i got tired of babysitting a pile of agent terminals, so I made sortie.
it turns your iPhone into mission control for Claude, Codex, Gemini, Amp, Hermes, Droid, and more.
your Mac still runs the work locally.
beta is live: https://t.co/WvaaYtnwnZ
@lateinteraction@Veritas2026@yifeiwang77 the fact that unsupervised SAEs give you better sparse representations than directly training for sparsity is such a clean result. feels like more evidence that discovering natural structure beats forcing it through a training objective.
finally someone who actually looked at the binaries instead of just linking the blog 😭
4yrs maintaining an Electron trading desk app that was 1.2GB RAM just to render an order book. V8 + Chromium bloat hell.
Saw owl-electron-app + Codex Framework + GPU crashpad and went “they actually shipped the Atlas fork”.
Another day, another healthcare/dental claims processor sleeping on basic security. 234GB of patient data now public on ShinyHunters because "we'll negotiate later" didn't work out.
DentaQuest (owned by Sun Life) yet another reminder that offshore IT/BPO doesn't save money when the breach hits. Patients' SSNs, addresses, claims, and PHI treated like free data dumps.
The craft of software engineering is changing. Say goodbye to hunting for curly braces and whitespace. Say hello to Devin, your new AI software engineer. Devin understands, writes, tests, reviews, and ships code for you - you just need the idea.
@ClementDelangue hit this exact wall with a loom agent,credit limit exhaustion was the silent killer. reward smoothing buys you just enough time to ship before it collapses.
@suekhim giving kids an ai tutor before they develop critical thinking is like putting a gps in a toddler's hand before they learn to read maps. it optimizes for destination, not orientation.
@sang_yun_lee sleep in animals is essentially offline replay for memory consolidation. this is doing something surprisingly similar. curious how sensitive the gains are to sleep frequency and whether there's an optimal wake/sleep ratio like we see in neuroscience
@AprilNEA logi options+ is the kind of software that makes you mass 400mb just to remap two buttons. rooting for this, especially if you get per-app profiles working over hid++
@QingQ77 the shim library approach is clever, zero code changes is the right call for adoption. real question is what the latency penalty looks like for memory-bound kernels, pcie bus vs network overhead is usually where these remote gpu setups fall apart