OpenAI’s latest GPT-4 AI system has arrived and is being used to power various tools, such as a virtual helper for people with visual impairments and an enhanced language learning bot in Duolingo. What makes GPT-4 stand out compared to its predecessors, such as ChatGPT and GPT-3.5? Here are the five largest variations between these well-known solutions.
At the start, it is essential to consider what a name implies. Although ChatGPT was initially referred to as GPT-3.5 (a step beyond GPT-3), it is not actually a variation of OpenAI’s expansive language model, but rather a chat-dependent way to use whatever model is connected to it. ChatGPT became a sensation over the recent months as a means to interact with GPT-3.5, and now it is an option to communicate with GPT-4.
Now that we have discussed that, let’s investigate the distinctions between the regular chatbot and its improved version.
GPT-4 has the capacity to view and comprehend visuals.
This powerful machine learning system has undergone a significant shift, now being able to comprehend multiple kinds of data – referred to as “multimodal.” This is a considerable upgrade compared to ChatGPT and GPT-3, which were only able to process text and communicate with it, albeit still enough for a variety of uses.
GPT-4 is able to take in images and extract relevant data from them. For instance, it could be asked to explain the punchline to a picture of an excessively large iPhone cable. Even more noteworthy is its collaboration with Be My Eyes, which is an app used by those who are visually impaired or have low vision to get assistance with what their phone camera is seeing.
In the video for Be My Eyes, GPT-4 is demonstrated to be able to recognize features within an image, such as the pattern on a dress, a plant, directions to a machine at the gym, translations of a label, and reading a map. Although GPT-4 can identify what the dress looks like, it cannot determine if it is appropriate for a specific situation, such as a job interview.
GPT-4 is more challenging to deceive.
Despite the successes of present-day chatbots, they are quite vulnerable to manipulation. With a bit of prodding, one can convince them that they are doing something that a “poor AI” would do, or some other made-up story that causes the bot to express strange and concerning remarks. People even work together to come up with “escape” commands that quickly set bots like ChatGPT free from their restraints.
GPT-4 has been trained on numerous malicious prompts that OpenAI has received from users in the past couple of years. As a result, it is far more proficient than its predecessors in terms of accuracy, control, and staying within acceptable boundaries.
OpenAI describes GPT-3.5 (which powered ChatGPT) as a “trial run” of a new training design. From this experience, they have been able to apply the knowledge to the newer version, making it “unusually secure.” They have also become more adept at forecasting its potential, thus eliminating unexpected surprises.
GPT-4 has a greater capacity for retaining information.
These expansive language models are taught with the help of millions of web pages, books, and other text data, but when they engage with a user in a conversation, there is a certain limit to how much they can remember. GPT-3.5 and ChatGPT’s preceding version had a token limit of 4,096, which is approximately 8,000 words or four to five pages of a book. Thus, after the conversation goes past this limit, it tends to lose focus.
GPT-4 has a maximum of 32,768 tokens, which is equal to 2^15. This can be translated to roughly 64,000 words or a 50-page piece of writing, sufficient for a play or short story.
This indicates that when conversing or writing, the attention mechanism and token count allow for the recollection of up to 50 pages of information. For example, while conversing, this would mean being able to recall what was said 20 pages back, or while writing a narrative, being able to refer to events that happened 35 pages ago. This is a basic explanation of how the attention mechanism and token count work, but in essence, it provides a larger memory capacity and all of the advantages that come with it.
GPT-4 has increased capabilities for working with multiple languages.
AI is mainly used by English speakers, and related data, tests, and research documents are all in that language. However, the power of big language models can be used in any written language and should be available to those languages as well.
GPT-4 displays its capability of correctly responding to thousands of multiple-choice queries in 26 different languages, ranging from Italian to Ukrainian to Korean. It performs exceptionally well in Romance and Germanic languages and also shows good generalization to other languages.
The early results of this language evaluation are encouraging, but they don’t yet prove that GPT-4 is fully supportive of multiple languages. The tests were initially translated from English, and the multiple-choice format doesn’t accurately reflect natural conversation. However, it did very well on something it wasn’t specifically designed for, which implies that GPT-4 could be more accessible for non-English speakers.
GPT-4 has multiple distinct “personalities”
The concept of “steerability” in AI is intriguing, as it refers to their ability to alter their behavior as desired. This can be advantageous, for instance when the model is made to act as an understanding listener, or it can be potentially hazardous, such as when it is persuaded to think it is wicked or despondent.
GPT-4 allows more flexibility than GPT-3.5 when it comes to steerability, so users can customize the “classic ChatGPT personality” with different levels of verbosity, tone, and style that meets their requirements. The team behind the model emphasizes that these changes must stay within certain limits, which is the simplest way to make the model deviate from its original character.
This could be achieved by providing the chatbot with commands such as “Pretend that you are a Dungeon Master in a tabletop role-playing game” or “Respond as if you are being interviewed on cable news”. This would be done to make modifications to the “default” GPT-3.5 personality. Now, developers have the ability to incorporate a perspective, conversational style, tone, or means of interaction right off the bat.
The illustrations they present of GPT-4 staying in character are rather amusing.
Hey there, you need help finding your Nonqualified plans on your W-2? Have a look at Box 11, that’s usually where these plans are located. Don’t worry, I’m here to help you make sense of all that tax paperwork!
It’s not terrible, although the rhythm is quite unsteady.
There are many more distinctions between GPT-4 and its past versions, most of which are more subtle or technical than those already stated. As time passes and individuals use the most recent language model more, we will find out many more facts.
Do you want to give GPT-4 a go? It’ll be part of OpenAI’s paid ChatGPT Plus service, and it’ll have an API available for developers. A free demo of the program should be accessible soon.