The launch of ChatGPT captivated users with its impressive natural language capabilities, relying on the previously established GPT-3.5 large language model. However, the arrival of the much-anticipated GPT-4 transformed expectations for AI, earning recognition as an early glimpse into artificial general intelligence (AGI).
What is GPT-4?
GPT-4 is OpenAI's latest language model, capable of generating text that closely resembles human speech. This model enhances the technology harnessed by ChatGPT, transitioning from GPT-3.5 to a more advanced system. "Generative Pre-trained Transformer," or GPT, refers to the deep learning architecture that utilizes artificial neural networks to emulate human-like writing.
OpenAI highlights that GPT-4 surpasses ChatGPT in three crucial areas: creativity, visual input, and handling longer contexts. In terms of creativity, GPT-4 excels in generating and collaborating on projects across various formats, including music, screenplays, technical writing, and even adapting to a user's unique writing style.
The ability to process longer contexts is significant as well. GPT-4 can handle up to 128,000 tokens of user input and can even interact with text from web links, which enhances its capabilities in producing long-form content and sustaining extended conversations.
Additionally, GPT-4 can process images, allowing it to respond to visual prompts. For example, it can analyze a photo of baking ingredients and suggest recipes based on what it sees. However, it remains unclear if video input is supported in the same manner.
Importantly, GPT-4 has been designed to improve safety significantly compared to its predecessor. Internally, it reportedly provides 40% more factual responses and is 82% less likely to engage with requests for disallowed content. This advancement is based on extensive training incorporating human feedback and collaboration with over 50 experts, particularly in AI safety and security.
In the weeks following its launch, users showcased remarkable applications of GPT-4—notably, creating new languages, designing complex animations, and even programming a functioning version of Pong in just sixty seconds using HTML and JavaScript.
How to Use GPT-4
GPT-4 is accessible to all users across OpenAI's subscription tiers. Free-tier users have limited access—approximately 80 chats within a three-hour window—after which they are transitioned to the less capable GPT-4o mini until the cooldown resets. To enhance access to GPT-4 and generate images with DALL-E, users can opt for the ChatGPT Plus subscription at $20 per month. Upgrading is straightforward: simply click on “Upgrade to Plus” in the sidebar within ChatGPT, enter credit card details, and toggle between GPT-4 and earlier language models.
For those hesitant to subscribe, Microsoft’s Bing Chat offers a way to experience GPT-4's capabilities for free. Microsoft integrates GPT-4 into Bing Chat, although some features may be missing and the service includes Microsoft's proprietary enhancements. While Bing Chat remains free, it is limited to 15 chats per session and 150 sessions daily.
A variety of other applications are leveraging GPT-4, including Quora, a popular question-and-answer platform.
When Was GPT-4 Released?
GPT-4 was officially unveiled on March 13, following pre-release confirmation from Microsoft. Initially available to ChatGPT Plus subscribers and through Microsoft Copilot, GPT-4 is also accessible as an API for developers. Companies such as Duolingo, Be My Eyes, Stripe, and Khan Academy have already integrated GPT-4 into their services. The first public demonstration of GPT-4's capabilities was live-streamed on YouTube.
What is GPT-4o Mini?
GPT-4o mini is the latest iteration of OpenAI’s GPT-4 model line, optimized for simpler, high-volume tasks that prioritize quick inference speed over the full model's capabilities. Released in July 2024, GPT-4o mini has replaced GPT-3.5 as the default model once users exceed their three-hour limit of queries with GPT-4o.
Is GPT-4 Better Than GPT-3.5?
Previously, the free version of ChatGPT was based on GPT-3.5. However, as of July 2024, ChatGPT operates on GPT-4o mini, which demonstrates superior performance even compared to GPT-3.5 Turbo. It comprehensively understands and responds to more inquiries, offers additional safeguards, provides concise answers, and is 60% more cost-effective to operate.
The GPT-4 API
For developers, GPT-4 is available as an API, provided they have made at least one successful payment to OpenAI previously. The API encompasses various GPT-4 versions alongside legacy GPT-3.5 models. OpenAI announced that while GPT-3.5 would remain accessible, it would eventually be phased out, although a specific timeline for this transition is undecided.
The API is primarily designed for developers creating new applications, but some users have experienced confusion. For instance, Plex facilitates integrating ChatGPT into its Plexamp music player, requiring a separate ChatGPT API key distinct from ChatGPT Plus. Therefore, signing up for a developer account is necessary to access the API.
Is GPT-4 Getting Worse?
While GPT-4 initially garnered praise upon release, some users have reported a decline in response quality over subsequent months. Observations from notable figures in the development community have surfaced, leading to discussions in OpenAI's forums. An OpenAI executive claims these concerns are largely unfounded, arguing that any perceived decline is merely subjective and stating that newer versions of the model are consistently improved.
However, a study indicates there may be truth to these concerns, as it noted a decrease in accuracy from 97.6% in March to 2.4% by June. While this finding isn’t conclusive, it supports claims that users are experiencing real changes in performance.
Where is the Visual Input in GPT-4?
One of the most anticipated features of GPT-4 is its ability to process visual input, transforming ChatGPT into a truly multimodal model. Uploading images for analysis is as simple as attaching documents; users need only click the paperclip icon in the context window, select the image source, and attach the image to receive meaningful interactions.
What are GPT-4’s Limitations?
Despite OpenAI's claims about GPT-4’s advancements, the model still faces certain limitations. Like its predecessors, GPT-4 grapples with issues related to social biases, inaccuracies, and challenges posed by adversarial prompts. Consequently, it is not infallible. Numerous instances online highlight these shortcomings, yet OpenAI asserts that it continues to work on resolving such issues. Overall, GPT-4 is less prone to inventing information compared to earlier models.
Another notable limitation is that GPT-4's training data only covers information available until December 2023, whereas GPT-4o and 4o mini cutoff at October 2023. Despite this, GPT-4’s web search capabilities allow it to locate and retrieve newer information from the internet. With GPT-4o now released, users can anticipate the forthcoming GPT-5 model.
Ranking
