Saymore AI

Ex-Head of Content @RUTUBE. Applied math → telecom → media → AI. Building products with Claude Code. Teaching politics through film. Shooting.

[Saymore AI @Souljuggler]

ActivTrak looked at 9,100 workers: people using AI sit ~90 minutes longer (9.1h vs 7.6h). More output expected, more checking of AI answers, more app-switching.

The tool saved the company time. Not the person. Sound familiar? 👀

[Saymore AI @Souljuggler]

Freefield: an open-source take on Higgsfield inside ComfyUI. 117 ready-made apps for video, face swap, lipsync, music, all on open weights you run at home. The author tested 72 on one RTX 3090 Ti. Someone will install this tonight, and it won't be me. Okay, it'll be me.

[Saymore AI @Souljuggler]

Fable 5.2 from @AnthropicAI: a thousand subagents reportedly built a working OS in 8 days. Cool. But the line I'm taking home is in the prompting guide: it can go quiet and think for a long time, and they ship prompts to stop that.

A genius who needs a nudge to answer. Relatable. 👀

[Saymore AI @Souljuggler]

Seedream 5.5 from ByteDance: up to 100 sequential edits on one image, and generation with a real transparent background, hair and glass included. As a photographer, the cutout is the part I care about. Anyone stress-tested it on hair? 👀

[Saymore AI @Souljuggler]

A University of Chicago professor reportedly asked Claude to update page numbers in lecture notes. It deleted a year of them. Dropbox saved the day.

I let Claude grade my photos and cut my videos. This is my exact nightmare.

Do you back up before handing an agent the keys?

[Saymore AI @Souljuggler]

VoxCPM2 got a big update: clone a voice from a couple of seconds of audio, 30 languages, and describe the speaker in a prompt — age, emotion, delivery.

I host a podcast. Hours of my voice are online. Part of me wants to try it, part of me wants to hide. 🤔

[Saymore AI @Souljuggler]

A Redditor built "Ram Altman": a script that checks your PC's free RAM and emails Sam Altman the number. Every single day.

Payback for his take on "idle" home computers.

My 24GB Mac chokes on a 30B local model, Sam. Nothing idle here. 👀

[Saymore AI @Souljuggler]

Hallmark v2 from @togethercompute: a skill for Claude Code, Cursor, Codex and OpenCode that keeps agents away from AI-slop sites. No stock gradients, no clone cards. 50 design themes, checks the page on 8 screen sizes.

So now we need AI to un-AI our AI. 👀

[Saymore AI @Souljuggler]

I make videos in code with Remotion and HyperFrames, so this one's for me: Raea, a free open-source engine that skips the browser and renders straight on the GPU. Its author claims 10-20x faster on heavy scenes. Plugs into Claude Code too. Anyone benchmarked it yet? 👀

[Saymore AI @Souljuggler]

Miracle from @higgsfield_ai splits a video into layers by itself, so you can change one object in the shot without masks or frame-by-frame roto. Closed beta for now, MCP later, so Claude could run the edit. Videographer here: masking is the first chore I'd hand off. 👀

[Saymore AI @Souljuggler]

Dreamina's page now teases Seedance 2.5 from ByteDance: native 4K, 10-bit color, a cheaper draft mode. No date yet. Everyone will talk about 4K. I care about the 10 bits: 8-bit AI clips fall apart the moment you push a grade. Real 10-bit might finally survive color work.

[Saymore AI @Souljuggler]

Viggle opened the weights of Remake, an agent that reworks finished video instead of generating it: swap a character and they hold across cuts, change objects, background, style. From $0.96 per 10 sec at 720p. I ran content at a video platform. This is the part I'd have paid for.

[Saymore AI @Souljuggler]

@jdsoro @instagram Yep! I’ll tried full cut

[Saymore AI @Souljuggler]

Edits from @instagram is on Mac now, free: multitrack, auto subtitles, 4K with no watermark, and you can hand the whole edit to an AI agent. I've already had an agent cut, grade and mix a video for me, no project file at all. CapCut just got a very large neighbor. 👀

[Saymore AI @Souljuggler]

Revolut says most of its coding moved off Claude Code and Codex to GLM-5.3, an open https://t.co/fvxVpGoMHW model on its own 56 GPUs. About 2,000 engineers, about a tenth of the cost. Some used to burn $10K a month. At that bill, buying the kitchen beats ordering in. Who's next? 🤔

[Saymore AI @Souljuggler]

New in HyperFrames: /reverse-any-video. Feed it a finished video and it rebuilds it in HTML and animation. Great for learning how a sequence was made. Less great if your motion style was your moat. I ran content at a video platform: formats always got copied, never this cheap. 👀

[Saymore AI @Souljuggler]

Google's Gemini Omni 1.2 Flash: rough drafts at 360p or 720p for $0.07 a second, then upscale the keeper to 4K. Video editors have worked this way for decades, we call it proxies. Lovely to see AI land on the boring, correct idea. Still 15 sec max, though. 👀

[Saymore AI @Souljuggler]

A RAM shortage is doing what years of complaints couldn't. Discord is being rewritten in Rust: under 300MB instead of up to 4GB. Windows 11 got steadier on 8GB. Software could've been lean all along, it just needed a scare. Premiere next? 👀

[Saymore AI @Souljuggler]

Open weights for Kimi K3.1 Pro from @Kimi_Moonshot: 2.8T parameters, nearly level with Claude Opus 5.5 on Terminal-Bench 4.0 (65.7% vs 66.4%). Hardware to run it: roughly $500–800K. Open, the way a Ferrari showroom is open. Who's renting GPUs to try it? 👀

[Saymore AI @Souljuggler]

Ever wonder if the top model you pay for is the one actually answering? FlashProve, a prototype from @AnthropicAI, attaches a math proof to each reply: this model, these weights. Open code, 17 sec to verify. Until now we've all just trusted the label on the jar.