GPT-4 is here: As Heise reported exclusively last week, the new generation of the AI system has now appeared. GPT-4 is no longer a pure language model, but multimodal – in addition to text input, it can also handle images. As indicated by the CTO of Microsoft Germany on March 9, 2023 at the digital kickoff event “AI in Focus” in front of business customers, it is actually a multimodal model that can handle different media. According to the current state of knowledge, GPT-4 is able to interpret more complex inputs than previously possible and at the same time to parse text and images.
According to OpenAI, the model should be more creative than the previous GPT 3 series and is probably more geared towards collaboration. In addition to text input, it should also be able to process visual input – although it can apparently only respond in text form. The existing problems that were known from ChatGPT are not resolved: the model continues to tend to confabulate and not always answer truthfully. The ability to generate violent or otherwise harmful content is apparently not banned.
What is now known about GPT-4
Sam Altman, the CEO of OpenAI, pointed out that the now published version of GPT-4 differs only very slightly from GPT-3.5 in terms of conversational ability. GPT-3.5 is familiar to most users as it is the model behind ChatGPT’s chat interface. For over a year, the AI scene had speculated on what architecture GPT-4 will have, and Altman himself had dampened expectations in an interview in January 2023. After the hype, the public will inevitably be disappointed. It’s not yet AGI (general artificial intelligence at human level).
In internal tests, GPT-4 is said to have a significantly lower probability than its predecessor models of generating unwanted content (reduced by 82 percent according to OpenAI) and have a 40 percent higher hit rate for facts than GPT-3.5, i.e. the known one Version behind ChatGPT. It appears to have outperformed ChatGPT on popular benchmark tests, and consistently scored better: in the top ten percent rather than the bottom ten percent on a simulated bar test.
Technical research report and safety trainings
According to the announcement, GPT-4 went through security training for six months and is said to have been readjusted for desired behavior through human feedback in reinforcement learning. A technical research report is available on the OpenAI website. As a result, the architecture of the model is the same as its predecessors, a pre-trained transformer model that predicts the next words by statistical probability and thus generates its outputs. More about the Research work for the model can also be found in a blog entry in the Research section.
GPT-4 is said to outperform existing language models in most NLP tasks and at least be able to compete with “the vast majority of known SOTA systems” (SOTA stands for state-of-the-art, i.e. the most powerful currently available AI systems, including others Offerer).
In the course of the announcement, Microsoft also announced that the new Bing was already using GPT-4. The assumption had already been circulating in the AI scene, since Microsoft had kept a low profile on the model version used. Microsoft recently had to limit its AI-assisted search to a limited number of search queries per IP address and day in order to avoid gaffes. Seen in this way, there are also initial user experiences with the increased creativity of the new model, according to OpenAI, which manifested itself in Microsoft’s Bing, above all, in increased “emotionality” in longer conversations and increased use of emojis.
In the technical report, the OpenAI team also warns that GPT-4 “poses new risks due to the increased capabilities” – the conclusion is silent on exactly what and how OpenAI intends to hedge them. There is still a lot to be done and GPT-4 is a significant step on the way to widely deployable and secure AI systems. More information can be the release notification of OpenAI remove.
You can read more about this here shortly.
(sih)
