Ola Krutrim had launched its first public-facing AI model yesterday, but users tinkering with the model have been met with some unexpected responses.
Ola Krutrim’s newly-launched AI model has told several users it’s made by OpenAI, the company behind ChatGPT. The model said this in response to being asked who it was created by. Several users reported seeing these results within hours of the model’s release yesterday.
“Something seems super fishy,” wrote X user Raghav Arora. “KrutrimAI says it was created by OpenAI,” he added. He shared a screenshot of him asking the model “Who are you”, to which the model replied “I am an artificial intelligence created by OpenAI”.
Another X user also reported seeing similar results. “Yes, my large language model is developed by OpenAI…OpenAI created me,” KrutrimAI’s model said.
At this point, some users began to troll Krutrim, claiming it was just a “wrapper” around OpenAI’s ChatGPT. This meant that Krutrim was simply passing on user questions to ChatGPT, and passing off its answers as its own.
Others questioned Krutrim’s valuation. Krutrim had raised $50 million at a valuation of $1 billion, and wrapper that simply passes off OpenAI’s responses would be worth very littel in comparison.
Ola Krutrim, however, took cognizance of the issue, and said the responses werre caused by a data leak from an open source dataset the model had been trained on. “Thanks for sharing this, it is very helpful and will help improve our product. We investigated the issue and found the root cause to be a data leakage issue from one of the open-source datasets used in our LLM fine-tuning,” Krutrim’s official X handle said. Krutrim appeared to have quickly fixed the problem — when prompted, Krutrim AI now says it was built by Ola Krutrim.
It’s unclear what caused Ola Krutrim’s model to say it was created by OpenAI. There are several ways today to build an LLM like Krutrim AI. One way would be to build a foundational model, which would mean a model that’s trained from scratch for several months using expensive hardware. Another way is to fine-tuning an existing model, which involves training an open-source model such as Meta’s Llama on a dataset that’s tailored for the model’s specific needs. Yet another way — which would also be the easiest — would involve simply paying ChatGPT to get its answers user questions, and passing them off as your own.
It’s unclear which of these approaches KrutrimAI followed. Its official response suggests that the reference to OpenAI came from a dataset that contained ChatGPT’s responses, and Krutrim happened to inadvertently train on those responses. This is entirely plausible, as ChatGPT responses are present in several open-source datasets. But this wasn’t the only reference to OpenAI that is present in Krutrim’s responses — when we asked Krutrim when it was built, it said it was built in 2016, which happens to be the year OpenAI was founded. And when a user asked what the current date was, Krutrim said it was September 2021, which happens to be the knowledge cutoff for ChatGPT. While the jury is still out on what Krutrim AI’s architecture is really like, the fact that India’s first AI unicorn has told users it was made by a different company isn’t the best look for India’s upcoming AI ecosystem.