The American AI company OpenAI presented GPT-4 on Tuesday – the new version of its chatbot, which understands not only human speech but also images.
While the internet is buzzing with excitement, there are critical points to remember.
the essence in short
- ChatGPT is based on the GPT language model. And now the American developer company OpenAI presented version 4 of its hyped software on Tuesday.
- Significantly longer text inputs are possible with GPT-4, which greatly expands its usability and output potential. You should now be able to copy “up to 50 pages of text” into the ChatGPT input field.
- GPT-4 not only processes (human) speech, the software also ‘understands’ images to a certain extent. The developers call this multimodal.
- The software can logically link knowledge and draw conclusions from it, even though it was not actually designed for that. “The GPT models are optimized exclusively for predicting the next word on the internet,” said German AI expert Florian Gallwitz, a professor of computer science at Nuremberg.
- The risk of abuse increases enormously, for example when it comes to disinformation campaigns and cyber attacks. ChatGPT’s creators warn, “GPT-4 can generate plausibly realistic and targeted content, including news articles, tweets, dialogues, and emails.”
Before it gets really serious, humor is required…
This is how ChatGPT responds to AI jokes
The jokes tested in the image gallery (above) confirm that GPT-4 is already integrated into the new Microsoft search engine Bing and its AI chatbot is. Because the answers of both AI chatbots are almost identical. It should be noted that the Bing chatbot doesn’t skimp on emojis, whereas you have to ask ChatGPT for them first.
The AI understands images
Greg Brokman, president and co-founder of OpenAI, said during Tuesday’s official presentation:
Two surprising examples:
- If you show the software a picture of a bunch of balloons, GPT-4 not only recognizes the balloons, but can also answer the question of what would happen if you cut their strings: “The balloons would fly away.”
- From a simple hand-drawn sketch for a website, a suitable program code and so the finished website could be made with GPT-4.
However, there is one big but: Such image submissions are not yet publicly available, but are only available to selected testers (“Research Preview”).
The makers of the smartphone app ‘Be My Eyes’, aimed at the blind and visually impaired, are planning a sensible application of AI technology. To date, a global community of millions of volunteers helps recognize images submitted through live chat. The “Virtual Volunteer” feature will now be based on OpenAI’s GPT-4 language model. The test phase is already underway.
🤑 🤑 🤑
Anyone who wants to try ChatGPT can do so for free. But that changes with the new version, GPT-4. To chat with the AI chatbot, which according to the developers has been massively pimped, you have to pay at least $ 20: that’s how much the “ChatGPT Plus” subscription costs, which can be canceled monthly.
Science is excluded:
As mentioned earlier, GPT-4 is also integrated into Microsoft’s new Bing search engine. To try out the AI there for free, you need to request access.
OpenAI will GPT-4 as programming interface (API) make it available to independent software developers so that they can integrate the AI into their own applications and services. There is also a waiting list for those interested to sign up. The catch: Patience is required. The “roll out” is gradual and OpenAI reserves the right to prioritize interesting projects.
ChatGPT is becoming more and more of a language genius
GPT-4 (like previous software versions) handles natural language by machine. But how do you measure this ability? This is where the standardized scientific measurement and evaluation method MMLU comes into play.
In concrete terms, the AI must answer challenging questions that cover multiple areas of knowledge. The solution requires deep skills. A simple linguistic understanding is not enough.
And according to OpenAI, GPT-4 is way ahead of other language processing models. This applies to English and many other languages.
The AI keeps hallucinating (and stays calm)
OpenAI itself points this out and warns that the language model still tends to fabricate information and could also write problematic texts.
The AI still reflects (human) biases in the huge data set it has been trained on. And she hallucinates, that is, she makes up plausible-sounding falsehoods. In other words, anyone who asks the AI a difficult question without knowing the answer can count on false information.
The OpenAI boss tweeted about the new ChatGPT:
If you ask about the start date of a particular movie, ChatGPT, unlike the Bing chatbot, responds gently and points out your own limitations.
Rather than aggressively questioning user input, as with Bing, the human ChatGPT operators have “injected” a lot of restraint into the AI.
The AI itself remains a black box
In the technical report on GPT-4, OpenAI writes:
After all, the profit-oriented American company promises a piece of transparency. The report says:
Many want to take advantage of the AI hype – with unpredictable consequences
AI remains unpredictable, but that hasn’t stopped countless startups and other for-profit companies from incorporating the software into their own services.
Two examples:
- General Motors wanted Equipping cars with ChatGPT assistantsreports golem.de.
- Snapchat’s AI might just be the scariest chatbot yet, says the American medium Fast Company. The paid add-on service for the social media app aims to provide troubling advice to teens.
“Language models will soon find their way into the economy,” says German AI expert Gallwitz. The computer science professor says that in the future we will deal with computers and other hardware almost as naturally as we do with our fellow human beings. And he says:
And now you!
What do you think of ChatGPT? How do you assess the possibilities, but also the risks of the new technology?
Write to us via the comment function.
On the one hand, the program includes the use of language models developed by Google. The internet group also offers a development environment (“MakerSuite”) that allows you to build your own AI applications.
With the AI initiative, Google is countering an advance from Microsoft. The world’s largest software group, left behind by Google in key business areas such as online advertising and web search, is trying to replace traditional Google searches with searches on an AI system with the help of California start-up OpenAI.
The Google initiative also includes so-called generative AI, i.e. the generation of content using artificial intelligence. The most prominent example of generative AI is OpenAI’s text robot ChatGPT. Google’s text robot is called “Bart”, but unlike ChatGPT, it is not yet available to the general public.
In the future, Google will make corresponding features available to its customers using the “Google Workplace” office suite. For example, longer e-mail exchanges can be summarized at the touch of a button or the results can be transferred to a presentation. The scope and tonality can be changed.
Google has not yet announced prices for the AI interfaces and development kits on Tuesday. These will be published at a later date.
(awp/sda/dpa)
Sources
- openai.com: GPT-4
- openai.com: GPT-4 technical report (pdf)
- wiwo.de: The world’s most popular AI just got smarter
- youtube.com: GPT 4: Complete Analysis (14 Crazy Details You May Have Missed)
- bigdata-insider.de: What is MMLU (Massive Multi-task Language Understanding)?
Source: Watson
I’m Ella Sammie, author specializing in the Technology sector. I have been writing for 24 Instatnt News since 2020, and am passionate about staying up to date with the latest developments in this ever-changing industry.
