OpenAI has kept itself in the headlines. The creators of ChatGPT and their attempts to integrate and revolutionize the way we interact with machines and the world around us with the use of artificial intelligence show no sign of slowing down anytime soon as OpenAI announces enthralling AI language model GPT 4.
Also Read: OpenAI Chatgpt 4 Will Let You Turn Text Into Video
After all the hearsay and speculations, Open AI released its latest and most advanced till date AI language model that streamlines human-level capabilities and responds to the inputs of images and texts. This new line up gears applications like Chat GPT and its latest addition to Bing.
Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment: https://t.co/TwLFssyALF pic.twitter.com/lYWwPjZbSg
— OpenAI (@OpenAI) March 14, 2023
“We’ve created GPT-4, the latest milestone in OpenAI’s effort in scaling up deep learning. GPT-4 is a large multimodal model (accepting image and text inputs, emitting text outputs) that, while less capable than humans in many real-world scenarios, exhibits human-level performance on various professional and academic benchmarks,” reads the announcement blog post.
As claimed by the developers, this ingenious model is more function and capacitive, and also more accurate than before. OpenAI stated about their efforts to scale up the advancements by saying that they spent six months into “iteratively aligning” GPT 4 using the lessons from an internal adversarial testing program as well as ChatGPT and that resulted in an unprecedentedly stable version of the eminent model.
“We’ve spent 6 months iteratively aligning GPT-4 using lessons from our adversarial testing program as well as ChatGPT, resulting in our best-ever results (though far from perfect) on factuality, steerability, and refusing to go outside of guardrails.”
OpenAI worked with Microsoft and developed a Super Computer from the ground to the Azure cloud to train GPT 4. In addition to that, OpenAI also revealed its partnership with a number of companies to integrate GPT 4 into their products, some of which include Duolingo, Stripe, and Khan Academy.
Where does GPT 4 stand out in comparison to GPT 3.5 for ChatGPT?
1. GPT 4 sees and understands images: This is the most anticipated and talked about change to the language model. Thanks to the fact that OpenAI partnered with Be My Eyes for this. As it is a “multimodal” learning system, it understands more than one form of information. Thus, in addition to responding to texts just like its predecessor, this year it takes a leap and processes images as well. This includes offering a recipe just by seeing a picture of the ingredients or explaining the proper form to use a machine by looking at the pic, and more. However, it only responds via text.
2. Hard to fool: “Refusing to go outside of guardrails”, this is what OpenAI claims for GPT 4. Based on their claims, GPT 4 cannot be coaxed and ventured hostile. It can handle any form of nuanced instructions in comparison to GPT 3.5. This has been trained on lots of malicious prompts using the data given by people to OpenAI over the past two years.
3. Longer Memory: Language models of GPT 4 have been trained on a substantial learning stack of web pages, books, and other text data. The academic benchmarks of the model result in a significant improvement in performance, as it scored in the top 10% of test takers for a simulated bar exam, in comparison to GPT 3.5 which scored in the bottom 10%. Moreover, when it comes to saving in mind, the limits are comparatively significant over GPT 3.5. The former one had a limit of 4096 tokens (nearly 8000 words or 4 to 5 pages of a book) but the latest translates around 64000 words or 50 pages of text.
4. Multilingual Efficiency: GPT 4’s Massive Multitask Language Understanding benchmark consisted of a suite of 14,000 multiple-choice problems that covered almost 57 subjects in 26 languages. This model performed superiorly in 24 out of 26 languages, including Italian, Latvian, Korean, Nepali, Urdu, and many more, against GPT 3.5 English-language performance.
5. Behaviorial Attributes: Open AI claims to have worked on the steerability aspect of GPT 4. Machine language models are equipped to alter their behavior and personalities on demand. This further enhances the role of a machine in synchronization with reality, often acting as a warm companion and sometimes a demonic ruler. The AI’s style, verbosity, tone, and task can now be prescribed by users in a more native way than GPT 3.5. Moreover, API users would also be able to customize it for their users’ experience but within bounds.
GPT 4 is open to OpenAI’s premium users via the $20 per month subscription to ChatGPT Plus as well as in API format for developers who can currently sign up for the waitlist. As far as pricing is concerned, it stands at $0.03 per 1000 “prompt” tokens (nearly 750 words) and $0.006 per 1000 “completion” tokens (nearly 750 words). Tokens represent raw text, prompt tokens are parts of words served to the model, and completion tokens are the content generated.
Ultimately, there is a lot that needs to be discussed, and the developers have been considerate of the limitations and risks of GPT 4 as well. They still keep their users in the loop with the statement that even though GPT 4 reduces hallucinations in comparison to GPT 3.5, their models are not fully reliable. They can hallucinate facts and produce errors, so should be taken care of when being used. Moreover, even after rigorous training in selecting, filtering, enforcing evaluation, and monitoring, harmful advice or incorrect information can still be generated.
Recommended: What is the Role of Artificial Intelligence in Cyber Security?
Now that we are slowly pacing towards more advanced forms of machine learning technologies and their applications in real life, things need to be taken in conscience, as there is always a better output backed by an unwanted consequence. That being said, as OpenAI announces enthralling AI language model GPT 4, we are highly charged to witness where it leads to in the form of innovations. Let us know your thoughts in the comments box below.
Source: Open AI Blog
