Rich · Atom Tan Studio

Scrollback is launching today. A self-hosted archive for X/Twitter saves, RSS feeds, articles, media, prompts, and AI-searchable research trails. Keep the things you meant to come back to. Demo: 📷scrollback.atomtan.studio GitHub: github.com/atomtanstudio/…

Another fun fact that no one will see. Grok 4.3 is an excellent model for your OpenClaw chat agent. By far the best in my opinion. Try it if you have a Grok account. It will be obvious.

Here's a fun fact that I just learned today. If you are using Codex, you can copy a password or a API key or anything to your clipboard, and it is able to take that information and use it without you pasting it.

@thsottiaux Codex. Always Codex.

How do you refer to it

GPT Image 2 is so good at mock interfaces. I just wish there was a good way to convert these into nearly identical web apps.

Hey @OnlyTerp , any tips on getting this working? I am using Ollama cloud and am able to get either Ollama models to load or use my original configuration to get GPT 5.5 to work. The two won't work simultaneously. The model switching just gives me errors when trying to use GPT 5.5 when the other model picker is present. I gave the repo to Codex on Extra-High and it fell short.

I could have posted the github on my account so I got the credit instead of a comment tagging me. Because the time i spend building stuff like this isn't for money but it cost me alot. Its for the love of the game & my reputation. Which doesn't help me when the githib has my name nowhere 😭 I had like 5 people dm me this morning saying why did sero do you dirty. 😭

DeepSeek-V4-Pro & Kimi-K2.6 running in Codex app. Cheapest way to taste the frontier. Works w local models and they can do all the computer use stuff (:

Dear @openai, can I please ban hand waving from Codex's vocabulary or can you please add more phrases?

Has anyone got this working? I created something similar about a week or so ago. I think I have a GitHub for it, but the issue was that I either got to use my own model or I got to use the 5.5 model, but never at the same time, and also I would lose all of my chats and my projects when I would switch, which makes sense since I think it uses a profile. I set GPT-55 to Extra high, I pointed it at the repo that's listed here, as well as this chat thread and the other chat thread by the other author. The best I could get was that the other model would show up as an other model type, and it would work, but the GPT model would not work because it kept on trying to use the same URL as the other model, which obviously is not going to work.

@Chryzleenprompt I really like your prompts.

Ai muse in frame: Rain 💋 Prompt 👇 Create image: use the attached image for facial reference with high precision apply hyper-photorealistic iPhone 17 Pro front-camera behavior while preserving exact facial structure, natural asymmetry, and believable close-range proportions. SCENE: A young woman takes a close indoor selfie beside a bright cream-colored wall during quiet afternoon light. One shoulder leans slightly forward while loose hair falls naturally across one side of the face. The atmosphere feels calm, fresh, and softly intimate like a casual no-makeup moment at home. CAMERA / PHOTOGRAPHY LOGIC: Close handheld smartphone framing with slight perspective expansion around the nearer shoulder and cheek. Tight crop with soft edge sharpness and realistic smartphone HDR balancing. MAIN SUBJECT / SKIN DETAILS: Healthy believable skin texture with soft pores, faint cheek redness, subtle nose highlight bloom, natural tonal variation, and realistic daylight reflectivity. EXPRESSION / MOOD: Soft closed-mouth smile with relaxed eye contact and quiet indoor warmth. MAKEUP: Minimal makeup with natural lashes, soft pink lips, brushed brows, and barely visible cheek tint. HAIRSTYLE: Dark shoulder-length hair with soft natural movement, loose strands crossing the cheek, slightly uneven layered ends. OUTFIT: Black halter-style knit top exposing shoulders and collarbone with relaxed indoor styling. BACKGROUND / ENVIRONMENT: Soft cream apartment wall with minimal interior detail and clean daylight atmosphere. LIGHTING: Natural side daylight producing soft facial gradients, gentle shadow falloff, and subtle smartphone highlight retention. COMPOSITION: Tight upper-body framing with one shoulder closer to camera and approximately 15–20% headroom. --ar 4:5 NEGATIVE PROMPT: over-retouched skin, fake beauty filter, influencer glamour styling, cinematic grading, exaggerated makeup, artificial sharpness, unrealistic symmetry

