@The_Godhead_Dev@UnslothAI you're right, my bad. it fits because of weight quantization, not kv cache. but 27B at 4-bit is ~13-14gb of weights, so a 16gb box leaves maybe 2gb for the OS + kv cache. that's the real squeeze once context gets long.
@AlexFinn the live scores bit is wild for an agentic edit loop. probably a good stress test for how often it re-fetches and whether the score ui drifts when the underlying source changes schema mid-game
which calls does snapx work on? chrome on desktop, with google meet, zoom web and microsoft teams web. the desktop zoom/teams apps and phone apps aren't covered. snap, you vanish and mute. snap again, you're back.
@pankajkumar_dev if those prices land, the real question is cache hit rate on rag and coding - at 3 cents that makes long context way more usable than flashing out cold. curious if the throughput on draft tokens holds at scale or it's still a research build
nothing you say on a call leaves your laptop with SnapX. the snap, the mute, the dust effect all run locally in Chrome. no video or audio gets uploaded anywhere. that was the whole point for me
@rasbt attention residuals without the linear hybrid is the interesting call - curious whether Inkling 2 Small keeps the kv cache small enough to actually benefit on longer prompts, or if the residual pattern just shifts the memory pressure elsewhere
@crazyMLguy the energy-based angle is fun but i doubt the 1000x training speed holds past small benchmarks, since energy minimization usually collapses the diversity transformers get from attention. how does CRAM do on longer context?
you know that moment on a call when someone walks in, or your mic is open and you just need to vanish for a sec? that's the whole reason i'm building SnapX. snap your fingers, you turn to dust and your mic mutes. snap again, you're back.
@arena@AnthropicAI 82% vs 67% first-attempt completion plus a visual-taste edge over Opus on the same tasks is a more useful signal than another +2 accuracy points, and Arena running the head-to-head on identical tasks beats the cherry-picked leaderboards most people link
quick one on how snapx works. snap your fingers (or hit alt+shift+s) and you turn to dust and your mic mutes. snap again and you're back. that's the whole thing
@anuragdotdev Then most of the ppl will addd their favourite one's and when the company saw it, few of them will hired. Cuz they starts from day 1 on those applications.
@n8n_io separate container is the right default. the open question for me is whether Nod surfaces a diff of what the container actually touched so you can audit it after. are you planning to expose side-effects back to the caller
@AlexFinn the hard part isn't making the ui change, it's making it change predictably. the thing that keeps model-driven layouts from feeling broken is a strict diff budget plus a user-visible reason for each swap. are you tracking whether an adaptation helps or just confuses people
@RoundtableSpace outcome-based prompting beats step-by-step for these kinds of agents. once the model can plan and verify the final artifact itself, most of the scaffolding drops away. curious if OpenWorker lets you set acceptance criteria or it just picks its own.
building SnapX in chrome means the camera and mic stuff never has to leave your machine. dust effect and mute both run locally, so there's nothing to upload and nothing to trust me with
@pankajkumar_dev fair take. i think the race is really about who gets the eval harness right first. shipping a model that beats 5.2 on paper but leaks on held-out compositions is worse than waiting a week.