What is GPT-4? & How Does It Differ From ChatGPT

What is GPT-4? & How Does It Differ From ChatGPT

OpenAI’s Generative Pre-trained Transformer 4 (GPT-4) is a multimodal big language model and the fourth in the GPT series. It was published on March 14, 2023, and is now publicly available in a restricted version through ChatGPT Plus, with access to its commercial API available through a waitlist. GPT-4 was pre-trained as a transformer to forecast the next token (using both public and “data licensed from third-party sources”), and was then fine-tuned using reinforcement learning from human and AI input for human alignment and policy compliance.

Microsoft revealed that GPT-4 was used in earlier versions of Bing before its official release.

According to OpenAI’s blog article unveiling GPT-4, it is “more trustworthy, creative, and capable of handling significantly more complicated instructions than GPT-3.5.” The group created two versions of GPT-4 with context windows of 8192 and 32768 tokens, which was a substantial advance over GPT-3.5 and GPT-3, which had context windows of 4096 and 2049 tokens, respectively. Unlike its predecessors, GPT-4 can accept both pictures and text as input; this allows it to describe the comedy in strange photos, summarise screenshotted material, and answer test problems using diagrams. Despite these additional capacities, GPT-4, like its forefathers, has the propensity to hallucinate solutions.

GPT-4 is “more trustworthy, innovative, and capable of processing substantially more intricate instructions than GPT-3.5,” according to OpenAI’s blog post announcing it. The group developed two versions of GPT-4 with context windows of 8192 and 32768 tokens, which represented a significant improvement over GPT-3.5 and GPT-3, which had context windows of 4096 and 2049 tokens, respectively. GPT-4, unlike its predecessors, can accept both pictures and text as input; this enables it to describe the humor in unusual photos, summarise screenshotted information, and answer exam questions using diagrams. Notwithstanding these enhancements, GPT-4, like its forebears, has a tendency to hallucinate solutions.

The firm announced the language model’s capabilities in its announcement blog, claiming that it is more creative and collaborative than ever before. Although GPT-3.5-powered ChatGPT only accepts text inputs, GPT-4 can also produce captions and analysis from photos. Nevertheless, this is simply the top of the iceberg. We describe the new language model, what it isn’t, and how it outperforms its predecessor.

Apart from the new capacity to interpret pictures, OpenAI claims that GPT-4 “demonstrates human-level performance on a variety of professional and academic standards.” Because of its better general knowledge and problem-solving abilities, the language model can pass a simulated bar exam with a score in the top 10% of test takers and handle tough questions with more accuracy.

It can, for example, “answer tax-related inquiries, plan a meeting between three busy individuals, or learn a user’s creative writing style.”

GPT-4 can also handle over 25,000 words of text, expanding the number of use cases to include long-form content creation, document search and analysis, and lengthy dialogues.

Microsoft researchers evaluated GPT-4 on medical issues and discovered that “GPT-4, without any specialist prompt crafting, outperforms prior general-purpose models (GPT-3.5) as well as models specially fine-tuned on medical knowledge (Med-PaLM, a prompt-tuned variant of Flan-PaLM 540B)”.

What distinguishes GPT-4 from ChatGPT?

Following are some of the significant distinctions:

*GPT-4 can now see images: The most visible difference in GPT-4 is that it is multimodal, meaning it can grasp more than one type of input. The GPT-3 and ChatGPT’s GPT-3.5 were only capable of textual input and output, which meant they could only read and write. GPT-4, on the other hand, may be given pictures and told to produce data.

It’s understandable if this reminds you of Google Lens. Lens, on the other hand, exclusively looks for information connected to a picture. GPT-4 is far more advanced in that it comprehends and analyses images. OpenAI offered an example of the language model conveying the joke in the form of an outrageously big iPhone connection. The only caveat is that picture inputs are still in the early stages of study and are not yet publicly available.

GPT-4 is more difficult to deceive: One of the most significant disadvantages of generative models such as ChatGPT and Bing is their proclivity to periodically go off the rails, creating prompts that raise eyebrows or, worse, frighten people. They can also jumble up facts and spread falsehoods.

According to OpenAI, it spent 6 months training GPT-4 using lessons from its “adversarial testing program” as well as ChatGPT, yielding the company’s “best-ever results on factuality, steerability, and refusing to venture outside of guardrails.”

GPT-4 has increased accuracy: OpenAI acknowledges that GPT-4 has the same constraints as earlier versions – it is still not completely dependable and makes logical mistakes. Nonetheless, “GPT-4 dramatically lowers hallucinations relative to earlier models” and performs 40% better on factuality ratings than GPT-3.5. It will be far more difficult to fool GPT-4 into producing undesired outcomes such as hate speech and disinformation.

GPT-4 understands languages other than English better: Because machine learning data is largely in English, as is the majority of content on the internet nowadays, training LLMs in other languages might be difficult.

Nevertheless, GPT-4 is more multilingual, and OpenAI has shown that it beats GPT-3.5 and other LLMs by correctly answering thousands of multiple-choice questions in 26 languages. It clearly handles English the best, with an accuracy rate of 85.5 percent, but Indian languages like Telugu aren’t far behind, with a rate of 71.4 percent. This implies that users will be able to utilize GPT-4-based chatbots to provide outputs in their native languages with increased clarity and accuracy.

GPT-4 has already been implemented for a variety of applications in products such as Duolingo, Stripe, and Khan Academy. While it is not yet free for all, a $20 per month ChatGPT Plus subscription can get you immediate access. Nonetheless, ChatGPT’s free tier is still based on GPT-3.5.

If you don’t want to pay, there is an ‘unofficial’ option to start utilizing GPT-4 right away. Microsoft has confirmed that the new Bing search experience now operates on GPT-4 and is available at bing.com/chat.

ChatGPT Plus gives you access to a GPT-4-backed version of ChatGPT for $20 per month, whereas the regular version is GPT-3.5-backed. OpenAI also makes GPT-4 available to a select group of applicants via their GPT-4 API waitlist; once accepted, an additional fee of 0.03 USD per 1000 tokens in the initial text provided to the model (“prompt”) and 0.06 USD per 1000 tokens generated by the model (“completion”) is required to use the model with an 8192-token context window; the prices are doubled for the 32768-token version.

Everything You Need to Know About ChatGPT

Top 10 ChatGPT Alternatives You Can Use in 2023

How Does ChatGPT Make Money