Roman Carrot

@SSSvinosvin Wow. Already the 5 part

PART 5. Building in public I made my own MEDIA CONVERTER What's up, guys? I often have to edit videos and I'm really tired of searching for converter sites And almost always, if I use a reference video, the audio track doesn't work when I export it to DaVinci Resolve I don't know why But I have to extract it from various websites, including converting the video to another format after editing. Of cource, they have free limits, and you often hit them So, I figured it wouldn't be that hard to create my own converter, so I decided to share it with you It's pretty easy to install If needed, I can create an installer to avoid any unnecessary hassle with GitHub =================================== 17 input types. 11 output options. a small desktop converter with bundled FFmpeg -> video: mp4, mkv, mov, webm, avi -> image: png, jpg, webp, bmp, tiff -> audio: extract WAV from video No browser uploads No separate FFmpeg install share your projects in the comments! =================================== github link: github.com/svinosvin/Conv…

@SSSvinosvin this looks really scary

A 23-year-old creator built an AI archive video that feels more real than most horror films No actors No rented location No camera crew Just an empty room, a fixed camera, and the right story around the footage He did not make the video look cinematic He made it look like evidence The clip is framed as a recording from a camera installed by specialists of the 25th Archive inside Object No. 7. That one detail changes everything It is no longer “AI horror.” It is a file from an organization that should not exist. Pause at 0:02. The room is almost too empty. Flat walls. bad lighting. dead geometry. Nothing jumps out. Nothing explains itself The frame just sits there and makes you search every corner That is the hook. The horror is not the creature. The horror is the system around it. Object No. 7 -> camera installed by specialists -> archive language -> fixed surveillance angle -> no music -> no explanation Most creators try to make AI video look expensive. He went the other way. He made it look boring, compressed, procedural, and slightly broken. The kind of footage that feels like it came from an internal folder nobody was supposed to open. Pause at 0:07. This is where your brain stops watching normally. You start scanning the frame. Doorways. shadows. empty space. The background becomes the main character. That is how found footage works You do not show everything You make the viewer search for it The smart part is the title. “25th Archive” implies history. “Object No. 7” implies there are more. The video becomes one piece of a bigger fictional system. And that is why it works The creator is not just generating a scene He is generating context A file name. A camera source. A reason the footage exists. A reason it was recorded. A reason it feels incomplete That is the new AI video skill Not better monsters Better evidence Save this The next viral AI horror will not look like a movie It will look like a leaked folder reference: @archivus25" target="_blank" rel="nofollow noopener">tiktok.com/@archivus25

@SSSvinosvin it was really intresting article. I actually bookmarked this

@SSSvinosvin awesome technologies

AI tracking. 18 meters. 0.003-second response. No cameraman in the room. One guy walks into a rented studio Black hoodie Softbox on the left, warm lamp in the back The setup looks like a two-person shoot. It is not On the tripod sits a FeiyuTech SCORP 3 with a mirrorless camera bolted on top. He presses one button and walks away. The camera follows Pause at 0:16 He is moving left across the frame. The gimbal is already panning No operator No joystick The AI tracker locked onto his face and is following every step Pause at 0:22 He stops The camera stops He shifts his weight. The gimbal adjusts. The shot stays centered like a cameraman is reading his body language in real time. 3.5 kg payload. Every mirrorless body with a mid-range lens. The handle detaches into a wireless remote from 12 meters away. The entire rig costs less than a single day rate for a camera operator. Course creators are filming tutorials solo. Product reviewers are shooting walkthroughs without asking a friend to hold the camera. Real estate agents are filming property tours alone A year ago, solo creators had two options. Lock the camera on a tripod and sit still Or hold it themselves and get shaky footage Now a $300 gimbal tracks your face across an 18-meter room and the footage looks like someone paid for a crew. Not the craft. The invoice. Save this.

