Artificial intelligence has been taking the world by storm in the recent decade and the advent of ChatGPT was just the start. Released in late November 2022, ChatGPT took the world by storm with its ability to seamlessly create essays, compose music, write code, and so much more in a matter of seconds.
OpenAI has now released the latest version of its AI technology - the GPT-4. Although it was introduced only on Tuesday, GPT-4 is already making waves amongst developers and the public alike due to its heightened abilities that have drastically improved on its predecessor, the GPT-3.5.
OpenAI describes the large multimodal program as being less capable than humans in many real-world scenarios yet able to exhibit human-level performance on various professional and academic benchmarks. So, what exactly are the new features of this advanced protégé?
The Features of the GPT-4
In the program’s release post, OpenAI declared that it had been working with Microsoft in designing a supercomputer within the Azure infrastructure to train the GPT-4 model, and like the previous ChatGPT versions, the program was modeled using licensed and public internet data – most importantly the feedback from ChatGPT. The GPT-4 boasts several improved and unique capabilities:
Visual Interpretation
The term “multimodal” refers to the program’s ability to extrapolate information from more than one “modal” or form which is why a key new feature of the GPT-4 is its ability to accept image prompts as well as text prompts. This allows the program to interpret, analyze, and respond to any image inputted into its system.
OpenAI has partnered with Be My Eyes to run this feature first before it can be available to all OpenAI users. In a blog post, Be My Eyes explained how the feature works precisely, stating that “if a user sends a picture of the inside of their refrigerator, the Virtual Volunteer will not only be able to correctly identify what’s in it but also extrapolate and analyze what can be prepared with those ingredients.” They go on to explain that the tool can then offer several recipes for those ingredients and also send a step-by-step guide on how to make each of them.
Sourced from the BBC
In a video introduction to the Virtual Volunteer, it’s demonstrated how the app uses AI chatbot technology to identify plants, describe the pattern of a dress, read a map, translate labels, and even tell the user which buttons to press on a vending machine.
Reliability
OpenAI has improved its chatbot technology substantially in terms of content, stating that “GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5.”
The company has tried to make the program safer and more aligned through the selection and filtering of pretraining data, evaluations, and expert engagement to improve modal safety, monitoring, and enforcement.
The company has also been using GPT-4 internally – for support, sales, content moderation, and programming.
Multilingual Capabilities
In its testing and design, the company took to using a series of benchmarks that simulated exams designed for humans as well as those designed for machine learning models. In a step towards reaching a non-English audience, the program can successfully answer thousands of multiple-choice questions with high accuracy across 26 languages.
While the questions themselves only form an MMLU benchmark and cannot substitute natural speech, the results still show an improvement in making the AI program more globally accessible.
Steerable Changes
The concept of “steerability” refers to the AI’s ability to adapt to demands and behaviors. With the GPT-4, users won’t have to deal with the “fixed verbosity, tone, and style” used for older ChatGPT versions as the new program can be prescribed a new personality or style as prompted.
The GPT-4 features “system messages” that allow the user to customize the way the AI responds – almost like a role-playing exercise.
Sourced from OpenAI
The company was quick to add that these prompts are only usable within the bounds of the AI’s usage policies and that improvements are still being made to the feature as they are fully aware that it is “the easiest way to “jailbreak” the current model.”
Not Easily Manipulated
While ChatGPT took quite a hit in terms of how its system could be manipulated, GPT-4 has promised to be more of a force to be reckoned with. Many people tested the limitations of the ChatGPT program after its release, figuring out ways to get around its guardrails – from using ChatGPT to write malware to many other malicious prompts.
The company asserted that the GPT-4 has improved in factuality, steerability, and refusal to go outside of guardrails. OpenAI noted that they’ve “decreased the model’s tendency to respond to requests for disallowed content by 82% compared to GPT-3.5,” and that the GPT-4 responds to “sensitive requests” - such as medical advice and self-harm - following policies 29% more often. An example of this is in the GPT-4’s ability to refuse requests on how to synthesize dangerous chemicals.
