Hub AI in practice
Artificial Intelligence is present in everyday life – from “googling” to facial recognition to vacuum cleaner robots. AI tools are becoming more and more elaborate and support people and companies more effectively in their tasks, such as generating graphics, texting or coding, or interpreting large amounts of data.
What AI tools are there, how do they work, how do they help in our everyday world – and how do they change our lives? These are the questions we address in our Content Hub Artificial Intelligence in Practice.
BodyArmor, a sports drink brand owned by Coca-Cola, is using AI-generated imagery in its regional Super Bowl ad campaign called "Field of Fake." The spot features AI-generated sports highlights with surreal imagery and a robotic voice. BodyArmor aims to differentiate itself from the competition by emphasizing that it contains no artificial sweeteners, flavors, or colors. The AI-generated content is used in this analogy as a negative contrast to the real footage, making it more about cultural commentary than the use of generative AI for ad production. The 30-second spot will air during the Super Bowl in 20 U.S. markets and Canada, as well as a Spanish-language version during Univision's coverage of the event.
Joseph Semrai shows on X how the small, large Mistral 7B language model runs on an Apple Vision Pro. This is a variant of the model with 4-bit quantization, which reduces the model's memory requirements, but also its accuracy. The performance requirements are reduced enough to run on a Vision Pro M2 with a total of 16 GB of memory. A 4-bit version of Mistral 7B Instruct is available here.
Video: Joseph Semrai
The Hong Kong branch of a multinational company lost HK$200 million (US$25.6 million) to deepfake fraudsters. The fraudsters used publicly available video and audio footage to create convincing digital representations of the company's finance director and other employees. They instructed a finance employee to carry out a transaction totaling HK$200 million. The employee made 15 wire transfers to five Hong Kong bank accounts before realizing it was a scam. Hong Kong police are investigating the case and warning of the growing use of deepfake technology for fraud. Senior Inspector Tyler Chan Chi-wing recommended verifying the authenticity of people in video calls by asking them to move their head or ask questions.
Meta introduced "Prompt Engineering with Llama 2", an interactive Jupyter Notebook guide for developers, researchers, and enthusiasts working with large language models (LLMs). The guide covers prompt engineering techniques, best practices, and showcases various prompting methods such as explicit instructions, stylization, formatting, restrictions, zero- and few-shot learning, role prompting, chain-of-thought, self-consistency, retrieval-augmented generation, and program-aided language models. The guide also demonstrates how to limit extraneous tokens in LLM outputs by combining roles, rules, explicit instructions, and examples. The resource aims to help users achieve better results with LLMs by effectively using these techniques. The Jupyter notebook is available from the llama-recipes repository.
Hugging Face has introduced a new Chat Assistant feature that allows users to create custom AI chatbots in just two clicks. Similar to OpenAI's GPTs, the Hugging Face Chat Assistant can be defined by its name, avatar, description, and underlying language model, such as Llama2 or Mixtral. Custom system messages can be used to control the behavior of the chatbot, and different message starters are available. The main advantages of Hugging Face Assistants over GPTs include the ability to choose from different open-source models, free inference provided by Hugging Face, and public sharing without a subscription. The feature is still in beta and has some areas that need improvement to match OpenAI GPTs, such as adding RAG and enabling web search. These features are on the roadmap. Another open-source GPT alternative is OpenGPT.

Adept recently introduced Fuyu-Heavy, a new multimodal AI model for digital agents. Fuyu-Heavy is the third most capable multimodal model after GPT-4V and Gemini Ultra, and excels in multimodal reasoning and UI understanding, the company says. It performs well on traditional multimodal benchmarks and matches or exceeds the performance of models in the same performance class on standard text-based benchmarks. The model performs similarly to Claude 2.0 on chat scores, and slightly better than Gemini Pro on the MMMU benchmark. Fuyu-Heavy will soon power Adept's enterprise product, and lessons learned from its development have already been applied to its successor. Â The following video demonstrates the model's ability to understand a user interface.