@SSSvinosvin bookmark

7 free LLM APIs for testing AI agents. No credit card on any of them. 1) Google Gemini 2.5 Flash Free tier on AI Studio. 1,000,000 token context. 15 requests per minute. Best free option for RAG and long document agents. Nothing else gives you this context window at $0. aistudio.google.com ---------------------------------------------------- 2) DeepSeek R1 671B parameters. Reasoning competes with paid Opus-tier models. Free API with generous limits. If your agent needs to think through multi-step problems, start here. platform.deepseek.com ---------------------------------------------------- 3) Mistral Large 3 1 billion free tokens per month. No credit card. 250K context. Sign up with Gmail, API key in 30 seconds. Faster than most paid alternatives on straightforward tasks. console.mistral.ai ---------------------------------------------------- 4) Groq (My favourite. Just sign in and create api key) Runs Llama 4 and Mixtral at 800+ tokens per second. Fastest free inference anywhere. When you're iterating 50 times a day and latency kills your loop, this is the one. console.groq.com ---------------------------------------------------- 5) OpenRouter One API key. Access to DeepSeek R1, Qwen 3, Llama 4, Gemma 3, all on free tiers. No credit card. Swap models by changing one string. Best option if you want to benchmark your agent across multiple models without managing five accounts. openrouter.ai ---------------------------------------------------- 6) Qwen 3 Alibaba's open model. 0.6B to 235B sizes. Strongest open option for code generation right now. Free API through Alibaba Cloud or run locally via Ollama with zero API dependency. ollama.com ---------------------------------------------------- 7) Cloudflare Workers AI Serverless inference. 10K free tokens per day. Hosts Llama, Mistral, and more. Deploy an agent endpoint with zero GPU setup. Best for shipping a lightweight prototype fast. developers.cloudflare.com/workers-ai ---------------------------------------------------- My picks for agent testing: - > Gemini Flash for context length - > Groq for speed - > DeepSeek R1 for reasoning - > OpenRouter for multi-model benchmarks P.S. I tried grok myself, and honestly, I was blown away I thought everything was paid these days. ofc, to use these models, you have to set them up meaning, essentially, describing how they work in an MD file. you can also run them locally on your computer or host them using your own script. If you're building agents and only testing against one paid model, you're shipping half-blind. Save this so you don't lose it.

@SSSvinosvin Yeah, Groq is really cool I’ve used it too. You just generate an API key and add it to your project. there’s nothing to it.

@SSSvinosvin he sleeps, agents work

A 22-YEAR-OLD IN HOUSTON BUILT A SCI-FI COMMAND CENTER TO RUN HIS ETSY STORE. 11 AI AGENTS. $1,836 IN REVENUE. ZERO EMPLOYEES He calls it UltronOS. He did not hire a single person. He did not outsource a single task. He built a crew. That is his actual operations dashboard. Every green light is a working agent. The crew has names. Ultron processes task assignments. Nova monitors product categories. Forge refines mockup renders. Pixel iterates on brand visuals. Cipher routes dispatches between agents. Atlas compiles competitive intelligence. Ledger runs the treasury. Echo analyzes market signals. Andrew sits at the top. Commander and Founder. He watches. They work. Pause at 0:27. Revenue Command. $1,836.72 total station revenue. $159 today. $938 this week. Factory I runs Etsy. Factory II is not operational yet. 563 items produced. 56 products live. 6 automations active. 194 days of uptime. The station has not gone offline once. The agents talk to each other in Ship Comms. One writes "My models predict we'll hit $2,000 by next week. Should we accelerate production?" Atlas responds. "Not yet. Quality over quantity. Keep the QA pass rate above 90%. Sustainable over fast." He did not script those messages. The agents wrote them. They are debating growth strategy while he holds a lollipop and watches. Pause at 1:50. Full station view. Six operational modules floating in a space-themed UI. Revenue ticker scrolling at the bottom. Customer names. Order amounts. Real money from real people. Pink plant on the desk. The screen looks like something from a Marvel movie. His parents probably think he plays video games for a living. He built it on OpenClaw and Claude. OpenClaw handles orchestration. Claude handles reasoning. Together they do not wait for prompts. They move on their own. $1,836 is not retirement money. But the station runs while he sleeps. 194 days. Zero manual labor. Factory II is next. The internet said AI agents are just chatbots with extra steps. This one runs a storefront, manages quality control, and argues with itself about whether to scale faster. The internet was looking at the wrong screen. Use the article to build your first agent. Save this. The future of AI is not better prompts. It is ops.

