This is the power of YOLO, trained on a laptop for ~1 hour, with a Kaggle dataset. Oh, and just ~100 lines of Python. I can make a startup on this and it took me literally a couple of hours.

Computer vision is wildly underrated! I've been diving deeper into CV for dental issues and trained a YOLO model in Python on a labeled dental X-ray dataset with caries annotations on Kaggle. A well-trained segmentation/detection setup achieves solid accuracy in caries

Go ad-free on X with Premium+ Includes access to SuperGrok.

We can do better than this but for a quick try on Colab is enough.

Trained a YOLO instance segmentation model from scratch on a custom dental dataset using my yolo-training-example. - Automates tooth numbering & segmentation - Faster diagnostics - Early detection of issues (bad placement, caries) - Reduces manual errors & time

My feed is full of these beautiful reachy mini and I like it😆

Let me write a different post, no tech stuff here just to say thank you. I have almost reached 10k followers and that's pretty crazy. Thank you all for the support, the likes, the retweets, and the great discussions. Really happy to see so many people interested in computer

Our new node-based editor for physical simulations is becoming a reality. Built on PhiFlow, this visual engine lets you compose complex physics simulations as graphs with no boilerplate. You can simply: - drag nodes, - connect fields, - iterate, and visualize results in seconds.

Still playing with Roboflow to train a model that recognizes a kettlebell during the exercise. I really like do exercise with the kettlebell because is a strong full body training. The main point is keeping the right trajectory. Having an application that keeps track of the

I think we can do better than this. We should squeeze every single byte out of these GPUs.

See you in six hours

Data centers in space is becoming a thing. Starcloud training the first AI model in orbit (using NVIDIA H100s) and big players like SpaceX, Google, & AWS pushing orbital computing for unlimited solar power. I built a Threejs 3D scene visualizing a futuristic space data center

MicroCAD in the browser is becoming real. I'm building a web-based μCAD editor where code is the model: - Write parametric geometry, compile, and instantly preview. - Export STL with no dependencies. - No heavyweight CAD. - An LLM chat to generate μCAD from text (still

New job position achieved: Context Switch Engineer

This single aerial image contains hundreds of detections, each one localized and classified in real time. Containers, vehicles, buildings, and port infrastructure are all recognized in one forward pass of the model. Why it matters: - Scalable monitoring for logistics - Real-time

What if programming in Python was like building with Lego? Instead of writing lines of code, you snap together blocks: - a function is a brick - inputs are studs - outputs are connectors You don’t fight syntax. You build systems.

Sometimes, things break when you're just doing things. The pothole detector hardware is still being developed, but I had to wait for a new camera. https://t.co/NZq5dGs3YE

IT IS WORKING!

This is enough to start your business and it costs like 10 cents a day.

This project is amazing. You can explore various robotics projects, like on GitHub. You can find all the required hardware for reproducing the projects, along with the firmware and software. Robotics is finally open source, and everyone can now join this beautiful discipline.

Most people fine-tuning 70B models are doing it wrong. If you’re teaching new behavior (code, DSLs, internal tools), small models win. A 3B code model + QLoRA on a couple of GPUs can learn things bigger models struggle with. Why small models matter: - faster iteration - cheaper

Just discovered schemdraw, a Python library that turns code into stunning electrical schematics! - No more dragging wires in clunky tools. - Pure code for resistors, logic gates, and more. - Fully customizable objects. pip install schemdraw and start drawing