Google wants to improve the accuracy of its AI models. To avoid “hallucinations,” the company is working with partners such as Moody’s, Thomson Reuters, and ZoomInfo who will feed the AI systems with up-to-date information. A new “confidence score” is also supposed to indicate how confident the AI is in its answer being correct.
With Gemma 2, Google is also releasing a new series of lightweight AI models. In addition to a 27-billion-parameter model, there is now a 9-billion-parameter variant that is particularly suited for applications on devices such as smartphones or IoT devices. The models are designed to give developers more flexibility and can be freely used and customized. Google also plans to release an even smaller model with 2.6 billion parameters in the future.
Google’s new Imagen 3 text-to-image model is now available on the company’s Vertex AI platform. The improved image AI is said to offer faster image generation, more accurate prompt understanding, photorealistic humans, and better text representation in images.
Google is also adding the Mistral Small, Mistral Large, and Mistral Codestral models to its Vertex AI platform. This follows the recent addition of Anthropic’s Claude 3.5 Sonnet, and demonstrates Google’s plan to providing businesses with a broad range of AI tools and models.
Meta, on the other hand, is testing user-generated AI chatbots on Instagram. The chatbots, created through Meta’s AI Studio, will initially appear in direct messages in the U.S. and will be clearly labeled as AI. The goal is to allow users to interact with AI avatars of their favorite influencers and topics. In the long term, businesses will also be able to create their own AI chatbots to improve communication with customers.
Meta also introduced “3D Gen”, an AI system designed to generate high-quality 3D models from text descriptions in seconds. The combination of “3D AssetGen” for 3D model creation and “3D TextureGen” for realistic texture generation is supposed to enable the production of complex objects at unprecedented speed.
Meta has also released pre-trained models based on multi-token prediction. This technique, which Meta first introduced in a research paper in April, allows models to predict multiple words at once, rather than just the next word in a sequence. This allows the models to be trained more quickly, while achieving better performance.
And with MobileLLM, Meta has developed a new language model optimized for use on smartphones and other devices with limited resources. MobileLLM outperforms comparable models and shows that even smaller models can be competitive in certain applications.