I am curious about your view on accessibility tooling. Often screen readers or speech to text tools would effectively be external pieces of software instrumenting for instance browsers through their APIs. And surely using local AI going forward could enhance the tooling while also making the users less dependent on a small market charging exorbitant prices for poor user experience. What viable paths do you see for users perfecting user agents for themselves when the big browsers further lose their plot and focus on what's fun to maintain or profitable to ship instead of optimising for every user individually?

It's just an unfortunate reality that people with disabilities live with.

Some of them are already prevented from operating heavy machinery because doing so would put everyone at risk, so there is precedent here. I'm in favor of making concessions to the very unfortunate, though at some point you have to look at percentages, expected impact and decide what is reasonable and what isn't.

To your other question, in the last week there have been entire CPUs embedded within weights. Clearly it's possible to publish "open models" with an agenda so then what assurances you have that the enhancements provided by local ai would be more focused on the plot than what the big browsers do?

Specifically how would a person with disabilities be able to tell the difference between improved instrumentation and conditional access to information, as we have in other media? Consider this in the context of "compared to superintelligence, we all have an intelectual disability"