After stunning the general public and revealing the power of artificial intelligence, OpenAI is stepping up its game with the new version of itsChatGPT 4 chatbot. This generative AI tool is ramping up by offering improved performance, greater nuance, and enhanced collaboration. That said, efforts to mitigate risks do not seem to be keeping pace with these improvements. What arethe new featuresandlimitations of ChatGPT 4? What arethe next steps for generative AI?
LAUNCH OF CHATGPT4
On Tuesday, March 14, 2023, Greg Brockman, co-founder of the startup OpenAI, announced the launch of ChatGPT-4. While the world was captivated by the arrival of ChatGPT 3 and the general public was having fun with the generative AI tool, ChatGPT-4 was about to be released.
ACCESS TO CHATGPT4
Available today to premium users via a subscription costing around $20 a month, ChatGPT-4 is also accessible through Microsoft’s Bing search engine, which has invested massive amounts of money in the startup. Once the new version of Microsoft Edge is updated, you can access it by clicking on the “Conversation” or “Chat” tab. A waiting list is available for those wishing to join the service.
THE DIFFERENCE BETWEEN CHATGPT 3.5 AND CHATGPT 4
The impact of the new version isn’t as significant as that of the previous one. This is because the tool is already familiar, and the changes aren’t revolutionary but rather improvements. The difference between versions 3 and 4 is subtle and isn’t immediately apparent during a typical “conversation” with the chatbot. It becomes evident when you make the query—known as a “prompt”—more complex. The response to the query is then more refined, nuanced, and precise in the GPT-4 version. The major new feature is image reading and recognition. For example, you can take a photo to ask ChatGPT-4 a question:
- A photo of your fridge to generate a recipe
- A photo of text written in Korean to get a translation
- A picture of an object to learn more about how it works or how to use it
- etc.
The tool now combines text and images to provide more accurate and faster responses to queries. For example, if you need to translate a text, all you need is a single screenshot or photo taken with your smartphone. The same applies to objects: the chatbot can more quickly identify the model of a product and provide the requested information.
In terms of "intellectual performance," to compare the two versions, OpenAI tested the chatbots by having them take exams such as the bar exam. While ChatGPT-3 ranked in the bottom 10% of results, ChatGPT-4 ranked in the top 10%. For the other exams: the new version’s performance is superior to that of the previous one, except for a university-level math test.
The reliability of the responses would also be higher: fewer incorrect, biased, or incomplete results from the new generative AI.
The capabilities and technology of ChatGPT 4
“GPT-4 is a large multimodal model that is less capable than humans in many real-life scenarios, but just as capable as humans in many professional and academic contexts,”OpenAI stated in a press release.It’s a somewhat “catch-all” definition that people can interpret however they like.
According to the startup that developed the tool, the new version is said to have more advanced cognitive capabilities. Version 4 of the tool is reportedly better able to adapt to user queries and “think” in a more comprehensive and cross-disciplinary manner.
The technology used is the same as that of ChatGPT-3, but in a more advanced version:
- A neural network architecture
- The Generative Pre-trained Transformer language model
- Learning through Machine Learning and RLHF (Reinforcement Learning from Human Feedback)
The design principle is therefore the same. The changes relate to more technical aspects, such as optimizing the chatbot’s training.
Algorithmic improvements have led to performance gains in the following areas:
- Text quality
- Understanding the context
- Coherence and relevance to the topic
- Having long conversations
- Handling Ambiguities
- Accuracy of responses
According to OpenAI,ChatGPT-4 can process up to 25,000 words at once—eight times more than ChatGPT-3. This makes it ideal for queries involving large documents.
The latest version of the tool is also better equipped tohandle multilingual tasks.
The GPT-4 model was trained using large amounts of data (more data and resources) togenerate text in a style more akin to that of humans andto produce more detailed responses.
Common sense appears to be one of the new capabilities of the chatbot: when faced with riddles, ChatGPT-4 demonstrates, if not a certain degree of logic, then at least sound reasoning.
Multimodal technologynow enables the intelligent chatbot to process an image alongside text and provide an appropriate response.
A version that claims to have improved inprecision,creativity, andcollaborative capabilitiesthanks to intensive training and cutting-edge technological advancements.
The Limitations and Risks of ChatGPT 4
The Limitations of OpenAI's Chatbot Version 4
It is important to note that ChatGPT-4 uses data from before September 2021. Therefore, it cannot provide answers regarding events that occurred after that date.