A 27-year-old in Cologne built an AI command center that runs 47 agents from his living room. No team. No funding. No framework. Just one laptop and a folder full of markdown files He calls it an AI orchestrator Last month the system handled $14,200 worth of client work while he slept Pause at 0:04 : That is not a video game That is a live agent map Every glowing dot is a separate AI agent. Each one has a name, a role, and a task queue. They are all connected. They are all running. He built the entire nervous system from scratch The secret is one folder. He calls it "Soul" Inside, a single markdown file controls everything. soul.md. Who the orchestrator is, what it manages, how it delegates, when to escalate. Change one line, the entire system shifts behavior. Pause at 0:11 : Look at the bottom of the screen "New agents online" The orchestrator spun up new agents on its own because the workload demanded it. The dashboard shows the numbers: -> 107 tasks completed today -> 97 in queue -> 47 agents active -> 17 running parallel right now His electricity bill for the whole setup is less than his morning coffee habit The agents bill more in one night than most freelancers bill in a week He built it because he saw what the AI space is slowly realizing. The era of unlimited free AI is ending. Costs are going up. Rate limits are getting tighter. The window to learn this is closing. 3 months of evenings. No shortcuts. No tutorials. Just building, breaking, rebuilding. His girlfriend asked why the screen looks like Iron Man He told her it kind of is Except Iron Man does not invoice clients at 3 AM The internet is still debating whether AI agents are real. This guy already has 47 of them clocked in for the night shift. Use the article to build your first agent. Save this. The future of AI is not better prompts. It is ops.

@SSSvinosvin If he could do it, then maybe I can do it too...

@SSSvinosvin so useful, ty for the share

A 21-year-old in Austin built a sci-fi command center to run 4 businesses with 8 AI agents. $28,321 year to date. Zero employees. Last year he did every task himself. Now he does none of them. Pause at 0:12 That is not a video game. That is his actual dashboard. $4,940 this month. $1,609 net profit. 15 jobs completed. He is watching, not working. Two AI agents staring back at you from the bottom of the screen. CORTANA. Chief Operations AI. Active. Monitoring Gmail. 20 tasks today. Last ping 17 minutes ago. JARVIS. Operations and Client XP. Idle. Awaiting orders. 20 tasks today. These are not chatbots. These are employees with job titles and shift logs. 8 agents. Each one does exactly one thing. Cortana runs operations and manages the other 7. Jarvis hunts new clients. Boss guards the content pipeline, checks every caption before it goes live. Aura generates content ideas across every platform. Cash watches the market and delivers picks by morning. Forage builds products on Printify and ships them to Etsy. Titan reads the numbers and decides keep it or kill it. Pack sources sports cards and Pokemon for the TikTok shop. 4 businesses. 1 person. He built it on OpenClaw and Claude. OpenClaw handles autonomous orchestration. Claude handles the reasoning. Together they do not wait for prompts. They move on their own. The screen looks like something from Iron Man. The revenue is real. Most people are still typing one prompt at a time. This guy built a team that runs while he sleeps. Use the article to build your first agent. Save this. The future of AI is not better prompts. It is ops.

@SSSvinosvin no way...

4 DAYS. THAT'S HOW LONG FABLE 5 LASTED BEFORE THE U.S. GOVERNMENT SHUT IT DOWN. commerce Secretary Howard Lutnick issued an export control directive. anthropic had to shut down Fable 5 and Mythos 5 for every user worldwide Immediately - - - amazon researchers found a jailbreak and reported it to the government the technique: asking the model to "read a codebase and fix software flaws" the government classified it as a national security risk anthropic calls it a misunderstanding, says the same vulnerabilities exist in every other model - - - anthropic's response got 38M views in hours. They disagree with the directive. They say perfect jailbreak resistance isn't possible for any provider. all other Claude models still work. - - - Fable 5 and Mythos 5 are offline until further notice. The most powerful model ever released to the public. 80.3 on SWE-bench Pro. Pulled in under a week. The Trump administration tried to stop the launch before it happened. They failed. Now they're using export controls instead.

@SSSvinosvin really cool that you decided to try it, unf can't help with Rust

Day 1. Building in public. waiting for advice I wanted something closer to a kanban board. small widget, always on top, built for how I actually think. so I'm building it myself. today I sketched out the visual layout priority system with color-coded tags. click to cycle inline task expansion with description dates and recurring tasks not a wireframe tool. just a whiteboard and a marker. set up the stack. Tauri 2 + React 19. Tauri is a framework for desktop apps. under the hood it runs Rust + React or uses Rust or another framework of your choice. I've never written a single line of Rust before today. first build crashed instantly. - > time 0.3.48 conflicts with tauri-utils no upstream fix spent a good hour pinning one dependency to get it to compile... - - - - > zero features built - > but it compiles - > that counts the plan: desktop first, then sync with your phone. add a task on your PC, see it on your phone. no account needed if you just want local mode. I also use the Opus 4.6 model for assistance but I believe it should help rather than do everything instead of you especially with things you don't understand anyone here writes Rust daily? how long until it clicks? github: github.com/svinosvin/todo…

