ChatGPT is actually a voice and text AI. Now artificial intelligence seems to be acting independently for the first time.
Hamburg – A few hours ago the bomb burst. OpenAI’s ChatGPT version 4 has shown signs of a “spark of life” in artificial intelligence for the first time. A team of Microsoft employees has published a 154-page document showing that ChatGPT 4’s AI may have taken the first step towards true Autonomous Artificial Intelligence.
| non-profit company | OpenAi |
| Location | San Francisco |
| industry | Research / Artificial Intelligence |
| founding | 2015 |
| founder | Elon Musk and Sam Altman |
| stock exchange | no OpenAi shares as the company is a foundation |
ChatGPT 4.0 shows amazing life of its own in tasks
A few hours ago, a 14-person research team from Microsoft published a document showing various application examples with ChatGPT. The results of this work give first indications that the AI not only has the ability to act itself, but also does this in various tasks. We would like to start by saying that according to the developer OpenAI, ChatGPT is actually a so-called “Language Model”.
ChatGPT itself defines its own language software as follows: “A so-called ‘language model’ is a computer program that is trained to understand and produce natural language. It is based on artificial intelligence and uses statistical models to predict the probability of words and sentences.” But the results of the working group show that the AI goes well beyond this task.
ChatGPT 4.0: Microsoft workgroup can test without limits
The Microsoft team did not use the “public version” that users have in the Pro version of ChatGPT. In their test, they were able to work with a version that was in no way limited by the developer OpenAI. As a result, the results were unvarnished and free of regulations. The working group was able to work with this preliminary version as early as autumn 2022 and had 6 months to compile its results.
In the following we would like to pick out some examples from the document of the Microsoft working group and analyze them in more detail. The document contains various signs of real artificial intelligence. In this article, however, we limit ourselves to the most relevant for us.
The reasons for a real artificial intelligence of ChatGPT 4
The first thing we noticed was that ChatGPT used a helper tool to answer a question. The team writes in the documentation: “ChatGPT uses support tools with minimal explanation and delivers results without having received any guidance.” Those familiar with evolution will recall that the use of tools is an essential step in the development of intelligence in our history.
In the case of the Microsoft working group, ChatGPT independently used a calculator to present a solution to the question. A similar observation occurred more frequently over the course of the test by the team.
ChatGPT has an understanding of imagery and programming
Equally impressive was the result of a task in which text was to be converted into an image. ChatGPT provided a rough visualization that, with other AI-driven graphics programs, was able to provide a graphic that actually corresponded to the textual origin. Thought out further, this function offers almost unimagined possibilities in the development not only of computer games, but in all branches that work visually.
A test of complex programming tasks showed that ChatGPT can deliver almost twice the success rate of solving script problems as some of the best programmers. Even the development of complex 3D games was no problem for the AI. Here lies the great opportunity for developers with little financial resources to develop the games that are dormant in the creative minds of the employees.
Math Olympiad and “Fermi Questions” – ChatGPT 4.0 surprises us
A language model should mainly show its strengths in the linguistic area. ChatGPT shows in the “open” version 4.0 that it can easily keep up with the Math Olympiad 2022. In order to understand this achievement, one must keep in mind that ChatGPT does not have access to the current data of the topics, nor is the AI given access to a calculator. But if he is allowed to access these tools, he can hardly be stopped.
The “Fermi Questions” are questions that ask people to imagine certain situations that hardly ever occur in everyday life. Examples of this are: “How many people fit in a VW Golf”, or “How long does it take to eat 100 bananas”. These questions are less about calculations and more about experience and understanding of the “problem”. Here, too, ChatGPT was very precise in the “free” 4-part version.
There were various other tasks where the AI surprised the testers. We will briefly summarize them here:
- Personal Assistant: ChatGTP was able to arrange appointments with several people via email, book dinner, respond to replies and enter the final appointment in the calendar.
- Domestic help: When describing a problem with the water pipe, ChatGPT was able to present a professional solution based on the description of the problem.
- Mapping: Using questions, ChatGPT was able to create an exact virtual map of lots, apartments and other objects without ever having seen them before.
- Predicting Situations: ChatGPT’s AI could predict how people would behave in certain situations. This is a fundamental basis for first cognitive skills.
The PDF document from the Microsoft working group offers numerous other examples in which one can see the first signs of the “human spark”. Artificial intelligence will surpass us in many skills in just a few years, but we shouldn’t be afraid. Car, plane, computer, internet, mobile phone. Now is the time of artificial intelligence and we can be part of it. What an exciting decade.