OpenAI has developed a technique called “instruction hierarchy” to address the issue of AI models being tricked by injecting misleading instructions. This technique prioritizes the developer’s original prompt and prevents the model from being influenced by unauthorized instructions. The new safety method has been implemented in OpenAI’s GPT-4o Mini model, making it more resistant to prompt injections.
Meta has released Llama 3.1, an open-source AI model that outperforms competitors like OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet. With 405 billion parameters, Llama 3.1 is the largest-ever open-source AI model and was trained with Nvidia’s H100 GPUs. Meta aims to make Llama the most widely used assistant by the end of the year and is working with companies like Microsoft, Amazon, and Google to deploy their own versions.
According to a survey by Capgemini, 90% of businesses are exploring generative AI, with IT, risk management, and logistics being the most common adoption routes. Generative AI adoption has improved productivity, customer engagement, operational efficiency, and sales. However, concerns about bias in models and a lack of clarity around model fairness and training data remain.
A recent study from the University of Pennsylvania found that while using AI tools like ChatGPT in schools can improve performance, it can also inhibit learning. The study showed that students with access to GenAI tutors performed 17% worse than their non-AI-assisted peers when the tools were removed. This research highlights the need for vigilance and technical understanding when using AI in education.
Why do we care?
Another day, another set of models. Meta really is leaving into open source in a big way. Moving beyond the model updates, right tool for the right job remains the theme. Everyone’s exploring… but not every use case produces the improvements hoped for. This is good news – the business of advice and measuring success is exactly where IT service providers want to be.