As a completely ignorant infra engineer, is AI not also a potential solution that avoids rewriting entirely? Like, it feels like translating between visuals and audio signals would be one of the few things it’s really great at.
Perhaps a bit wasteful, but feels more likely than getting the entire software industry (including companies with inverse financial incentives) to get on board.
The two problems I see with the AI-screen-reader approach is wastefulness and confabulations. It is one thing to have to deal with your screen reader not reading something. However, it is quite a different thing having to hope that your screen reader doesn't outright lie to you. And they do. I recently had BeMyEyes (uses OpenAI vision model) describe an iOS screenshot to me. It imagined an error which definitely was not on-screen. I cliamed I have a storage full warning, for a 256GB iPhone. Thing is, storage is not full, and the iPhone has 128GB. I really dont want such technology to be used for screen reading, its bad at best, and dangerous at worst.
No. There are AI systems currently in-use, and I expect they could be improved significantly with a purpose-built model (e.g. one that describes what's actually there, rather than describing how it's supposed to make you feel), but there's no way to bring the friction down without compromising accuracy (even further). Describing an image involves editorial decisions that just aren't present when doing language-to-language translation, there's no way to avoid that, and transformer models (etc) are bad at editorial decisions more complicated than "rank the salience from most to least".
https://www.bemyeyes.com/ serves a similar role: higher-latency, different privacy concerns, way better accuracy (most of the time). You can see why this isn't the same genre of thing as "make computer systems that people can actually use without assistance".
Perhaps a bit wasteful, but feels more likely than getting the entire software industry (including companies with inverse financial incentives) to get on board.