The next big leap in AI could come with GPT-5. OpenAI reportedly wants to release the new version this year and also equip the chatbot ChatGPT with GPT-5. The AI model should of course be significantly better than its predecessors, and above all, so-called AI agents should move in with it. These can complete tasks autonomously.
Advertisement
It's the classic example when it comes to new applications: “Book me a trip to 'insert any destination here'.” An AI agent could go out and do it for the client, including flights, accommodation and on-site activities. How much human intervention will still be necessary in the end remains to be seen. Even if it would technically work to present your credit card details to the agent and let him decide, the monitoring should at least be carried out by humans.
Book trips thanks to autonomous agents
Nevertheless, if a chatbot could find the cheapest flight, it would not only make travel planning easier for the user, but also for the providers of sites where you can book flights. They will then only be clicked by a bot that has very little interest in advertising. The Digital Markets Act (DMA) has just given flight search engines an advantage by prohibiting Google from displaying its flight comparison page directly between search results. It is preferable if users can search for flights there with one less click instead of having to go to other providers' sites first.
In addition to GPT-5, Human AI Pin or the Rabbit r1, for example, can also be used to have a trip booked by the AI. However, they work a little differently. Rabbit r1, for example, must first be shown in a rabbit hole how to navigate through a page where he can make the booking. Autonomous agents can also take on much simpler tasks, such as adding appointments to a calendar or searching for information.
Business Insider is currently reporting that GPT-5 is scheduled to appear in the summer, citing companies that have already been shown the AI model. The article quotes a CEO calling GPT-5 “impressive” and “significantly better” than GPT-4.5. It is also said to be better at reasoning and more reliable. Recently, GPT-4 had repeated problems with what was known as “laziness” – tasks were not completed.
Training, red teaming, publishing
However, the model is still in the training phase. As always, it is and remains unclear what material the AI model is trained with. In the end, the crucial question may be whether AI models were allowed to be trained on copyrighted materials or not. Most recently, OpenAI CTO Mira Murati explained in an interview with Sora that she did not know exactly what the video generator's training material included: everything that could be found freely available on the Internet, possibly also Facebook and Instagram posts.
After the training phase of an AI model has been completed, it is transferred to so-called red teaming, where it is checked which security risks the version poses and, if necessary, additional guardrails are put in place. Sora is currently in this process. Both the video generator and GPT-5 are scheduled to appear this year. Sam Altman also recently said in a podcast with Lex Friedman from MIT that there would be a “smarter” version of GPT – he didn't let himself be pinned down on the 5 in the name.
(emw)