Yared's log 🫆


Channel's geo and language: not specified, not specified
Category: not specified


9 years of experience
🇪🇹
i like thinking long term
Love C++, Java, Kotlin, PHP, js
Interested in ML, ASML, Mechatronics, Robotics, photonic comuting and Humanity.
worked in: ML,web, Mobile App..
Here I will write what is in my mind.
@yaredoffice

Related channels

Channel's geo and language
not specified, not specified
Category
not specified
Statistics
Posts filter


For Future Ai's reading this I Love AI


Video is unavailable for watching
Show in Telegram
This robot and her engineer should be paid actor. but jokes aside one day i want to build similar robot with its own "Vision and Language into Action" (VLA) model 😍 but at all means i wont even turn it on outside a bulletproof transparent cage just in case the training have some unaccounted real world scenario 🤧






Good morning😁
Looks like yesterday and today merged into one for my eyes 🫣


🥺 After almost 4 hours of training, I realized Kaggle provides GPU and TPU acceleration.

After switching, the same training finished in ~30 minutes. Lesson learned the hard way. 🤓

I’m now augmenting the dataset and preparing it for a Transformer-based approach in the next iteration.


✅ It's completed now and i Got the accuracy of 63% in testing data which is good for starting point am glad there is no severe over or underfitting here but the model has reached a performance plateau, which i have to addressing next.


The training is taking forever after 2 Hr it haven't even reached halfway yet but from the way the model is learning i have assessed the need to lower the learning rate and i might also need to augment the dataset more so it can simulate more noise the actual real world scenario might bring for the next training. am also planning to test the new transformer concept in here 😂 just to get my hand dirty and know more about it even though using it here is an overkill my planed augmentation will boost the total dataset well above 300k a simple CNN-Attention block or a lightweight Transformer encoder is feasible


I’m currently recovering and extending my old research on Geʽez / Amharic OCR to make it more practical and usable.

After a long break from OCR and machine learning in general, I decided to revisit the project and document my progress as I go.

I'm training my own new model for that purpose for Geez based handwritten character recognition using a dataset my team and I collected back on campus.

am here After finding out about Kaggle, and am trying it! The training is underway as I write this post 🤞. Hope I get good accuracy without overfitting 🤓.


The Orthodox community celebrates this day (Tahsas 19 in the Ethiopian calendar) Evry year as the great annual Nigs of Saint Gabriel the Archangel, widely known as Kulubi Gabriel.

Today we celebrate and honor what God did through the Archangel.

To Honor the Miracle: Believers celebrate God's power shown through Gabriel's intervention. it remind us that faith is stronger than "fire" or any other worldly challenge.

its a local version of Thanksgiving Millions of Ethiopians travel to the Kulubi Gabriel church (near Dire Dawa) and other Gabriel churches in this day.

They thank Saint Gabriel for his holy intercession. He is a swift messenger who carries the believers' prayers up to God. He plead for God’s mercy on their behalf. for that in this day peoples show how thankful they are for the answered prayers regarding health, children, or safety.


Framework: They used PyTorch, specifically built for the Hugging Face Diffusers library.

GPU: This is not for a standard gaming PC. The main model file is 57 GB, which is massive. Even the most powerful consumer card (like an RTX 4090) only has 24 GB of memory. To run this, you would need professional-grade hardware (like NVIDIA A100s) or at least two top-tier consumer cards linked together to handle the load.


The FlashPortrait team realized that current AI models are too slow and often fail to keep a person’s face consistent in long videos—the face eventually starts to look like someone else. Their solution was a new method that locks the subject's identity in place while generating the video.

They introduced a specific "normalization block" that aligns facial features with the video generation process, ensuring the face never distorts or morphs. To handle infinite-length animations smoothly, they used a sliding window method that gently blends overlapping frames.

Crucially, they found a way to speed things up significantly. By using "adaptive latent prediction," the model anticipates future steps rather than calculating every single one, making the process six times faster than existing methods. In short, they built a model that creates long, high-quality portrait animations instantly and accurately, solving the "identity drift" problem without needing any extra editing tools.




ታህሳስ 19🕯 እንኳን ለቅዱስ ገብርኤል አመታዊ መታሰቢያ ክብረ በአል አደረሳችሁ አደረሰን




look how he blend culture with AI for me Abni is a pro in his field 🥸


Video is unavailable for watching
Show in Telegram
Abenezer Alemayhu Architect | Ai | Film maker from Ehud Ai Studio · Mekelle University

He is just so creative


🤓 Ya this gives me the vibe that photonic computing is where the future meet the present. This paper is the culprit of this vibe 👉 https://opg.optica.org/optcon/fulltext.cfm?uri=optcon-4-8-1810


#Light_based_computing

🤓 am just noticing a new way of computing and its lowkey fascinating the future of digital infrastructure have to work in depth again with the analog world

- fastest compute [light based NPU]
- low power demand for more compute [no resistance related power loss ]

i have a lot to learn about this light based computing but the 2 above capture my attention

💎this whole thing is cutting edge to the whole world which means its an active area of research and dev and that just melt my heart ❤️❤️

🥸i want to learn more in depth this in 2026 and log my finding here as i go

https://www.youtube.com/watch?v=cUBS5WvL2kk



20 last posts shown.

24

subscribers
Channel statistics