Microsoft launches Phi-3, its smallest AI model yet

Pivot 5: 5 stories. 5 minutes a day. 5 days a week.

1. Microsoft launches Phi-3, its smallest AI model yet

Microsoft has launched its smallest AI model, Phi-3, which measures 3.8 billion parameters and is trained on a smaller data set compared to large language models like GPT-4. The model is now available on Azure, Hugging Face, and Ollama. Microsoft plans to release Phi-3 Small (7B parameters) and Phi-3 Medium (14B parameters). Eric Boyd, corporate vice president of Microsoft Azure AI Platform, says Phi-3 Mini is as capable as LLMs like GPT-3.5 in a smaller form factor.

Small AI models are often cheaper to run and perform better on personal devices like phones and laptops. Microsoft's competitors have their own small AI models, most of which target simpler tasks like document summarization or coding assistance. Developers trained Phi-3 with a "curriculum" inspired by children learning from bedtime stories, books with simpler words, and sentence structures that talk about larger topics.

Read the full story here 

2. Hugging Face releases a benchmark for testing generative AI on health tasks

LLMs have the potential to revolutionize healthcare by understanding and generating human-like text. However, they face challenges in medical applications like medical question-answering and dialogue systems. The Open Medical-LLM Leaderboard aims to address these challenges by providing a standardized platform for evaluating and comparing the performance of various LLMs on diverse medical tasks and datasets.

The platform includes a variety of tasks and uses accuracy as its primary evaluation metric. The goal is to foster the development of more effective and reliable medical LLMs, contributing to better patient care and outcomes.

Read the full story here

3. Meta says Llama 3 beats most other models, including Gemini

Meta's Llama 3 language model, released to cloud providers like AWS and model libraries like Hugging Face, performs better than most current AI models, according to Meta. The model features two model weights with 8B and 70B parameters, offering text-based responses and showing more diversity in answering prompts, fewer false refusals, and better reasoning.

Llama 3 also understands more instructions and writes better code than before. Benchmark testing is imperfect, and human evaluators also marked Llama 3 higher than other models.

Read the full story here 

4. Microsoft's AI tool can turn photos into realistic videos of people talking and singing

Microsoft Research Asia has unveiled an experimental AI tool called VASA-1, which can create lifelike talking faces from still images and audio files. The tool can generate facial expressions, head motions, and lip movements for speech or song.

Despite potential misuse, the researchers believe the technology can enhance educational equity, improve accessibility for communication challenges, and provide companionship and therapeutic support. VASA-1 was trained on the VoxCeleb2 Dataset.

Read the full story here

5. Can AI image generators be policed to prevent explicit deepfakes of children?

Child abusers are using AI image generators to create explicit deepfakes of their targets, leading to a cycle of sextortion. One of the largest training datasets, Laion, contains millions of images of child sexual abuse material. Despite being illegal in the UK, there is little global agreement on how to police AI image generation.

The creation of more explicit images is easily accessible, as explicit imagery is built into the foundations of AI image generation. Despite governments taking action, the damage has already been done, and systems trained on Laion-5B continue to use the illicit training data.

Read the full story here

Advertise with Pivot 5 to reach influential minds & elevate your brand

Get your brand in front of 50,000+ businesses and professionals who rely on Pivot 5 for daily AI updates. Book future ad spots here.