OpenAI has finally unveiled GPT4, a formidable new AI model with image and text understanding capabilities, marking the latest breakthrough in the company's journey to scale up its flagship generative pre-trained transformer model. Excitingly, GPT4 is already available to ChatGPT PLUS subscribers with a usage cap, while developers can join a waitlist for API access.
 |
| GPT-4 AI Breakdown How Smart vs AGI GPT-5 + IQ TEST + PRICE + NEW FEATURES |
As for its pricing, the cost is 3 cents per 1,000 prompt tokens, or approximately 750 words, and 6 cents per 1,000 completion tokens, which is also around 750 words. And as it turns out, GPT4 is already being used by Microsoft to power its Bing Chat technology as well as other large tech names, including Stripe, Duolingo, Morgan Stanley, and Khan Academy, powering a wide range of business and consumer applications.
Additionally, in a significant enhancement over its GPT 3.5 predecessor, GPT4 can process both image and text inputs and is able to generate text at a human-like level across a range of professional and academic benchmarks.
For instance, it massively outperforms its predecessor, scoring in the top 10% on a simulated bar exam, which is the licensure test to become a lawyer, with GPT 3.5 only managing to score in the bottom 10%. Over a six-month period, OpenAI meticulously refined GPT4 based on findings from an internal adversarial testing program and ChatGPT, achieving unprecedented results in 40% higher rate of factuality, steerability, and adherence to guardrails, with GPT4 being trained on publicly available data such as public web pages and licensed data.
This was made possible by the collaboration between OpenAI and Microsoft to create a custom supercomputer in the Azure Cloud, which was used to train GPT4.
What's incredible is that GPT4's multimodal capabilities to understand text and image inputs enable the model to accomplish a range of new tasks across a jaw-dropping number of domains.
For instance, due to its improved accuracy and understanding, GPT4 can now write functional code for applications like Discord bots while simultaneously identifying and resolving error codes along the way.
In another impressive example, GPT4 can even convert hand-drawn sketches of a website layout into a functional site, complete with working HTML and JavaScript.
In addition to these technical feats, GPT4 can even tackle complex mathematical tasks that include reasoning like deciphering intricate tax codes and performing complicated calculations.
With its almost infinite creativity, GPT4 is an invaluable tool that is now capable of providing innovative solutions that are pushing the boundaries of artificial intelligence.
When compared to GPT3.5, its incredible level of performance becomes extremely obvious when tasks increase in complexity, with GPT4 displaying greater reliability, creativity, and adaptability to nuanced instructions, as well as its ability to handle 25,000 characters as input, whereas GPT3 only took 3,000.
Furthermore, GPT4's ability to understand images in addition to text is a breakthrough feature that will forever change the lives of visually impaired individuals.
For instance, GPT4 can not only caption complex images but also identify specific items, such as a lightning cable adapter from a photo of a connected iPhone. This image understanding capability is currently being tested with a single partner called Be My Eyes through their new virtual volunteer feature to assist the blind, but other applications are potentially limitless, such as in the areas of accessibility services, social media content moderation, creative design applications, and much more.
But that's not all, because another notable leap forward is the model steerability, with OpenAI Introducing System Messages as a new API Capability
To allow developers to provide specific instructions on style and task, we are introducing System Messages. These messages set the tone and establish boundaries for AI interactions and will be introduced to ChatGPT in the future. This opens up a wealth of new opportunities for developers to create highly customized artificial intelligence-driven applications, allowing them to tailor the model's interactions and responses to specific use cases, leading to more efficient and user-friendly AI solutions across a range of sectors.
Despite these advancements, OpenAI acknowledges that GPT4 still has imperfections, such as hallucinating facts or making reasoning errors with high confidence. The AI model may also lack knowledge of events occurring after the majority of its data cuts off in September 2021, resulting in simple errors. However, the company has highlighted massive improvements in several key areas, such as GPT4 being less likely to respond to dangerous requests. In fact, it's 82% less likely to generate disallowed content compared to ChatGPT and adheres to safety policies 29% more often regarding sensitive requests like medical advice and violence, meaning that jailbreaking it won't be easy.
Finally, it's noted that as time goes on, the model will continue to improve, meaning that users will enjoy increasing functionality like gaining access to the soon-to-be-released image input feature, as well as increased accuracy and speed.
Despite the progress made with GPT4, OpenAI remains vigilant about the potential drawbacks and challenges posed by AI technology. The company is committed to mitigating risks and maintaining a responsible approach to AI development, as evidenced by its ongoing efforts to involve the public in the decision-making process and the emphasis on transparency.
But what about GPT5?
The next iteration of OpenAI's groundbreaking model could possess a host of new and advanced abilities that further revolutionize the field of artificial intelligence. Building upon GPT4's text and image understanding capabilities, GPT5 could potentially understand and process multimedia content such as video and audio, enabling it to analyze and generate contextually accurate transcriptions, translations, and even video summaries. One can envision GPT5 developing an advanced level of context awareness, which would allow it to engage in more natural, intuitive, and meaningful conversations while continuously learning from its interactions. This would enable GPT5 to adapt and evolve, providing even more personalized and effective responses.
GPT5 could also enhance its problem-solving capabilities, addressing complex challenges across a wide range of domains, such as medical diagnostics, scientific research, and policy analysis. By integrating vast amounts of domain-specific knowledge and reasoning abilities, GPT5 could assist experts in making more informed decisions and accelerate the discovery of new insights. Furthermore, GPT5 might possess the ability to collaborate with other AI models, combining their unique strengths to create a synergistic effect, resulting in more sophisticated and efficient solutions.
As the potential capabilities of GPT5 unfold, we can only imagine its far-reaching impact on society and the future of artificial intelligence.
So, this brings us to our last question: Could GPT4 qualify as an artificial general intelligence? Well, this depends on which definition of AGI you're using. According to the requirements of the Turing test, some people could argue that the model already qualifies as AGI. In summary, the Turing test requires GPT4 to answer questions meaningfully enough to trick a human into believing they're talking to another human. To truly answer this question, you probably have to ask yourself: Could GPT4 fool you into thinking it was a human when reading its answers to your questions?
GPT5 could also enhance its problem-solving capabilities, addressing complex challenges across a wide range of domains, such as medical diagnostics, scientific research, and policy analysis. By integrating vast amounts of domain-specific knowledge and reasoning abilities, GPT5 could assist experts in making more informed decisions and accelerate the discovery of new insights. Furthermore, GPT5 might possess the ability to collaborate with other AI models, combining their unique strengths to create a synergistic effect, resulting in more sophisticated and efficient solutions.
As the potential capabilities of GPT5 unfold, we can only imagine its far-reaching impact on society and the future of artificial intelligence.
So, this brings us to our last question: Could GPT4 qualify as an artificial general intelligence? Well, this depends on which definition of AGI you're using. According to the requirements of the Turing test, some people could argue that the model already qualifies as AGI. In summary, the Turing test requires GPT4 to answer questions meaningfully enough to trick a human into believing they're talking to another human. To truly answer this question, you probably have to ask yourself: Could GPT4 fool you into thinking it was a human when reading its answers to your questions?