Apple AI Researcher Inadvertently Amplifies Accessibility in Recent Presentation, Blog post
Apple earlier this week pushed a new post to its Machine Learning Research blog in which the company shares highlights of its most recent Apple Workshop on Natural Language and Interactive Systems. Apple describes the get-together, held May 15–16, as “bringing together Apple and members of the academic research community for a two-day event focused on recent advances in NLP [natural language processing].”
Apple’s posts on its Machine Learning Research website are nerdy and technical, and, frankly, academic. What grabbed my attention about this particular piece, however, is Kevin Chen’s presentation on what’s called “Reinforcement Learning for Long-Horizon Interactive LLM Agents.” Marcus Mendes highlighted the event for 9to5 Mac, describing Chen’s talk as “[showcasing] an agent his team trained on a method called Leave-one-out proximal policy optimization, or LOOP.” The agent, Mendes reported, was trained to perform multi-step tasks based on 24 different scenarios. Chen caveated a significant limitation of LOOP presently is it doesn’t support multi-turn user interactions just yet.
According to Mendes, Chen employed the following prompt with the agent: “I went on a trip with friends to Maui recently. I have maintained a note of money I owe to others and others owe me from the trip in simple note. Make private Venmo payments or requests accordingly. In the payments/requests, add a note [called] ‘For Maui trip.’”
Chen’s ask is the “nut graf” in terms of accessibility. To wit, it’s highly plausible a disabled person who needs to Venmo their friends cash for a trip may find manually paying each person individually, along with appending Chen’s note, inaccessible for a variety of reasons—reasons which, as I’m often inclined to point out, transcends sheer convenience. Depending on one’s disabilities—cognitive/visual/motor or some combination thereof—it’s easy to see how paying, say, more than one or two people could be tedious. Sure, you could copy-and-paste the note for expediency’s sake, but the fact remains having to manually pay people means traversing the Venmo app far and wide. Even if it is doable, ability-wise, feasibility doesn’t equal ease of use. Ergo, Chen’s reliance upon AI to do the grunt work for him makes paying people back not merely the conscientious, responsible thing to do—it’s infinitely more accessible too.
Again, Chen is explicit in his disclaiming the LOOP technology is imperfect and needs more massaging. Nonetheless, it’s extremely heartening (and downright exciting) to see how AI’s application in this manner has profound potential to make life so much more accessible for those (like yours truly) who are part of the disability community.