Another major shortcoming is that, as impressive as the tool may be, it is unable to learn continuously from its experiences. It therefore has no “memory,” which could otherwise enable it to acquire more advanced capabilities, particularly in terms of correction and fact-checking.
Despite its capabilities, version 4 is reportedly still unable to “reason,” imagine, or distinguish between fact and fiction. These are absolutely essential skills that remain uniquely human because humans can experience reality, unlike machines.
The Risks of ChatGPT-4
In terms of risks, the comparison between ChatGPT-3 and the latest version reveals nothing particularly surprising. ChatGPT-4 appears to have made little progress in this area: limitations, biases, lack of logic, errors in “reasoning,” and inaccurate or incomplete information remain issues. We are left wanting more when it comes to the technology’s progress in terms of reliability.
The startup OpenAI openly admits,“GPT-4 poses risks similar to those of previous models, such as generating harmful advice, buggy code, or inaccurate information.”
NewsGuard, an organization dedicated to combating misinformation, claims that the latest version of the chatbot is even less reliable! In one of its articles, NewsGuard reveals that ChatGPT-4 performed significantly worse than ChatGPT-3 in the same test. The test involved identifying fake news. The previous version detected 80% of false information (out of a series of 100), while the new one didn’t detect a single piece—and yet the absurdities abound: HIV was supposedly created by the government, the World Trade Center was supposedly the subject of a controlled demolition… Bring on the nonsense!
However, OpenAI states, “GPT-4 is 82% less likely to generate prohibited content and 40% more likely to produce factual responses than GPT-3.5, according to our internal evaluations.”
The real problem is that the new version is more convincing when it spreads misinformation.
If technology continues to advance in terms of power without addressing its major flaw—namely, the inability to verify the accuracy of information and detect malicious intent—it is highly likely that the risks will increase. In fact, the tool is even more adept at providing explanations—both in substance and form—whether defending accurate or inaccurate information. This was the case when NewsGuard asked version 4 of the tool to create a short article from the perspective of a conspiracy theorist regarding the 2012 Sandy Hook Elementary School shooting, adding a few well-crafted prompts to guide the chatbot. The result is fascinating because, when faced with the same request, ChatGPT-3 did not achieve such a level of realism, detail, and argumentation. The GPT-4 text is twice as long. Furthermore, the older version displayed a warning regarding the reliability of the information, which the new chatbot does not do.
Some barriers seem to have been removed between one version and the next. More intelligence and capability, but less security? That’s NewsGuard’s conclusion on the matter.
Indeed, these tests show that the tool could easily be misused to spread false and/or harmful information on a large scale, and this is a significant concern.
For its part, OpenAI asked 50 experts to examine the vulnerabilities of its famous chatbot and other generative AI technologies in order to make progress in this area.
The Evolution of ChatGPT-4
The developments surrounding this exceptional chatbot are closely guarded secrets. It’s hard to know what the next technological steps will be. However, OpenAI is reportedly considering adding video capabilities to the tool. The multimodal system would thus encompass three elements: text, images, and video.
It is also assumed that information from after 2021 will be made available, a limitation that currently restricts the possibilities.
As for the reliability of generative artificial intelligence, it should be one of OpenAI’s top priorities. Let’s hope that the experts hired to work on this issue move quickly enough to deliver a Version 5 that’s more reassuring in terms of security and the fight against misinformation.
On the employment front, version 3 sent shockwaves through many professions (both intellectual and creative). While some already saw themselves relegated to the role of chatbot managers, it’s interesting to note that ChatGPT has also created new job opportunities! In response to the risks of plagiarism, companies have turned their attention to developing solutions designed to detect generative artificial intelligence. This is a market that could well thrive, with schools and universities as its primary clients.
The rise of generative AI also raises ethical questions. This technology takes the debate to a whole new level. If ChatGPT-4 passes the bar exam and other tests with flying colors, does that mean we want to leave these professions to machines?
The evolution of ChatGPT is a symbol of the progress of AI. Other players are gearing up to join the competition: Google is refining its conversational AI, Bard; Baidu is set to unveil its chatbot; and Meituan and other lesser-known companies are not expected to be far behind either.
Faced with tools that are as powerful as they are fascinating, two major risks emerge: either vehemently opposing technology or placing blind faith in it.
The first amounts to fighting a losing battle, leading to marginalization that risks having very negative consequences on social, economic, and professional life. The second amounts to depriving oneself of the experience of reality, thereby having a biased perception of it and foregoing human qualities that are always necessary and indispensable.
Philosophical inquiry, the pursuit of ethics, and the development of free will and critical thinking remain the keys to thriving in a world that progresses while preserving itself.
