I got tired of paying $20/mo for meeting recorders that upload everything to the cloud. So I built Listener for Mac: records, transcribes live, editable summary. 100% on-device. $5 once, no subscription. https://t.co/gaBhr5Uzqr
that 2011 siri ad is going around again. flat tire, reminders, weather, one smooth chat
15 years later the small stuff works. ask for anything real and it's still web results
meanwhile apple's on-device model on my mac does decent summaries. the gap is siri, not the chip
elorian just shipped glimpse. a visual model that zooms in or pulls video frames while it thinks, instead of turning the image into text once
small team, a few hundred gpus, post-trained on open weights. frontier labs set the bar, open weights let small teams build on it
@Madrobot_ sms parsing is such a good fit for a small on-device model. short input, fixed shape output. and bank sms is exactly the data nobody wants sitting in a cloud
deepseek shipped v4.1 flash 1017 today. 77.0 on deepswe for under $0.20 a task, open weights on hugging face.
the "flash" one is 763b params. best part for me is the shorter outputs. same task, fewer tokens. that helps at every size, even the tiny on-device models i build on.
iphone duo pre-orders opened today. foldable, starts at $1999, siri ai with apple intelligence on a 7.6" inner screen.
everyone's talking about the hinge. i'm watching how personal the on-device ai gets when the screen is that big — keep siri open next to an app, ask about what's on it, and that context doesn't need to leave the phone.
gemini 4 argon just landed in antigravity for ai pro, with its own weekly quota.
spending tonight on the next version of my mac app with it. real test for me is swift, most models still get swiftui state wrong. will report back.
yeah the memory is the wall, a quant on a mac studio is probably the most realistic. on my side it's way smaller: apple's on-device speech model plus the apple foundation model in macos 26. i use them for a mac app that transcribes and summarizes meetings without anything leaving the machine
reflection dropped beam 1.0 today. 501b params, 23b active, 1m context, apache 2.0, weights on hugging face.
not running on my macbook anytime soon. but a frontier-scale model you can own and fine-tune, with a license that doesn't fight you, is a big deal.
listener now has a free trial. 3 meetings, no account, no card. it records, transcribes and summarizes on your mac, nothing leaves the machine. if it's useful, $5 once unlocks it forever.
meta dropped muse spark open today — 816b open weights that run on two m5 ultra mac studios with mlx.
frontier-class models living on a desk used to sound like a stretch. now it's just a setup guide.