We saw the emergence of a new Generative AI behemoth, OpenAI, with the introduction of ChatGPT on November 2022. ChatGPT serves as a chat-oriented interface for the underlying model it employs. The first version of the chatbot relied on a large language model called GPT-3.5, which sure blew all our minds.
But on Tuesday 14th of March 2023, 4 months after the first launch of ChatGPT, OpenAI is pepping things up with the release of GPT-4, a more capable model than previous versions. Initially designed for engaging with GPT-3.5, the chatbot now functions as a medium to interact with GPT-4.
https://twitter.com/sama/status/1635687853324902401?embedable=true
What can GPT-4 do?
The same thing that GPT-3.5 can do, but more. The company said in a blog post,
“GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5.”
Since 2018, OpenAI has been disseminating GPT language models, but of all Open AI's huge language models, GPT-4 is not just the newest, but also the most potent. OpenAI reported that it used Microsoft Azure to train the model. Here are a couple of ways GPT-4 compares to GPT- 3.5
-
GPT-4 accepts visual inputs, and GPT-3.5 doesn’t
Given that GPT-4 is "multimodal," as specified by OpenAI, and can comprehend many "modalities" of information, it can be prompted to produce content by both visual and textual cues while GPT-3.5 was essentially able to read and write text. Nevertheless, the company is not deploying its image-description capability immediately, and the GPT-4 version remains available only to subscribers of OpenAI's ChatGPT Plus users and developers, and only supports text.
-
Compared to GPT-3.5, GPT-4 has a larger buffer
Simply put, GPT-4 has a longer memory than the previous model. With GPT-3.5 and the previous iteration of ChatGPT, the maximum was 4,096 "tokens," or roughly 8,000 words, while the maximum token count for GPT-4 is 32,768, or around 64,000 words. That’s 8 times more memory capacity.
-
GPT-4 has a variety of "personas." or steerability, unlike GPT-3.5
Steerability refers to the ability to control or influence the output of a language model in a specific direction or style. In other words, it allows users to guide the language model to produce text that fits a particular tone, style, or topic. While GPT-3 does not have built-in steerability features, GPT-4 does. With GPT-4, developers and ChatGPT users can specify their AI's style and task by giving such directions in the "system" message, as opposed to the traditional ChatGPT personality with a defined verbosity, tone, and style. The company stated that although this feature is currently only accessible to developers, it will soon be made available to ChatGPT users as well.
-
GPT-4 has a broader language range.
This means that GPT-4 can understand and generate text in a wider variety of languages than its predecessor GPT-3.5. This is a significant advancement in the field of natural language processing. According to OpenAI, GPT-4 beats the English-language performance of GPT-3.5 and other LLMs in 24 out of the 26 languages they tested. As a result, GPT-4 can respond accurately to a large number of multiple-choice questions in 26 different languages.
-
GPT-4 has lesser ‘hallucinations’ on facts
Even while it does not fully stop hallucinations or making up facts, GPT-4 considerably lessens them in comparison to earlier variants. The company said the updated language model continues to "hallucinate" facts and exhibit flawed logic, though not as bad as GPT-3.5. Hallucinations in the context of language model software refer to the generation of text that appears to be coherent and meaningful but is not actually supported by the input data. This can occur when a language model generates text that is not grounded in the context or is not consistent with the semantic meaning of the input text.
In other words, a language model may sometimes generate text that is not accurate or plausible based on the information it has been trained on, resulting in what is commonly referred to as "hallucinations." The model also incorporates social biases and adversarial prompts in addition to the above limitations. OpenAI’s blog post stated,
“GPT-4 scores 40% higher than the latest GPT-3.5 on internal adversarial factuality evaluations.”
According to The Washington Post, LLMs have learned to mimic human prejudices of race, gender, religion, and class as a result of their training on internet text and visuals. Evidently, AI systems have the potential to perpetuate and amplify existing biases in society, rather than helping to mitigate them.
GPT-4 also has additional restrictions similar to those found in GPT-3.5, one of which is that it still lacks knowledge of events that occurred after September 2021. The ability of others to teach it new things is also limited by the fact that it does not learn from its experiences. Regardless, GPT-4 is still smarter; as it is more accurate, more precise, and has better expertise.
The language model has also shown that it is capable of acing exams. According to OpenAI, it scored at the 90th percentile on a mock bar exam, the 93rd percentile on an SAT reading test, and the 89th percentile on an SAT math test.
Where is GPT4 currently being used?
GPT-4 has the potential to be used in a wide range of applications that require natural language understanding and generation. The language-learning software Duolingo, the payment provider Stripe, the online learning platform Khan Academy, and the financial services company Morgan Stanley were among the companies mentioned in OpenAI's statement that are currently using the new model.
Microsoft also reported that its chatbot, BingAI has been using GPT-4 all along, even before the model's public release.
“We are happy to confirm that the new Bing is running on GPT-4, which we’ve customized for search, If you’ve used the new Bing preview at any time in the last five weeks, you’ve already experienced an early version of this powerful model.”
Yusuf Mehdi, corporate vice president, and chief marketing officer for consumers at Microsoft, wrote in a statement.
https://www.youtube.com/watch?v=hdhZwyf24mE&embedable=true
GPT-4 "excels in tasks that require advanced reasoning, complicated instruction understanding, and more creativity," according to OpenAI.
Users have reported using GPT-4 in a variety of inventive ways in the brief time since its release:
https://twitter.com/skirano/status/1635736107949195278?s=20&embedable=true
https://twitter.com/YosarianTwo/status/1635780666632687617?s=20&embedable=true
The possibilities for this new tool are believed to be endless.
https://twitter.com/mckaywrigley/status/1635740909383061504?s=20&embedable=true
How can GPT-4 be accessed?
Fundamentally, to access ChatGPT, you must first register on the OpenAI
Since Microsoft has confirmed that the new Bing runs on the new model of GPT, it is obvious then that Millions of Bing users can already access it through the chatbot.
https://twitter.com/JordiRib1/status/1635694953463705600?embedable=true
The new model is also accessible via API, the company stated that they are prioritizing API access for developers that give outstanding model assessments to OpenAI Evals during the slow rollout of GPT-4 in order to learn how they may enhance the model for everyone.