Monthly Notes: AI friend or foe, satellite imagery, and data collection insights from the source.

This month we interviewed IDinsight's Data Science Director Sid Ravinutala . In our Q+A below he shares which social sector AI innovations he's excited about, which jobs will stand the test of time when AI changes the workforce, and how we can mitigate risks of AI upending life as we know it.

Q. What excites you about recent AI innovations and their application in the social sector?

Sid Ravinutala, Director, Data Science: Working on data science projects in the social sector is to work under constraints – we are often up against tight budgets and insufficient data to train models. So many leaders are beginning to appreciate how machine learning and data science can add value to their work. Yet many are still unable to reap these advantages because of small operating budgets and slim margins. Two innovations that I’m most excited about make it substantially cheaper for a social sector organization to take advantage of machine learning.

First, is the MOSAIKS paper by Rolf et al. By pre-processing satellite imagery, it substantially reduces the effort and time needed to build models.

The second is the availability of large language models and libraries to work with them. We no longer need to train or maintain large models for working with natural language or text. Now, we can simply use OpenAI’s services and open-source packages like langchain and llama-index to quickly build AI-driven chatbots and applications. Read more about our chatbot work here.  

These advances make it significantly less expensive to build AI-driven applications that bolster the impact of our partners.

Q. What application(s) have been the most useful for clients with whom we work?

Sid: We have been working with Educate Girls for almost 5 years now. Each year we help inform Educate Girls ' expansion strategy by using a machine learning model to predict the number of out-of-school girls in each village. 

We have considered using satellite imagery to find out of school girls for a number of years, but it was not cost-effective until this year. Using the MOSAIKS methods, we can now generate satellite imagery features in days and build and test a model in weeks instead of months. 

Similarly, we worked with Reach Digital  to answer health questions from mothers in South Africa from a database of FAQs. We trained our own models and experimented with fine tuning large models like BERT. Now, with large language model services being highly performant and cheap, we can improve the performance of the question-answer matching service while reducing its operation cost. A win-win like that was not possible a few years ago.

Q. What is your take on whether AI could destroy us?

Sid: Coincidentally, this has been the topic of our reading group for the past few weeks. My personal take is that though there is a non-zero chance of this happening, the threat is not as proximate as, say, climate change. 

The existential threat from recent chatGPT-type models might be a little overblown. But there is a danger that we release an objective-maximizing agent into the world without sufficient constraints and it ends up destroying humanity as a side effect. Such models are not sophisticated enough to warrant immediate concern. Having said that, things may change quickly and we should think about what regulations we want in place on releasing such agents in the wild. 

I worry a lot more about the existing over-use of machine learning for making high-stakes decisions like who gets an interview or who gets off on bail. Before it destroys mankind, uncontested and biased black-box models could destroy many lives.

If we are to use ML for high-stakes decisions, at a minimum we must have a human-in-the-loop. At IDinsight, we have invested a lot in an ethics process with both internal and external reviewers to scrutinize our models for bias and potentially negative consequences.

Q. If AI doesn’t destroy us but does transform industry, are any skills irreplaceable?

Sid: We used to use human calculators before electronic calculators were invented. Today, we are focused on what we want to calculate while trusting a machine to do the number crunching correctly. 

My take is that the AI transformation will be similar. Higher level tasks that involve critical thinking and creativity will continue to be valued. Some of the execution of tasks may get automated or assisted by AI. 

I should add “for now” at the end of each of these sentences. It is a fast changing world and difficult to predict what it will look like once the dust settles.

IDinsight's Monthly Notes is the organization's newsletter with the latest insights, reflections, and news from IDinsight. Read this month's content and stay up-to-date with stories, news, events, and more from us by subscribing today.