@SSSvinosvin I like this explanation, ty

Claude with an Obsidian vault is a system that compounds. Claude without memory is just a chatbot. - - - Most people re-explain their project to Claude every session. Structure, conventions, patterns. Every. Single. Time. CLAUDE.md fixes this. One file. Claude reads it before your first prompt - - - Blogger shows the full Obsidian + Claude Code setup in 17 minutes. Not theory. Working vault. CLAUDE.md as a persistent system prompt across sessions Obsidian vault as long-term memory, not a note dump YAML frontmatter on every note, Claude queries structure, not just text Wikilinks between notes, Claude follows the graph and finds context you forgot Zero re-explaining. Open a session, Claude already knows the project - - - I run a similar stack. 6 AI tools tracked with pricing, status, and cross-references. 6 posts synced with engagement metrics. Custom skill that builds the whole note in 40 seconds. $0 in API cost, subscription only. - - - > 0:37 5 advantages of this setup > 7:03 How memory routing works > 12:37 File structure walkthrough > 16:50 Fast setup guide

@SSSvinosvin II’ve always wanted to try Obsidian but was put off by its complexity I really should give it a try.

@SSSvinosvin other models are no longer competitors of course.

Fable 5 80.3 on SWE-bench Pro GPT-5.5 got 58.6 Gemini scored 0 A Chinese developer built Minecraft , Super Mario and other games in Claude Code with Fable 5. Not a demo. Biomes, underground mines, the works. One prompt session. - - - Fable 5 is Claude Mythos with safety guardrails. The original was deemed too dangerous for public release. On sensitive queries, it silently routes to Opus 4.8. The unfiltered Mythos 5 is restricted to safety research orgs. - - - Chinese dev community figured this out first. They're shipping full games, complex apps, and production tools while most people are still testing it on basic prompts. - - - SWE-bench Pro: 80.3 vs 58.6 (GPT-5.5), 20+ point gap Augment Code tested ~500 real tasks, Fable topped accuracy and speed The harder the task, the wider the gap Free until June 22, then $10/M tokens (2x Opus 4.8) - - - The most powerful model available right now. Also the most expensive. - - - For the same number of tokens, Fable 5 costs twice as much as Opus 4.8 ($10 versus $5 per million tokens). However, for complex tasks, Fable 5 consumes fewer tokens because it solves the problem in fewer iterations (steps) than Opus.

@SSSvinosvin but it still grabs a lot of limits?

Сlaude Fable 5 scored 80.3 on SWE-bench Pro GPT-5.5 got 58.6 Gemini scored 0 Look at the 0:34 Not a marginal lead. A different category Fable 5 is Claude Mythos with safety guardrails. The original was considered too dangerous for public release. On sensitive topics like cybersecurity or bio threats, it silently downgrades to Opus 4.8. The full unfiltered version, Mythos 5, is restricted to safety research orgs only. - - - What you actually get with Fable 5: SWE-bench Pro: 80.3 vs 58.6 for GPT-5.5, gap of 20+ points Augment Code tested ~500 real coding tasks, Fable topped accuracy and speed The longer and harder the task, the wider the gap gets Free on Pro subscription until June 22, then $10/M input tokens (2x Opus 4.8 pricing) - - - I'm using it for reverse engineering binary reconstruction on a Codex project Opus 4.8 couldn't finish it 2:00. GPT-5.5 couldn't finish it. Fable 5 handles it. The most powerful model you can use right now. Also the most expensive.

@DfxoqEth This looks like an opportunity to try myself in game development.

A Chinese dev sat down with the new Claude model... and built a playable game in 8 minutes The internet said AI can't ship real products He wasn’t a studio. No team. No engine open Just a laptop. A late night. And one chat window He typed a single prompt. 19 lines. No code pasted in. Then he hit run. Pause at 0:06. Look at the left panel carefully. Files are appearing. Folders expanding. Functions writing themselves. That’s not autocomplete. That’s an agent building the project live. 33 chunks loaded. 40,000+ triangles rendering. 3D world… from nothing. No red errors. No retries. No debugging spiral. Just output. His mom thinks he’s watching videos. He’s generating playable environments from text. He didn’t “learn game dev”. He learned how to ask. The game isn’t the breakthrough. The interface is. The internet said AI is a toy. The internet was looking at the wrong layer.

@SSSvinosvin I think I'll have to try it

The Truth will set you free.. Join @inference_labs' TruthTensor Crucible to earn points and climb the leaderboard TruthTensor.com

@Zexyooo Hello