puran thapa

Dev by profession 💻 | AI Hype Tester 🧪 | Built 2 great apps | Love JARVIS 🤖 | Building products

[puran thapa @Iampuranthapa]

DeepSeek V4.1 Flash 1017 vs Opus 5.5 - same job, same prompt.
Make a full video song in just 35 minutes..
Deepseek burned 22.5 MILLION tokens, $0.20(Off-peak).

The song was created using Google AI Studio, while ChatGPT wrote the script and provided the instructions.

[puran thapa @Iampuranthapa]

@simplifyinAI Instead of using Voicebox, use OmniVoice directly. The 646 languages you're seeing are provided by OmniVoice, not Voicebox. You'll save disk space as well, since Voicebox is just a UI wrapper.

[puran thapa @Iampuranthapa]

If you use it with Opus 5.5 medium effort or Sonnet 5.5 high effort, it will work perfectly. Timeline trimming and clipping. Light adjustment is working perfectly. Bounce: use api[.]openverse[.]org for 5.2 million free audio files in your Claude. It will add audio according to your video.

[puran thapa @Iampuranthapa]

@VikParuchuri I tested it and it is really good. I checked a 100-page PDF, scanned image PDF, and distorted words PDF, but it works perfectly. And it is blazing fast. You nailed it. Without any GPU, 100 pages in seconds 🔥.

[puran thapa @Iampuranthapa]

Your ChatGPT or Claude Opus might know you better than almost anyone else.

Here are 4 powerful questions to ask your AI that could change how you think, grow, and improve your life. 🧵👇

Question 1. Alright, based on all our conversations so far, tell me 5 things about me that I never directly mentioned, but you picked up on from my patterns. And don’t sugarcoat it just to make me feel good.

Check its reply—it will blow your mind. Now, the next question.

Question 2. Do you notice any recurring patterns in the way I think and make decisions? And tell me that one thing that’s probably staring me right in the face, something I see, but refuse to admit to myself.

What it tells you will surprise you. Don't stop here—now ask the next question.

Question 3. One more question: if my current habits and patterns stay exactly the same, what do the next five years of my professional and personal life actually look like? Then run the scenario again: if I finally break through the one thing holding me back, how do those same five years turn out?

It will predict your future, and in some ways, it might be right. Now, the last question could change your life.

Question 4. Based on everything you know about me, write a short daily morning paragraph I can read to stay grounded. It should keep me laser-focused, stop me from chasing shiny, unnecessary opportunities, and drill into my head that finishing what truly matters is the only thing that counts.

Copy its answer, read it every day, and follow it as closely as possible.

[puran thapa @Iampuranthapa]

Some seriously impressive open-source reverse-engineering and clean-room projects floating around GitHub right now:

• Wispr Flow Teardown: Deep dive into how Wispr captures audio, transcribes, and injects text at the OS level: https://t.co/s1dMmwjuEj

• ArtCraft: Clean-room open-source alternatives to Adobe's creative suite: https://t.co/n4NSrKu1e4

• AnyPS5: Porting PS5 executables to run natively on Windows & Linux: https://t.co/ujVwYdlsme

• xtool: Cross-platform Xcode alternative to build iOS apps on Linux and Windows: https://t.co/NHMCXERE4N

Bookmarking these for weekend code reading.

[puran thapa @Iampuranthapa]

Working on an open-source project. Arriving soon!

Organize your bookmarks like a pro. It works just like its name.

[puran thapa @Iampuranthapa]

Five GitHub repos that will make your day. Bookmark them before you miss out!

1. Create short and long videos and upload them to your social media:
https://t.co/JYcTuWGSsL
2. An open-source alternative to CapCut that supports MCP, so you can connect it with your AI agent and automate editing:
https://t.co/Y2wPF5IIqs
3. Create polished demo videos in minutes:
https://t.co/27ugDItbcX
4. Make a promo video for your product:
https://t.co/6SHFSk5mXi
5. Motion design skill for Claude:
https://t.co/w0GJf75Mal

If your job involves editing videos and making money from them, these tools can help you automate a lot of your work.

[puran thapa @Iampuranthapa]

@nahid_pro09 I just checked it out, and it’s not working perfectly, but it’s better than ChatGPT. It also makes grammatical mistakes.

Here are the results I got from my own testing.

[puran thapa @Iampuranthapa]

https://t.co/l19aniUvMv

[puran thapa @Iampuranthapa]

X has just started a new program to help creators grow faster on the platform! 🚀

And the bonus payouts are $500, $1,500, $3,000, and $5,000! 💰

Read till the end—I'll share how you can increase your chances of winning this. 👇

Video Kickstart Bonus 🎥
All you need to do is post 15 videos on 15 separate days within 30 days, and that's it!
You can get a bonus payout plus X Premium+ for a full year. X may also boost your account once you join the program and start posting videos. From what I’ve seen, the bonus amount varies by account, so everyone might see a different offer.

