Last month, Nesta held a generative AI exploration day. Colleagues from across Nesta and BIT as well as Best Beginnings, the RSA and Kairoi came together for a few hours to learn more about generative AI and build prototypes that could help increase Nesta’s impact across our three innovation missions.
The main goal of this day was to connect people with technical expertise and domain experience in our mission areas of early years, sustainability and food environments. So far, our generative AI project has focused primarily on early childhood education, so it was exciting to bring together colleagues from different missions and practices. We hoped to identify new problem areas that could be addressed with generative AI and spark ideas for prototypes.
Overall, we had almost 30 people taking part in-person or remotely. At the start of the day, participants formed small teams – everyone had a chance to introduce themselves, pitch their ideas and learn what other participants were interested in. We also provided some practical and legal guidance related to various generative AI tools such as ChatGPT and OpenAI’s application programming interface (API). Participants could also seek guidance from Nesta’s two AI residents.
Our teams explored a variety of ideas, as outlined below.
We created a prototype chatbot that can answer questions about labour or childbirth, using verified content from a popular parenting app. The team, which included one of our AI residents and the developers of the parenting app, used OpenAI API and leveraged a method called retrieval-augmented generation to ground OpenAI’s large language model (LLM) in its high-quality content. The team’s impression was that the responses from the chatbot had a similar tone to the app; at the same time they will need to test rigorously the accuracy of the chatbot outputs. It was exciting to see this team’s prototype, as Nesta has been working on similar use cases as part of our generative AI project.
Another team experimented with asking ChatGPT to come up with ideas for healthy eating startups. The team also asked AI to suggest successful startups and describe their business models. This use case is relevant to Nesta’s Mission Studio, which creates, spins out and scales new tech startups that tackle some of the UK’s most pressing social challenges. ChatGPT did a good job of applying those business models to the problems we gave it, as well as generating some convincing personas.
One of the teams tried out platforms that help identify and summarise research insights, such as Elicit and Scite. Both platforms seemed to do a reasonable job of rounding up relevant academic research, as well as providing information about research funders. Literature reviews and qualitative analysis of research interviews have come up as some of the most time-consuming areas of Nesta’s work where LLMs might make the biggest impact. The team also tried using ChatGPT to summarise research project updates into tweets, and to identify potential research gaps, both of which worked well. Finally, the team also experimented with tools such as Litmaps and ResearchRabbit, which don’t necessarily use generative AI but visualise the connections between different academic publications, making it possible to appreciate the impact of the research.
We learned about how to use ChatGPT’s Advanced Data Analysis function (formerly known as the Code Interpreter). It’s possible to upload a spreadsheet of data, ask for hypotheses, create graphs and findings, and even write a simple scientific paper on the results. ChatGPT, however, struggled to integrate datasets in different formats and it can’t yet read image-based PDFs (but it is likely only a matter of time until it can).
We tested OpenAI API to match data on food and drink purchases with their nutritional information. The challenge here was that products have slightly different names in each database. The few-shot learning capabilities of ChatGPT (ie, learning from a very small number of examples) allowed the team to find matches between both datasets. It wasn’t perfect, but worked most of the time. This could be very helpful for projects that further Nesta’s healthy life mission, which aims to improve our food environments through, for example, identifying the most impactful products for food reformulation.
One of the teams tried its hand at a wide range of generative AI tools. From large language models such as ChatGPT (state of the art), Bing Chat (which has access to the internet and can cite sources) and Claude (which can interpret PDF documents), to platforms for generating images such as Adobe Firefly and Midjourney. The team also tried a ChatGPT extension for Google Sheets, which makes it simpler to process large amounts of tabular text data. As their example problem, the team leveraged all these tools to design a social marketing campaign for a nursery and tailor the messaging to different types of families. We used Claude to analyse a PDF report about a local authority’s demographic situation, Bing Chat to create a marketing plan, and other generative AI tools to draft initial marketing copy and illustrations for different personas. While this particular example could be relevant for the fairer start mission goals, these workflows apply to a host of other tasks across the organisation.
With the hope of supporting Nesta’s healthy life mission, we also tried asking ChatGPT to come up with healthy recipes that contain hidden vegetables. Curiously, it often just replaced the name of the vegetable with the word ‘hidden’ in the ingredients list. It also didn’t quite understand how to make substitutions like courgette for butter. This limitation confirmed the recent observation that this technology might not yet be ready for generating new recipes. Nonetheless the team came up with a theoretical design for a better system that could do just that.
This team asked ChatGPT to write stories about heat pumps in the style of Shakespeare or JG Ballard, to see if it might encourage people to buy one. More seriously, however, Nesta’s sustainable future mission is focused on making it easier for people to use clean, green sources of energy to heat and power their homes. To achieve its goals, it is important to convey the information about environmentally-friendly solutions such as heat pumps in a clear and engaging manner. The ability of LLMs to generate good-quality text might help with exploring a larger number of initial ideas, which can then be narrowed down and further developed by professional writers. In the team’s exploration, ChatGPT also came up with some clever ideas for television programmes based on a range of presenters.
Our final team explored ChatGPT to draft guidance for dealing with crises, describing aspects such as who should take decisions and suggesting how to use available resources.
Overall, participants of Nesta’s generative AI exploration day found the experience beneficial and even transformative. They particularly appreciated having some protected time for trying out new tools, learning through doing and testing methods that deviate from their usual problem-solving approaches. Participants also enjoyed the opportunity to collaborate with new colleagues, and the presence of AI resident experts to provide guidance and advice was invaluable.
Finally, it’s important to note that in this event the emphasis was slightly more towards appreciating the potential impact of generative AI – but the participants were also mindful about the limitations and shortcomings of this technology.
If you have more ideas for how we could be using generative AI in our work, or you’d like to share examples of how you’re using it, get in touch!