ChatGPT’s Adoption is a Lesson in AI Usability
Sam Altman mentioned they had the model for ChatGPT ready “10 months or something” before they actually launched a chat app around it. The mass adoption of ChatGPT is a lesson in usability in AI. The original Open AI Generative Pre-trained Transformer (yes, GPT) has actually existed since Summer 2020 and the playground to use it since 2021, where you define multiple control parameters to use the model e.g. temperature (randomness), result length, frequency/presence, stop sequence — too technical for a majority of the public.
What has helped take this model to the masses and really drive value for users (other than training the model on millions of tokens from public domain knowledge) has been an incredibly easy to use chat interface that allows people, even the non-technical ones, to tap into the powers of the pre-trained model without several control knobs that they don’t understand.
In my view, the next generation of most AI software tools will be building layers of abstraction on top of this powerful lower layer of abstraction (ChatGPT and also other pre-trained models) to make AI systems even more usable. It’s interesting how this is analogous to how computer systems have generally evolved over time, for example, higher level more-usable programming languages created on lower level languages: Machine Code -> Assembly -> C -> Python.
I remember coming across this anonymous quote/sentence somewhere in my early days in Computer Science, which is relatable in this context:
Power lies at the lower level of abstraction. Beauty lies at the higher levels of abstraction
(Beauty in this context, being “Usability of AI systems” and power being the underlying pre-trained models. It’s fascinating along with power, how increase in usability, has allowed these systems to create value for a much larger audience versus it’s initial iterations).
