In order for a virtual assistant to be useful, the agent should do more than just information retrieval and basic chit-chat. Rather than pattern recognition on the response level, the agent should be able to perform pattern recognition on the discourse level so it can mimic human-reasoning (even as true understanding remains an elusive goal). If a model were to reason about an utterance, it must have been trained to do so. Furthermore, we argue that such training must be explicitly performed through (weakly) supervised learning, rather than implicitly extracted from a large pre-trained LM (eg. through careful prompting).
The first step to truly becoming an AI-first company is to adopt a data-centric view, which naturally implies taking data collection seriously as a core competency of the business. Even before involving any sophistcated algorithms to improve data quality, there are already many best practices to consider when performing manual data collection. At a high level, this can be broken down into improvements
Data is the new oil. It underpins an undeniable aspect of growth in the popularity and dominance of deep learning. But is all data created equal? What makes some data points worth more than others? How could we go about calculating what each data point is worth? If data is so important, we should certainly want to have a proper way to assess its value. To do so, we should begin by recognizing that
This post will explore meta reinforcement learning with only minimal math. We attempt to dive into the core concepts without making it too complicated.
When the web first appeared, many folks adopted the new technology by porting over existing physical assets (ie. magazines and newspapers) to appear on a website with almost no changes. When smartphones began to take hold, those same people ported over their existing websites onto a smaller screen, again with no discernable changes. Rather than embracing the new medium, people simply applied the content of the old onto the new. In hindsight, it seems obvious that the builders and makers should have been developing for the new medium, but perhaps paradigm shifts aren’t so obvious when they are happening. What lessons does that teach us about the shift into building AI-first companies?