GPT-3 and A11Y | Web Accessibility
My startup Tenably Labs https://www.tenably.app got access to OpenAI GPT-3 within 1 week of apply.
Since then we have developed several features to help businesses achieve compliance and meet WCAG standards.
The coolest feature IMHO opinion at the moment is code generation. We provide you with the accessible code so you don't have to go looking for a solution and continuously test each to see if they work.
GPT-3 generates it for you.
Would love to chat with others in web accessibility about this features and the many others we have developed with GPT-3.
Cheers
The main question I always have with anybody building anything on GPT-3 is what do you folks do when in a few months time or maybe six months a better model comes along and basically smokes GPT-3 in terms of all performance metrics. Which is anything than far fetched because many many Big Labs are working on this. Besides there is also GPT-Neo community effort.
Isn't then all that prompt engineering than you have build up basically for nothing? Because a new model will behave very differently.
Seems to me like a really risky move to build a business on top especially since it could be very short lived.
The idea isn't to build a business on top of GPT-3. But utilize GPT-3 within a Business/product.
But you can hardly abstract from that unless you do something more generic NLP sentence analysis and simply replace model A with better Model B anything that is based on generating text and prompting is 100% tied to GPT-3 architecture so all that knowledge is gone once a better model comes around you start more or less from scratch. Or am I missing a trick here?
No integration should be tightly coupled. De-coupled classes solve this without changes to core code. Integrations are swapped out without affecting the core platform.
Sure I'll get that I just don't see how you are going to do that, as the prompting for text generation in GPT-3 is so tightly coupled with the model itself.
I don't think you can actually take much of the core and move it to a different model sure you can take the rest of the infrastructure but building that sounds pretty straightforward anyway. The core knowledge is in the model and how you poke it to produce something useful. At least that's my naïve understanding.