It has been about a year since I was summoned to help build a virtual talking head based on OpenAI GPT-3.5. Here is a reflection on the impact of Generative AI after a year of the technology.
On Hype
When AI is marketed on toothbrush packages, you know we’ve hit peak hype. Generative AI is an exciting technology that will, over time, permeate all of society, but currently, we are somewhere near the top of the hype bubble. Deriving sustainable value from these tools is much more complex than most people realize, despite all the cool (scary?) demos you see in the media.
I look forward to seeing the hype dissipate so we can focus on products of real value.
On Value
GenAI is most valuable in specific, narrow use cases. For example, using generative AI to tweak an article you’re writing will give you good results (most of the time). Using generative AI to schedule your life, attend meetings, and respond to emails might sound appealing, but it will result in something terrible. Keep use cases as niche as possible and focus on building robust systems.
On Production
Once you’ve found your specific and narrow use case, there is one more elephant in the room: production. Putting Gen AI applications into production is extremely difficult and much more akin to traditional software engineering than traditional ML engineering. Hallucinations, lack of observability, and unreliable LLM APIs are just a few. I am most excited to see how this space evolves over the coming years. We are indeed just scratching the surface of the tooling to make deployment more accessible and more reliable.
On Future
Despite the negative perceptions, I am optimistic about generative AI’s long-term potential. Generative AI can unlock tremendous value, but we must ensure that we approach adoption carefully and thoroughly. As a society, we share a responsibility to shape how generative AI is used to benefit us all.