GPT-4: What We Know So Far About OpenAI’s Latest Model

OpenAI recently launched its latest artificial intelligence (AI) model, GPT-4. Since its release, the tech community has been buzzing with excitement about its potential capabilities and implications for the future of AI technology.  

The AI company considers the ChatGPT successor its “most advanced system” as it produces “safer and more useful responses.” Sam Altman, CEO of OpenAI, believes that GPT-4 and AI have the potential to be “the greatest technology humanity has yet developed.” 

Here is what we know so far about the latest model: 

What can GPT-4 do? 

GPT-4's biggest difference from its predecessor is its ability to process both text and image inputs.  

For example, users can give the language model images of their ingredients and it will respond with specific recipe suggestions.  

GPT-4 also showcases improved creative capabilities - making music, screenplays and technical writing for its user, as well as the ability to learn their writing style accurately. Users may also send GPT-4 a web link and it can interact with the text present on the page. 

One Twitter user even prompted the AI model to “make as much money as possible” with only a capital of $100. 

I gave GPT-4 a budget of $100 and told it to make as much money as possible.

I'm acting as its human liaison, buying anything it says to.

Do you think it'll be able to make smart investments and build an online business?

Follow along 👀 pic.twitter.com/zu4nvgibiK

— Jackson Greathouse Fall (@jacksonfall) March 15, 2023

Another notable feature is GPT-4's ability to process 25,000 words, or around eight times the number of words ChatGPT can process.  

When it comes to taking tests, GPT-4 presented “human-level performance,” scoring in the 90th percentile in a Uniform Bar Exam. Meanwhile, ChatGPT only garnered an average of C+ in law exams when the University of Minnesota professors tested its abilities.   

GPT-4 nearly perfected an SAT Math test, and can also write computer code. In a demo showcase of the language model, the demonstrator uploaded an image of his drawing and prompted GPT-4 to turn it into a functioning website. The language model sent back a code which was pasted into a preview showing how the imagined website could look like. 

With a mix of HTML and JavaScript, another user commanded the language model to create a classic game of Pong, and it was able to do so in under 60 seconds. 

I don’t care that it’s not AGI, GPT-4 is an incredible and transformative technology.

I recreated the game of Pong in under 60 seconds.
It was my first try.

Things will never be the same. #gpt4pic.twitter.com/8YMUK0UQmd

— Pietro Schirano (@skirano) March 14, 2023

How safe is GPT-4 compared to ChatGPT?  

OpenAI reported that GPT-4 produces 40% more factual responses and is 82% less prone to “respond to requests for disallowed content” than its previous iteration.  

The AI company also took six months “iteratively aligning” refining its responses and safety features, working with over 50 experts for early feedback.  

Despite this, OpenAI warns users that the new generative AI is still vulnerable to “hallucinations,” making reasoning errors and factually incorrect statements – issues the company is working to address. 

"The thing that I try to caution people the most is what we call the 'hallucinations problem... the model will confidently state things as if they were facts that are entirely made up,” Altman explained. 

The CEO also acknowledged that AI models could fall into the wrong hands, fearing it will be used for “large-scale disinformation.”  

“Now that they're getting better at writing computer code, [they] could be used for offensive cyberattacks," he added.  

While he thinks that the AI models are still “very much in human control,” safety limits may be disregarded and society has a “limited amount of time” to figure out how to react to and regulate the new technology.  

Where is GPT-4 available?  

While OpenAI tests the language model and prepares it for a wider rollout, GPT-4 is available to paid subscribers of ChatGPT Plus for $20 a month.  

Interested users may also test it via Bing Chat, Microsoft’s free chatbot engine running on the new language model since its launch five weeks ago. However, Bing Chat is only limited to 15 chats per session and 150 sessions a day.  

An API version of the language model will also be available to developers via a waitlist to aid them in building applications and services. Companies such as Duolingo, Be My Eyes, Stripe and Khan Academy have already jumped on the wagon and integrated GPT-4 into their services.  

Microsoft said its chatbot search engine Bing Chat ha already been running on the new AI model since its launch five weeks ago.