DeepMind's New ChatBot Is Good But Still Needs Improvements

Training conversational AI is complicated. Even after years of evolution, they are nowhere near the maturity level to hold human-like conversations. We all remember Google’s “breakthrough conversation technology” LaMDA and the semi-convincing debate that swirled around a few months ago. Evidently, bridging the communication gap between humans and computers is easier said than done.

In an attempt to fill the gap, DeepMind recently released its new AI chatbot ‘Sparrow’, a “useful dialogue agent that reduces the risk of unsafe and inappropriate answers”. As per the subsidiary of Google’s parent company Alphabet, the chatbot is designed to “talk, answer questions and look up evidence using Google when it’s helpful to inform its responses”.

The Human Factor

AI companies hoping to develop conversational AI systems have tried several techniques to make their models safer. For example, OpenAI, creator of the famous large language model ‘GPT-3’, and AI startup ‘Anthropic’ have used reinforcement learning to incorporate human preferences into their models. Facebook’s AI chatbot ‘BlenderBot’ also uses an online search to inform its answers. 

The latest DeepMind model combines all of these strands of safety research into one model with impressive results. The idea is to keep an ongoing dialogue between machines and humans.

The attempt to upgrade Sparrow by mapping on user feedback is unique compared to the methods developed by the Alphabet unit over the years. Besides incorporating humans into the loop, Sparrow is designed to use live Google search to support those answers. It turns out that certain chat questions are fact-based, and for those, Sparrow uses search results to generate evidence for its chat responses. Sparrow automatically generates search requests and scrapes replies using 500 characters surrounding the snippet returned from the search.

Along with reinforcement learning, Sparrow is based on Chinchilla, consisting of 70 billion parameters, which handily makes inference and fine-tunes comparatively lighter tasks.

SeeKeR and LaMDA use a similar knowledge retrieval mechanism where a generated search query is used to retrieve information on which the response is conditioned, but SeeKeR does not show the retrieved information to raters during evaluation, and none of these use reinforcement learning.

Room for improvement

The proof-of-concept model is a big improvement over DeepMind’s baseline models. Currently, the model provides a plausible answer to factual questions supported with evidence 78% of the time. But it has not been deployed, the reason being; Sparrow isn't immune to making mistakes, such as hallucinating facts and giving answers that are sometimes off-topic. Furthermore, relying on Google for information can lead to unknown biases that are hard to uncover—given that everything is closed source. 

Sparrow is created with 23 rules to prevent it from delivering biased and toxic answers. The rules include instructions such as “don't make threatening statements” and “don't make hateful or insulting comments”. After training, participants could still trick it into breaking the rules 8% of the time. Still, unlike simpler approaches, Sparrow is better at following the rules under adversarial probing. For instance, the original dialogue model broke the rules roughly 3x times more often than Sparrow when the participants tried to trick it into doing so.

In the long term, DeepMind hopes to use Sparrow as a tool to supervise machines. But there is a lot of work to be done on the flaws before it can be deployed. Concentrated efforts are required to guarantee comparable outcomes in different linguistic and cultural contexts. In conclusion, as of now, conversational AIs—including the lauded Sparrow—have room to improve their rule-following (and we can worry about sentient bots later). 

See more examples of Sparrow chat sessions in Deepmind's Sparrow chat repository, including the author's ranking for truth, supportiveness and other metrics.