@LearnOpenCV
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
CEO, https://nitter.cf/t.co/CzUdJlxzJM. CEO, https://nitter.cf/t.co/et3i4a7oGF Entrepreneur. Ph.D. ( Computer Vision & Machine Learning ). Author: https://nitter.cf/t.co/olraDEG5Ue
San Diego, CA
Joined June 2008
- Tweets3.9K
- Following810
- Followers14.7K
- Likes4.6K
My limit just reset. I have 3 banked resets left.
So I am running a task that is consuming about 1% of weekly limit every 5-7 minutes. It's a business task that uses complex reasoning and a lot of computer use.
Set a Goal with GPT 6 Astra + Ultra + Fast.
The AI Model That Returns a Decision, Not a Paragraph
TypeSafe's Jev returns a typed decision (a choice, a score, or a probability) instead of prose, so your code can act on it directly. Useful for routing a vision pipeline: review, read the label, or reshoot. But first ask if a fixed threshold would do, and test decisions against known outcomes.
What is the optimal strategy to use a banked Codex reset?
Wait for a normal weekly reset. Exhaust your limit as soon as you possible (say in half a day), and then use the banked reset.
You're welcome!
The Service You May No Longer Need: VLMs Inside OpenCV 5
The interesting part of OpenCV 5 running VLMs is not the model. It is that a supported VLM can run inside the same app as your image processing. No separate Python service, network hop, or second deployment. At the edge, that is real simplification. At cloud scale, keep your serving stack.
opencv.org/opencv-5/
OpenAI DevDay 4 days away. I have 3 banked resets.
If they come up with a $500/mo version, I will subscribe and the banked resets will convert to the higher tier.
Never in my life had I imagined that AI agents will rule my life in 2026.
Sigh!
Is OpenCV 5's New DNN Engine Faster, and Does It Work on GPU?
The biggest win is compatibility, not raw speed: the new engine loads modern ONNX graphs, dynamic shapes, and transformer models the old one rejected. GPU still works via the classic backends or ONNX Runtime providers. But loading is not running on GPU, so check your engine and backend, and benchmark on your own hardware.
opencv.org/opencv-5/
Testing OCEC and BlinkCNN in one OpenCV app. DEIMv2 Atto finds the eyes; either classifier estimates open/closed. Both run close to 30 FPS in this Mac webcam demo. Watch the eye crops and closure scores change as I switch models.
cc @PINTO03091
Skills vs MCP vs RAG vs Memory: What Each One Actually Does
Four questions, four concepts. How should it work? Skills (the playbook). What can it connect to? MCP (the tool/data standard). What knowledge does it need now? RAG (retrieval into the answer). What should it retain? Memory (kept for later). They overlap, but that is the core split.
Satya Mallick retweeted
Is GPT-6 Luna worse "vision" model than GPT-5.6 Luna?
Yep.
- extraction (high): ↓ 81.79% → 66.67%
- counting (high): ↓ 70.72% → 64.41%
- reasoning (high): ↓ 65.56% → 60.71%
- detection (high): ↑ 62.29% → 64.12%
link: playground.roboflow.com/shar…
↓ more examples
Satya Mallick retweeted
It's mind bending walking around the same place simultaneously in two timelines
Inside the frame it's winter, and outside the frame it's summer - until you step inside.
Recorded this footage in the winter and localized it against a massive 3d scan of the same ranch in the summer. You can follow the camera from the outside, then step into the moment I recorded.
We spend so much time looking at memories through 2d rectangles in our camera roll. It's a strange feeling to walk around one instead. Like navigating through space instead of scrolling through time.
The geospatial memory palace continues to come to life!
Trusting AI detection tools is foolish.
I'm finishing a paper I wrote with my colleagues. I used LLM to help with the abstract, and to improve a few sentences (maybe 5-10 sentences in a 73 pages long paper) in order to increase readability. Pangram says the paper is 63% AI, and this is total BS. I have checked a poem written by Victor Hugo, and it flags as 100% AI! Serious question: what will happen if people/journals/entities start to use these AI detection tools to evaluate our work? Will people be wrongly accused of using AI to their papers? This needs an urgent discussion among academics!
Tell me if you would watch this documentary, and I'll give you a lower bound on your age.
I am watching.
Migrating YOLOv3 from Darknet to ONNX for OpenCV 5
OpenCV 5 removed the Darknet importer, so it can no longer load .cfg and .weights directly. Your weights are fine, the route is ONNX. But conversion is step one: check class ordering, preprocessing (size, normalization, RGB vs BGR), and output decoding. A successful load is not equivalent behavior.
opencv.org/opencv-5/
Why the Ball Box Disappears: YOLO11n vs YOLO26x
Same basketball clip, 545 frames. YOLO11n found the ball in 36% of frames, YOLO26x in ~66%. But that is coverage, not accuracy. The box drops because the ball is small, blurred, or hidden, and a missed ball and a hidden ball look identical in the output. Bigger helped, but size is not enough.