Google introduces new model PaLM 2

GWith announcements at its I/O developer conference, oogle has once again fueled the exchange of blows that has been going on for months with rival Microsoft and the ChatGPT developer Open AI, which it finances. It will be a “busy year for artificial intelligence,” said Google boss Sundar Pichai and presented the company’s new large language model on stage for the first time: PaLM 2. Such language models are the basis for applications such as ChatGPT. They are trained with vast amounts of data and are therefore able to answer even complex questions at a high level. Pichai promises that the new model should be able to do this faster, better and more efficiently.

In addition, it has new capabilities compared to its predecessor: while the current models were mostly fed primarily with English data, according to Google, PaLM 2 is explicitly trained in more than 100 languages ​​at a high level. In this way, artificial intelligence can understand the nuances of individual languages ​​and formulate poems and puzzles. A developer explains this using an example: the German phrase “I only understand the train station” was taken literally in previous models. PaLM 2 would understand the real meaning.

Because the model was specifically trained with mathematical and other scientific content, the program also has better logic skills. The programming skills are also on a new level with PaLM 2, Pichai promises. The new model is available in four different sizes, the smallest also runs on smartphones.

Google announced 25 new products and applications based on PaLM 2 at the I/O. This also includes Google’s own AI chatbot Bard, which is finally set to pick up speed after a difficult start. Google Bard recently switched from the previous model LaMDA to PaLM 2 on the sly. Google also announced that it would make Bard freely available in more than 180 countries. So far, only select users in the United States and Great Britain have been able to access the program. Bard also now understands Japanese and Korean, and 40 other languages ​​are to follow quickly with the help of PaLM 2. And Bard should also be able to process images soon. Responses from Bard can now also be exported directly to Gmail or Google Docs.

Watermarks for AI images

Other applications such as GMail and Google Maps are also to receive new functions based on artificial intelligence. For example, Pichai demonstrated how a program can automatically write an e-mail to refund airline tickets. A new image editing software should enable users to have artificial intelligence supplement parts of the image that are not visible in the original photo.

In his presentation, Pichai repeatedly emphasized the focus on the security of the technology. Images created by Google AI, for example, would in future all have watermarks and metadata that make it clear that the images were created by artificial intelligence. Pichai was also indirectly reacting to the fact that AI pioneer Geoffrey Hinton left the company last week after more than ten years to warn of the risks of the technology.
The current pace of development is enormous – not only in the large American corporations. More and more freely available language models are also appearing that reveal how they work. This increases the pressure on Google and Microsoft. Pichai already pointed out the upcoming AI model “Gemini”, which was developed multimodally from the ground up and “will open up completely new possibilities”.

Google is making a new attempt at an attack on ChatGPT. The first attempt ended in fiasco. When the program was first presented in February, Bard claimed the James Webb Space Telescope had captured the first images of a planet outside of Earth’s solar system. But it didn’t, as it quickly turned out. Google’s share price promptly fell by 9 percent, and the PR damage could no longer be averted.

The AI ​​writes the code

The pressure on Google is high. Finally, Open AI was not idle either, releasing GPT-4, a new version of the model on which ChatGPT is based, in March. The model is multimodal for the first time, so it can also process images. Microsoft already offers selected industrial partners access to the language models of Open AI via its Azure cloud offering. The German group Siemens is one of the participants. “There is a risk that Microsoft is already securing a supremacy in the use of generative AI in companies,” said Jörg Bienert, CEO of the AI ​​Federal Association, the FAZ

Google is now also counteracting this in the cloud sector. Among other things, Google announced the “Codey” application at the I/O for companies, which can complete code in more than 20 programming languages ​​or even program it based on natural language commands all by itself.