In the past week, business experts including Allens, McKinsey, Deloitte and KPMG have launched their own proprietary versions of tools based on generative Artificial Intelligence (AI). AI is going mainstream and is moving towards the centre stage of public and government attention. Some experts emphasise the immense potential of AI, while others are deeply troubled about the ramifications the technology may have on humans. It is accepted that AI has the potential to open up employment opportunities, but also to replace many jobs. So how do we ensure that AI tools are both effective and safe?
An ensemble is a group of individuals that is viewed as a whole.
The best approach for creating effective Artificial intelligence (AI) empowered digital products is to approach the design as you would conduct a choir.
When songs are sung by choirs the power of collective sound magnifies, giving the sensation of a single voice but with great depth and volume.
Astrid Jorgensen’s Pub Choir recently brought together an ensemble of 18,812 people across 15 cities with 37 musicians to create one song. Here are the results.
Ensembles of specialised artificial intelligence solutions deliver a more harmonious result than monolithic generic models. When combined with an iterative delivery approach, the result is more useful digital products and an increased likelihood that the digital product will be released to market.
ChatGPT is an ensemble of multiple large language models that generates text content. The risk of a system with such a large span of knowledge is that it can be difficult to ensure the response is factually correct and appropriate for the user.
GoSource’s experience in delivering artificial intelligence solutions has taught us how to deliver AI-powered digital products by breaking the topics or models down into ensemble parts. Put another way, we focus on specialisation over generalisation.
For example, this mirrors the way people engage with the medical system in Australia. If you have an illness or injury you approach your General Practitioner (GP). Then, if more serious you are referred to a specialist, then a surgeon, radiographer or other specialists to diagnose and treat the illness or injury.
GoSource took this approach to create a powerful AI solution that could provide specific answers to detailed questions on medical treatment costs, based on the content of specialist medical selectedon a professional services websites.
The ultimately successful approach was not training a model to understand all service costs. Instead, we trained hundreds of topic models that completely understand one service topic. We then integrated these topic models into an ensemble chatbot system.
Restricting the model to one specific topic not only improves the accuracy of the information generated, this approach puts guard rails around the individual components. This narrows the subject focus and is then less likely to draw on individual opinions from the Internet and provide a better ability to fact-check the answer, before presenting it to the user.
An iterative approach to AI success
Specialised AI systems produce a more harmonious result than monolithic models. Breaking down the delivery of AI empowered digital products into smaller iterations is the key to success.
Taking any large task or goal and breaking it down into smaller, more definable and manageable activities is a proven path to increase the likelihood of success.
The Standish Group’s analysis of over 50,000 projects over 30 years has proven that smaller projects are more successful than larger projects. Projects with budgets above $10 million only have a 10% success rate vs 76% success rate for projects broken down into $1 million chunks.
Better digital products produce more data because more people want to use them. Data engineering transforms the stream of data into higher-quality data. More and better data enable better AI models and better AI models empower more intuitive and accurate digital products. Then the cycle repeats, a positive upward spiral.
To quote the African proverb: If you want to go fast go alone, if you want to go far, go together.
To go far in AI, integrate many specialised models together with an iterative delivery approach. Be the conductor in your organisation.