 4K model input tokens are now four times more economical, with the price dropping from 1.2 US cents to 0.3 US cents per 1000 tokens. Output tokens follow suit, becoming 2.7 times cheaper, with the price decreasing from 1.6 US cents to 0.6 US cents per 1000 tokens. The training price remains consistent at 0.8 US cents per 1000 tokens.
These pricing adjustments aim to make advanced AI models a cost-effective choice for developers.
4. GPT Goes Multi-Modal: Image Prompts & Text-to-Speech
OpenAI is gearing up to introduce “GPT-4 Turbo with vision” enabling users to prompt the model using images directly in the chat box. This capability will empower GPT-4 Turbo to generate image captions and provide descriptions of the visual content. Additionally, it will facilitate text-to-speech conversion, expanding its utility further.
5. More Affordable Pricing
OpenAI has adjusted the pricing of GPT-4 Turbo to make it more accessible to developers. The cost of GPT-4 Turbo input tokens via the OpenAI API has been reduced to one-third of its previous price, now standing at 1 US cent per 1000 tokens (down from 3 US cents). Output tokens have also seen a reduction, now available at 3 US cents per 1000 tokens, halving the previous cost of 6 US cents per 1000 tokens.
How to Access GPT-4 Turbo
Access to GPT-4 Turbo is open to all paying developers. If you have API access, you can simply specify “gpt-4–1106-preview” as the model name in the OpenAI API. Likewise, for GPT-4 Turbo with vision, you can designate “gpt-4-vision-preview” as the model name.
It’s important to note that these preview models are not yet recommended for production use. Nevertheless, Altman has assured that a production-ready version will be made available in the near future.
For non-developers, GPT-4 Turbo is expected to become accessible to ChatGPT Plus and ChatGPT Enterprise users in the coming weeks.
Rate Limits
Access to GPT models through the OpenAI API is subject to rate limits, meaning there is a maximum number of requests allowed per month. OpenAI has provided clear guidelines on these rate limits to ensure that applications do not face unexpected interruptions. Furthermore, the rate limits for GPT-4 have been doubled.
As GPT-4 Turbo is currently in the preview phase, the rate limits are set at 20 requests per minute and 100 requests per day. OpenAI may consider adjusting these limits once a public version becomes available.