Halle Winkler

Audio Intelligence also proves my personal theory that if the leading advertisement for a new privacy-affecting feature is helping people with a disability, it's dystopian and the company is aware. In a perfect world this would be a sign of inclusion, but in this world it's a sign of framing concerns as anti-inclusion.

Very big fan of accessibility and inclusion, have enough real-world exposure to the technical realities to feel OK noting that there are certain recurring contexts where it is presented as the first-order concern.

Last thought: I generally feel positively about the value that Apple place on privacy and also on accessibility, under normal circumstances. But, the four times I thought they made the wrong call, they all involved speech, and half of those cases involved what felt like accessibility pretexts! There's a blind spot.

I've been saying "but maybe I hallucinated it" for, IIRC, ~35 years, so don't give me any guff about my word choice 😂

@stooovie@mas.to Yep, I have also experienced this temptation and I know it's unhelpful.

I had a great paper about leakage through tool use (without adversarial intent or backdoors or prompt injection) in one of my newsletter issues but I can't find it so maybe I hallucinated including it. Anyway, here is a paper about how tool use can leak data purely via argument and other artifacts that are easier to forget to set boundaries around than the actual inputs/outputs/prompt/"thinking" and can also propagate into logs, etc. I've personally seen a (local, thankfully) agent "deduce" an exploit chain via a brief commit message without being allowed to see the commit, which was alarming and made me think that it's totally possible that what looks like data leakage between agent instances might be...data leakage between agent instances. https://ieeexplore.ieee.org/document/11569042

RE: https://hachyderm.io/@cate/117330995142104460 Savage 😂

@akareilly@hachyderm.io Yes, before I did this kind of stuff I made an offline speech SDK, and these were the use cases I was constantly asked about (but large vocab was premature for on-device at that time) – for instance, medical practitioners who wanted to make notes on the go, which meant good accent recognition, which meant adaptation per-user. I feel like we've swept right past the (boring) human-helpful use cases into the (more exciting to product designers) dystopian ones.