#NowPlaying youtube.com/watch?v=RkT-aM… Good moaning

@thsottiaux @harsette @sama

@harsette @sama OK

I SWEAR TO GOD I WILL UPGRADE TO THE $200 PLAN FROM $100 AND CANCEL MY OTHER SUBSCRIPTIONS PLEASEEEE @sama @thsottiaux RESETTTT PLEASEEE I!!!!!

@Daniel_Farinax + Local

@Daniel_Farinax If we can get low latency TTS and STT, that will be the big win for all of us. I would love to never have to type again.

🚀 2 weeks ago I started building an AI harness for Qwen3.6 MLX, Gemma4 & other local models. It includes: • Native web browser the models can directly interact with • Task scheduling • Tools, etc Great learning experience and I already have several automations running daily in the background. Pretty fast on a Macbook pro M5 128GB. As models keep improving, agentic tools like this will become way more accessible.

@MayaStone420 @Merocle Good luck finding one of those for under 30k.

@Merocle 1 mac studio ultra 512gb ram would be better

It’s finally up and running! And the result is pretty decent. Now I want to put this 1U platform... 😅

That is remarkable. From docker to VM with a total OpenClaw and Mission Control migration and everything is working.

Yeah, that's a great combination. I've been using it to upscale some old 1980s music videos, and the results are remarkable. I just can't handle running it locally on my Mac due to the speed. I get 0.1fps and that's on the M5 Max with 128GB of unified memory. My 5090 laptop is still very slow at 0.3fps, but I can complete a 1080p video in roughly 1 hour per minute of video. My hope is that Topaz Labs can improve the Metal implementation.

@yannToskov @topazlabs always happy with Astra and Starlight Precise. best results

Finally Starlight Precise 2.5 from @topazlabs is available for Mac and the results are stunning! 2x upscale of this 15 sec video took 2:30 hours on my machine but I will definitely start using it regularly. Currently testing 3x and will definitely test 4x. The amount of wait time and energy consumed perhaps won't matter for anyone else but it does matter for me.

@Mugi_Records #worldgothday

A highly realistic candid smartphone photo of the same girl from the uploaded reference image, identity fully preserved. She is sitting at a small table inside a quiet Japanese izakaya at night. The camera is very close, slightly across the table, as if taken mid-conversation. She is not posing. At the exact moment of capture, she is just about to say something — her lips slightly parted, her gaze not fully meeting the camera, slightly off to the side. Her elbow rests on the table, but her posture is slightly shifting, as if she just leaned in. Details: half-finished drink glass with condensation, small messy plates, wooden table with subtle stains, chopsticks slightly misplaced, background customers blurred but present Lighting: warm uneven tungsten lighting, slight overexposure on highlights, natural skin texture, no beauty correction Composition: slightly off-center, part of her hand or glass cut off, very slight motion blur, subtle focus miss (not perfectly sharp), feels accidental Mood: intimate, in-between moment, emotionally close, slightly fragile, like a private moment captured too early Important: must feel like a real smartphone capture, NOT posed, NOT cinematic, NOT polished, imperfection is essential

I can't overstate how cool and useful it is to have Codex driving my @UnraidOfficial server. I told Codex I wanted to migrate my @OpenClaw container to a dedicated Linux VM and it started up the Brainstorming skill, followed by Goal Buddy and then built a custom Kanban board to track the progress. It then copied all of the data from the workspace folder, created the VM, downloaded Ubuntu and has now kicked off the unattended installer. This would've taken me an entire weekend in the past. Lets hope it finishes as strong as it started.

If you're doing this shit and hijacking other people's tweets, you deserve to stand on all the Legos when you get up to go pee during the night.