Aleksandar Perisic

One phone selfie → a 3-camera edit. My real shot stays as cam A. Open-weight models on @ComfyUI generate B & C angles, lip-synced to my real voice, then it auto-cuts. My face stays mine. Open source: github.com/apsolut/one-shot-… #OpenSourceHiggsfield

Email is quietly becoming a conversation with an AI before it ever becomes a conversation with a human. More and more inboxes now surface a short AI summary first. The full email is secondary. That shifts the job of the message: the summary has to carry the value, not just the subject line or the design. The practical constraint is simple. AI summaries work on real, selectable text. They still struggle (or skip entirely) when the important words live inside images, banners, or designed graphics. Beautiful layouts that hide the offer in a hero image are increasingly invisible to the first layer of attention. So the opening of the email is doing heavier work than it used to. The first couple of sentences need to state the action, what the person gets, and when it matters - clearly enough that a summary can lift them without loss. Everything decorative or exploratory can come after that. Teams that treat the top of the email as pure branding are starting to see lower engagement even when open rates look fine. The open still happens; the useful signal just never reaches the reader. Worth auditing your next few campaigns against that standard: if an AI only saw the first 150–200 characters of plain text, would the point still land?

Weekend project: I turned Starnberg into a Halloween zombie game 🎃 Run the Seepromenade at night, dodge zombies in Tracht and Dirndl, take on the Lake King. Runs right in your browser. #starnberg pumpkin.apsolut.dev/

Matrix bullet time from a normal phone clip. Freeze the moment, LTX-2.5 orbits 360 around you, then back to real time. Open weights, runs on @ComfyUI. Made with @ltx_io LTX-2.5. Free to try: bulletcam.apsolut.dev

was just having some fun with Qwen for Halloween (will delete later) halloween.gastarbajter.si/?l…

Higgsfield-style multicam, but open source. 🎬 One phone clip in. Close-ups, side, wide and low angles out, cut on your words. Your real face stays camera A. No lip-sync guessing. Built in @ComfyUI with LTX-2.3 @ltx_io #OpenSourceHiggsfield xitter.gq/apsolut/status/2115181…

Try it live, no setup: multicam.apsolut.dev/ Upload a short selfie clip (up to 30s) where you're talking. It picks the cut points and generates the other angles on @ComfyUI Cloud. Free runs are limited, but you can paste your own Comfy key for more.

If you`re obsessed with AI, let’s be friends. #buildinpublic

Wanted to try it again, was using Cursor for couple months last year! and... I dont mind many things, I can adopt - but .. 80% time I dont know do I need to answer something, is it working in background ... @grok how we went from great tool into this ? #cursor

Lovable wrote the best code of the four: real TS modules, a unit-test suite. It was also the only one that broke the single-file rule to do it. Its own NOTES.md calls it a "platform deviation".

Inspired by the Tesla Roadster hover demo. What came out: - 3 of 4 chose one HTML file with an inline script. Nobody told them to agree. - Replit built an Express API + Postgres schema, then imported none of it. - Bolt declared Supabase, React, Tailwind. Never called them.