How to participate:

- You need an invitation from X since this program is only open to selected users. If you're eligible, you'll spot Video Kickstart Bonus under the Programs section in your Creator Studio.
- When you click on it, you'll see the exact bonus amount X is offering you.
- Hit the Enroll button, and you're in! Just make sure to enroll before October 30.

Rules:
- Post 15 videos on 15 separate days within the 30-day campaign period.
- Videos must be under 3 minutes and in vertical format (9:16).
- Only original videos are eligible. The content has to be yours, though you can still post your own videos on X even if you've already shared them on other platforms—as long as they meet the program's requirements.

How can you increase your chances? 👀
Lately, Elon Musk and the X team seem really interested in AI-generated videos, especially those made with Grok Imagine.

So, creating videos using Grok Imagine or Grok might give you an edge. Based on recent activity on X, Grok Imagine and Grok Bot-related content are getting a lot of attention right now, and several creators are seeing much faster growth from it.

Are you eligible for this program? Check your Creator Studio and let me know! 👀👇

[puran thapa @Iampuranthapa]

@Nandakishorm1 I’m a small developer, and my hobby is testing small size AI models. I’ve checked the Laya, and I can say it’s really awesome!

I’ll try Vega soon as well.
Is Vega available in MLX, like Laya?

[puran thapa @Iampuranthapa]

🚀 Indra OCR: A powerful OCR AI model that runs locally right on your CPU!

Key features:

- Only 1M parameters
- Just 18 MB, runs on ONNX Runtime (no PyTorch required)
- Reads up to 49 pages per second on a CPU
- ~230 ms per page on 4 x86 vCPUs
- Scores 41.3 on OmniParseBench Text
- Supports multiple languages

My testing experience:I tested it locally on my MacBook Pro M3, and it scanned a 100-page PDF in under 2 seconds! 🤯
I also threw scanned image-based PDFs and distorted text at it, and it extracted the text perfectly.

It outputs in plain text, Markdown, lines, and JSON formats. You can easily use the JSON output to prep text for vector embeddings and store them straight in your vector database.

How to install and run it locally:

1. Clone the repository
2. Create a virtual environment using Python 3.13: python3.13 -m venv .venv
3. Activate the virtual environment: source .venv/bin/activate
4. From the cloned repository folder, install the package
5. Launch the Chrome viewer using cammand: indra-ocr-viewer
6. Open the web viewer: http://localhost:8765

Small OCR models running locally can be surprisingly powerful! 🔥

[puran thapa @Iampuranthapa]

You can check out Humanizer yourself!
https://t.co/PCY3VYsw9Y

[puran thapa @Iampuranthapa]

Can AI write like a human and get 0% AI detection? 👀

I just ran a quick test comparing MLX-Humanizer 12B (Q4), Gemini Argon 4, and ChatGPT using the same ~200-word ChatGPT-generated story.

The scores on the QuillBot AI Detector were pretty surprising! 👇

🤡 ChatGPT: 92% AI detection
🥱 MLX-Humanizer 12B (Q4): 50-70% AI detection, plus a few grammar mistakes
☠️ Gemini Argon 4: 0% AI detection, with zero grammatical errors

I ran MLX-Humanizer locally on my Mac, and honestly, it didn't perform as well as I’d hoped. I'm wondering if the Q8 or BF16 quantizations do a better job, since I was only able to test the Q4 version.

Has anyone here tried the Q8 or BF16 versions of Humanizer 12B? How did they hold up for you? 👇

[puran thapa @Iampuranthapa]

Honestly, it’s super impressive and covers 83 languages, but the hardware requirements are brutal.
You’ll need at least 48GB of VRAM
64GB of RAM
45GB of free storage just to run it.
If your rig can handle that, definitely give it a shot. Otherwise, you're much better off just sticking with ElevenLabs.

[puran thapa @Iampuranthapa]

I’m creating an educational app that fetches news every 3 hours. Can @bot automate this entire process for me?

I’ve created a team of agents:

1. One fetches the news.
2. The second fact-checks it.
3. The third summarizes and rewrites the articles into short content.
4. The fourth creates multiple-choice questions and answers.

Finally, everything is saved to my server, making the entire process fully automated.

[puran thapa @Iampuranthapa]

@plotarmordev @ElevenLabs @FishAudio It’s really good! It supports 83 languages, but it requires a minimum of 48GB GPU VRAM, 64GB RAM, and around 45GB of disk space.
So, if you have a system with these specs, go for it. Otherwise, ElevenLabs is a better choice. 😅

[puran thapa @Iampuranthapa]

@kushblowblack @RoundtableSpace As you can see, it supports MCP. I highlighted that part in the repo URL.

[puran thapa @Iampuranthapa]

@EmiliOlivi_ai Yes, there’s a complete video guide in the repo explaining how to create your own. If you face any problems, leave a comment, and I’ll create a demo video showing you how to do it.

[puran thapa @Iampuranthapa]

Claude is crazy!

I just created a song with Claude, for the animated music video, Google AI Studio for music generation and ChatGPT for the lyrics and instructions.

3 AI Tools. One song. Full AI generated music video. 🎵🤖

Want to make your own? Full guide in the GitHub repo. Check the first comment! 👇