The future of report and recommendations — a glimpse

Large language models, like openAI’s GPT3, are very powerful AI models.

You might have encountered the above statement sometime during last year. I did too. And moved on. Till a course in nautral language processing brought me back to things that it deals with.

Can such models, trained on huge amounts of online data, produce good recommendations for a totally out of context scenario? Say for bureaucracy in India? Like for a small, mostly unknown service called the Indian Trade Service, which has hardly 100 officers working in it? I expected that the training data might not have any mention of this whole thing called ‘Indian Trade Service’. A significant part of information about this service is contained in one small wikipedia page (mostly, and decreasingly, maintained by me). Thus, after trying whatever I was trying, I tried out some bureaucracy recommendation stuff with GPT3.

And it was a great surprise. GPT3 not only managed to pick up stuff about the service, but was also able to display great amount of understanding (in whatever sense ‘understanding’ occurs in AI). Somewhere in the huge n-dimensional space, it could see that Trade is very close to foreign affairs, a fact that some humans miss. It also had wider suggestions for bureaucracy itself.

I fed the following questions to GPT3’s text-davinci-002 — the largest and most knowledgeable model — as a prompt completion task. Sample the answers yourself : (zero editing of answers from my end. I have placed the best answer among three trials — every trial leads to random generation of answers which may or may not resemble one another. I have pasted the alternative answers too, at times. Every letter, punctuation, and content is that of GPT3)

— — — — — — — — — — — — — — — — — —

Q: How should the government of India manage the Indian Trade Service. Some suggestions?

GPT3:

1. The government of India should consider merging the Indian Trade Service with the Indian Foreign Service.

2. The government of India should create a dedicated Indian Trade Service cadre within the Indian Foreign Service.

3. The government of India should establish a separate Indian Trade Service Academy to train Indian Trade Service officers.

4. The government of India should increase the intake of Indian Trade Service officers from the open category.

Q: Why should Indian Trade Service be merged with Indian Foreign Service?

GPT3: There are several reasons why the Indian Trade Service should be merged with the Indian Foreign Service. First, both services are responsible for promoting and protecting the interests of India in the global economy. Second, both services have similar mandates and functions. Third, the Indian Trade Service is a small service with a limited number of officers, while the Indian Foreign Service is a large and well-established service. Fourth, the Indian Foreign Service has a better reputation and is better resourced than the Indian Trade Service. Finally, the merger would create a more efficient and effective Indian diplomatic service.

Another trial produced:

Q: What ails Indian bureaucracy? Some suggestions to improve.

GPT3: There is no one-size-fits-all answer to this question, as the bureaucracy in India faces a range of challenges that vary depending on the specific context and location. However, some suggestions for improving the functioning of the Indian bureaucracy include:

• increasing transparency and accountability
• strengthening anti-corruption measures
• promoting meritocracy and introducing performance-based evaluations
• rationalising the number of layers and levels within the bureaucracy\n-simplifying procedures and paperwork
• providing adequate training and capacity building for government employees
• Encouraging more public-private partnerships.

Another trial produced:

Q: What are some case studies on best practices on trade disputes. A list:

GPT3:

1. The US-Canada softwood lumber dispute

2. The US-EU banana dispute.

3. The US-Mexico trucking dispute.

4. The US-China trade dispute.

5. The US-Japan trade dispute.

6. The US-Korea trade dispute

7. The US-Vietnam shoe dispute

8. The US-India cotton dispute

— — — — — — — — — — — — — — — — — — — — —

So, this is the state of the art today. It was raw GPT3. I didn’t even finetune the results. GPT3 doesn’t ‘understand’ anything yet. It is relying on correlations in a huge n-dimensional space to arrive at concepts that are similar, and then piecing together the words and sentences. I never mentioned to GPT3 that there’s something called Indian Foreign Service, or provided any information that the trade service has very few officers. It picked up these things from years of commoncrawl, webtext and wikipedia data it was fed on. GPT3 can generate fictional content too. In fact all or any of the above content might have been cooked up by GPT3 on the go.

10 to 20 years down the line, any human who uses internet to research and create reports would possibly be redundant. Figures, Plots, and graphs generation will get automated too.

The above recommendations on Indian bureaucracy could have come from any DoPT official. But it came from a large language model.