The subject that has been put under the science communities’ spotlight, a topic that has been a hot topic in recent years, artificial intelligence, is multi-purpose and is being developed to envelope even more regions of our lives. From asking general questions to driving cars, creating PowerPoints to writing complete documents, AI has got it covered. Well, how about using this jack-of-all-trades technology to create more of it? That is exactly what the creators of Cappy wanted to achieve. The purpose of Cappy is to take responses from Large Language Models (LLMs) and score them on how well they performed on the task provided to them by the user. Cappy surpasses LLMs in its category by large margins and will hopefully reduce computational costs.
We use AI like Google’s Gemini and OpenAI’s ChatGPT daily, whatever the reason may be. But how did they come to a point where we can generally trust them? Obviously, they are coded to function properly, but most importantly, they are then trained to “learn” how to complete their objective in an efficient and correct manner. The process of teaching the LLM is rough, as it takes several to hundreds of billions of parameters to function, which leads to the monstrously high memory capacities of GPUs and TPUs. Not only does this bring about storage challenges but also expensive and inefficient training. Certain strategies involving prompt tuning and adapters significantly improve the storage problems, though they only stay there and do not help with the training itself.
Cappy’s novel approach to addressing this issue is an automatic check to see if the LLM’s response is in line with the request. Cappy receives both the input and output entries during the training stages of a LLM and grades the LLM’s response with a score between 0 and 1. This acts as powerful feedback to the training subject and reduces the time and resource demand of the training. Since it does not require the LLM’s parameters, it is compatible with open-source as well as closed-source LLMs, solving another difficulty other software has encountered. Cappy is also not limited to training protocols: it acts as an auxiliary component to LLMs. It enhances their own capabilities by prechecking their responses, which then sparks an increase in the quality of response in the LLM’s answers. It is also worth noting that Cappy is either equal to or higher performing than any other similar software with much fewer parameters. For example, OPT-IML with 175 billion parameters slightly underperforms Cappy, which contains only 360 million parameters; this is a clear sign of how efficient Cappy is.
To sum up, Cappy introduces a great way to enhance current methods for developing and utilizing AI. This is a serious improvement and will reduce the stagnation of new research in the AI industry. Hopefully, with more advances like this coming up, we will move on to a different age of technology.
