Key takeaways
- Adarsh Shirawalmath successfully obtained GPT-3 access from Greg Brockman in 2020.
- He aimed to use GPT-3 to predict questions for the JEE exam.
- Shirawalmath's experiment exemplified early instances of few-shot prompting and retrieval-augmented generation.
- The experiment involved inputting previous exam questions to generate potential future questions.
- Shirawalmath did not save the results but noted the experiment's surprising effectiveness.
Though he is a little older now, Adarsh Shirawalmath, the creator of Kannada Llama and founder of Tensoic, revealed in his post on LinkedIn that he mailed Greg Brockman from OpenAI in 2020 asking for GPT-3 access, and he got it as well.
The most interesting part is that he wanted to predict JEE exam questions using GPT-3, and had also mentioned that to Brockman.
At that time, GPT-3 was only known to people who were deep into AI/ML space. “The peculiar idea of generating the next words (tokens) given an input was fascinating. Essentially ‘predicting anything’ felt possible,” said Shirawalmath.
He had already applied for the API access earlier, but given the rush to prepare for the exam, Shirawalmath mailed Brockman for an early access to the model, and surprisingly within a few days got it. Earlier, Shirawalmath had also told AIM about this in an episode of Tech Talks.
“Now that I look back I do realise it just might be one of the first instances of few shot prompting and even RAG. I made it work by passing in a few questions from the previous year papers and asking it to predict what could be the next set of questions. Surprisingly, it did work to some extent,” said Shirawalamth, but he did not save the results of the experiment.
But the repository for the JEE question paper predictor still exists. Click here to check it out.
We at AIM had also tested out ChatGPT and Ola Krutrim if they can pass the UPSC Prelims exam.
