Trust and alignment
#89 - Nov.2023
Trust will continue to be the backbone of good product experiences. This is not different for AI-based products.
AI models, in particular LLMs, bring a bigger challenge to serve common affordances. Is the so-called alignment problem. The alignment problem in AI is basically reduced to how we create an experience - call it a copilot, agent, AI buddy, etc - to behave in accordance with what we want.
It seems a fair ask for a generative technology with the power of creating content in multiple forms such as text, audio, images, or video. Even more important if we expect the system to behave in accordance with our personal values and ethics.
The HHH framing is a good starting point (Askell et al. 2021). This framing defines alignment in the context of how helpful, honest, and harmless can a AI-system be.
Aligning to our expectations starts with ensuring the system performs the desired task concisely and as effectively as possible (helpful). The task/answer should be accurate or acknowledge the level of uncertainty (honest). Finally, AI systems should not be offensive or discriminatory, or promote dangerous behaviors (harmless).
PS: Interesting paper of potential solutions on how to solve the alignment problem:
