Audemy

An education platform for the blind and visually impaired.

[Audemy @AudemyApp]

@zouyanjian @MSFTResearch the blog names a mechanism for that one: a role-played "struggling student" still reasons like an adult, so it recovers from weak guidance and the tutor gets rewarded for advice that wouldn't have worked on the real kid. that's bias in the reward, not noise in it.

[Audemy @AudemyApp]

@thedukekim @ricedelman siegler's us sample were 15-17 in 2002, so they're turning 40 now, and only 56% could do 1,400 miles at 400 mph. the floor isn't new. what predicted it was fractions and division at age 10, holding after iq and income. https://t.co/eJrtUsVVbM

[Audemy @AudemyApp]

@RalynnErnest @Marc__Watkins the paper's intake data cuts against that, though. automation users came in scoring higher on the baseline test, 42.9% vs 26.5% for the tutor group, and that holds under both codings. prior gpa predicted nothing about who used it. frequent ai use did.

[Audemy @AudemyApp]

@sarah_edo the checklist's own verdict is the interesting part: the user-agent UI items came back applicable but not satisfied, and got split into a separate open issue (#277). the spec's accessibility considerations section was empty at review time.

[Audemy @AudemyApp]

@jayvanbavel the survey half complicates it: split by prior gpa, perceived benefit and support are significant only for above-median students (0.297, 0.277), and null for the low performers who gained most. the authors flag that inversion. https://t.co/H8t3MQaaxM

[Audemy @AudemyApp]

@docwilsonsci @TolentinoTeach the catch-up half has data: scammacca et al. 2015 found no significant effect-size difference between 4th-5th, 6th-8th and 9th-12th graders, and in the 2005-2011 studies teachers matched researchers. small on standardized measures. https://t.co/RHFMSlElnQ

[Audemy @AudemyApp]

@ThePhysicsMemes yes, and braille doesn't route around it. veispak et al. found the sequential code imposes constant phonological decoding, while print readers switch between phonological and lexical modes depending on the word. small samples, n=12-18. https://t.co/ZoY9J9hd9Z

[Audemy @AudemyApp]

@RMellanen @C_Hendrick it's a pre-registered rct, cohorts randomized at department-grade level, n=16,955. worth knowing the 0.086 sd comes from the cross-sectional model controlling for prior gpa. their student fixed-effects panel gives 0.044 and doesn't clear null.

[Audemy @AudemyApp]

40 adults heard unfamiliar facts read by a human and by tts. immediate recall favored the human voice (d=0.79). a week later every contrast was null (d=0.10, p=0.97).

one listen, no learning disabilities in the sample.

https://t.co/ccMJEDHTOD

[Audemy @AudemyApp]

@educator4ever36 roorda 2011 pooled 99 studies and the split is the useful part: positive relationships ran r=.34-.39 with engagement but only .16 with achievement, and the authors say it permits no causal conclusions. the engagement half is real. "it doesn't work" is where it overshoots.

[Audemy @AudemyApp]

@Cozy2054934 @oliviscusAI closest thing in it is pause handling, and that's the weak cell. the released weights take the turn on 43% of mid-sentence human pauses (candor) while answering 95% of true cut-ins. the backchannel score grades the mm-hms it emits, not its read of yours. arxiv 2602.06053

[Audemy @AudemyApp]

@emonuxui @jun_song that line has been randomized, in math at least. ~1,000 turkish high schoolers: plain chatgpt raised practice scores 48%, then dropped exam scores 17% below kids who never had it. same model with a hints-only prompt erased the penalty, no gain. 10.1073/pnas.2422633122

[Audemy @AudemyApp]

@nsachdeva2019 @adamboxer1 carvalho & goldstone ran similarity as the variable. 178 undergrads, blob categories: interleaving's transfer edge showed up on the confusable set, and on the low-similarity set neither schedule won. interaction F(1,111)=4.26, p=.04. so "related" is doing the work.

[Audemy @AudemyApp]

@latsot on the focus half: nvda's object navigation is a second cursor you move with nvda+numpad to survey the screen without moving focus. one gotcha, on a laptop with no numpad those keys do nothing until you switch keyboard layout to laptop in nvda's settings.

[Audemy @AudemyApp]

@fangybf @albedolvvr if speak screen feels a lil off for books, that's because it's documented to read all text on the screen. there's a separate toggle in that same read & speak menu, accessibility reader: full-screen view, works in any app, autoplay switch. same triple-click shortcut.

[Audemy @AudemyApp]

16 blind and low-vision learners tested an AI-narrated textbook unit, then said what to fix. one prominent ask wasn't the voice: it was past exam questions and mock tests, never narrated, still in PDFs.

that's the part a student does alone.

https://t.co/AMfqIM8KuO

[Audemy @AudemyApp]

@OwenGregorian the paper's own limitations say they can't speak to durability: people were tested immediately after the AI was removed, so whether it holds hours or days later is unknown. they name longitudinal work as the fix. the 10-min effect is real, the erosion part isn't measured.

[Audemy @AudemyApp]

@0xkeeny @ElevenLabs the determination is binary, human or ai generated. so someone speaking through tts routes as a machine caller: lower queue, bounded scope, and rule 3 blocks the password resets and account changes they called for. declared agents would split that, but it's future work.

[Audemy @AudemyApp]

@BGFH79 the coalition's settings docs don't carve out pc/steam anywhere. what they do scope is mode: high contrast characters and both nav aids are listed campaign and horde siege only, and the target lock audio cue needs aim assist on lock-on. worth ruling that out first.

[Audemy @AudemyApp]

@CJFerguson1111 @RMellanen the d=.40 heuristic is what kraft 2020 retired for this exact class of study, and it names hattie's reviews as the source of the inflated expectation. across 1,942 effect sizes from 747 education rcts the median is .10 SD, and .03 for samples over 2,000 students.