Coding Games
According to Al Jazeera, some Twitter users have found a way to use GPT-4 to code entire video games already as well. The publication includes an example of a user recreating the game Snake without any prior JavaScript knowledge.
Sourced from Twitter
The Twitter user noted that while the game had some issues at first when prompted, the GPT-4 fixed them immediately.
Evaluated Progress
While OpenAI is confident in its newest technology, it accepts and embraces the need to constantly improve and assess its AI models. This is why the company open-sourced OpenAI Evals as a framework for automated evaluation of its AI model’s performance.
The feature allows anyone to report shortcomings to help guide further improvements. The program creates and runs evaluations that use datasets to generate prompts, measures the quality of completions provided by an OpenAI model, and compares performance across different datasets and models.
Limitations
As with everything created, GPT-4 is not without its drawbacks. OpenAI was the first to point out the limitations that its newest AI creation suffers from. According to the company, the program is prone to “hallucinating” the same way its previous GPT models would – meaning that it can make up facts and commit simple reasoning errors. This is probably due to the AI having a limited knowledge set that cuts off in September 2021 along with its inability to learn from experience.
The company says that the program can also be “overly gullible” when accepting obvious false statements from users and can fail at hard problems the same way humans do. The AI can also be “confidentially wrong” in making its predictions due to its inability to double-check work even when it’s highly likely a mistake was made.
However, the GPT-4 model has still improved significantly in comparison to previous versions in terms of reaching accuracy - scoring 40% higher than GPT-3.5 on internal adversarial factuality evaluations.
The AI poses similar risks as its predecessors – including generating harmful advice, buggy code, cybersecurity risks, or inaccurate information.
OpenAI does not deny the potential risks that this technology may create either, noting that “GPT-4 and successor models have the potential to significantly influence society in both beneficial and harmful ways.”
GPT-4 is Already Being Used
The GPT-4 model has already been put to use in some places, with Microsoft being among the first to have its hands in the pie. The computing giant invested $1 billion in OpenAI to support the building of Artificial General Intelligence (AGI) and revealed on Tuesday that its new Bing search engine has been running on the GPT-4 model for the last 6 weeks already.
Online payment processing company, Stripe, has also leveraged the GPT-4 program to “streamline the user experience and combat fraud.” The AI was called “a game changer” by Stripe’s Applied Machine Learning Product Lead, Eugene Mann, and is used to optimize websites, answer support questions, and detect fraud.
The language learning app, Duolingo, has also taken to using the GPT-4 AI for its newest subscription tier called Duolingo Max which gives users the ability to get thorough explanations and has roleplay capabilities.
Morgan Stanley wealth management has also revealed its usage of GPT-4 to “deliver relevant content and insights into the hands of Financial Advisors in seconds, helping drive efficiency and scales.” Andy Saperstein, the Co-President and Head of the establishment said that the integration of AI technology will allow its financial advisors to harness “knowledge and insights in ways that were once never thought feasible.” The company is currently the only wealth management provider receiving early access to OpenAI’s new products.
Can I Use GPT-4?
The GPT-4 version is currently not available to users of the free ChatGPT service. GPT-4 is only available to OpenAI’s ChatGPT Plus subscribers who pay $20 per month for premium access to the service within a limited usage cap while developers wishing to use the API can sign up for the AI on a waitlist. However, at the rate that companies are integrating the AI newcomer into their services, GPT-4 could be everywhere sooner than you think.
While we're excited about the GPT-4, we should also pay attention to the threat to cybersecurity that the advanced Chatbot would bring. As we discussed in a previous article, the extensive data and natural language capabilities of ChatGPT make it an attractive tool for cyber criminals who are new to the field or simply too lazy to create their malicious code/email and want to carry out a cyber-attack all the same. Read this article to learn about what cyber security risks are associated with ChatGPT and how to protect against them.
Sangfor Technologies offers the best quality cybersecurity and cloud computing technology for your network. For more information, please visit www.sangfor.com.
