Satya Mallick retweeted
The fact that two Huangs are involved in an affair, but only one takes the blame is commonly known as Jensen's Inequality.
25
163
2,115
56,581
AI & Computer Vision News (Nov 2–6): Demo vs Real Data
Seven stories, from Sekai's event memory to annotator-free object detection. Authors report 60.8 AP on COCO and 47.4 on LVIS, zero-shot, without human labels on target data. We haven't reproduced these results; the founn supervision.
Sources and footage: Sekai (arXiv 2611.06685); huggingface/lerobot and allenai/toyblocks (Apache-2.0); hanyucc/trainspotting (2611.06849); OpenBMB/MiniCPM-OCR (Apache-2.0); VLALagBench (2611.03103); Annotator-Free Object Detection (2611.06195); OpenAI and Google. Office and factory scenes are illustrations made with Gemini Omni 1.2 Flash. No endorsement implied.
#ComputerVision #AINews
Seven stories, from Sekai's event memory to annotator-free object detection. Authors report 60.8 AP on COCO and 47.4 on LVIS, zero-shot, without human labels on target data. We haven't reproduced these results; the founn supervision.
Sources and footage: Sekai (arXiv 2611.06685); huggingface/lerobot and allenai/toyblocks (Apache-2.0); hanyucc/trainspotting (2611.06849); OpenBMB/MiniCPM-OCR (Apache-2.0); VLALagBench (2611.03103); Annotator-Free Object Detection (2611.06195); OpenAI and Google. Office and factory scenes are illustrations made with Gemini Omni 1.2 Flash. No endorsement implied.
#ComputerVision #AINews
2
2
2
408
A lot of algorithms, especially in geometry are difficult to code and very delicate.
Now, thanks to frontier models we can try those ideas out.
Not everything needs to be solved using a big hammer of neural networks.
Now, thanks to frontier models we can try those ideas out.
Not everything needs to be solved using a big hammer of neural networks.
The dogma of the last decade in computer vision is that all classical algorithms should be replaced by neural networks. This has been very productive but, counterintuitively, with the latest frontier models, I find myself sometimes turning back to classical methods.
If you have the training data, the dogma still holds. But there are situations where data is hard to get, there exist good classical algorithms for solving a problem, optimization methods are just better, or where some simple code just does the thing you want.
When writing code was time consuming and error prone, training everything made sense. Now, that a frontier model can code up and test solutions quickly, getting a solution through coding can be faster than gathering training data and training a model. Consequently, I think there will be more hybrid approaches that blend trained models and code.
We now have generic, high-level, planners that can execute sequences of lower-level tasks to solve large problems. These lower-level tasks may be solved by a neural network, a classical method, or specialized code written on the fly.
It never pays to be religious about technology. The future is probably hybrid. Pick the thing that works for you.
1
0
5
546
We tested Pangram: "Unsure", yet the heatmap found the AI half. Small demo, not a benchmark.
Chess: Bubba73 (Jud McCranie), colour adj. Hohum. Cropped, Qwen-edited, composited, annotated. Video: CC BY-SA 3.0.
commons.wikimedia.org/wiki/File:Ches…
creativecommons.org/licenses/by-sa…
Chess: Bubba73 (Jud McCranie), colour adj. Hohum. Cropped, Qwen-edited, composited, annotated. Video: CC BY-SA 3.0.
commons.wikimedia.org/wiki/File:Ches…
creativecommons.org/licenses/by-sa…
2
0
3
349
I asked Qwen Image 2.1 Turbo to move a knight from g1 to f3. Count the white knights.
Chess: Bubba73 (Jud McCranie), colour adj. Hohum. Adapted, CC BY-SA 3.0 creativecommons.org/licenses/by-sa… commons.wikimedia.org/wiki/File:Ches…
Clock: Envato Elements. Music: CorporationMusic.
Chess: Bubba73 (Jud McCranie), colour adj. Hohum. Adapted, CC BY-SA 3.0 creativecommons.org/licenses/by-sa… commons.wikimedia.org/wiki/File:Ches…
Clock: Envato Elements. Music: CorporationMusic.
1
0
3
335
Our Claude maze agent reached 9.8 MB per request by move 100. Using message threads, that dropped to 104 KB: just the new screenshot and the previous response ID. Token use stays the same, and threads expire after 5 minutes. We measured upload size, not speed.
0
1
3
429
Satya Mallick retweeted
Plot twist: I was Vapnik's Department Head between 1996 and 2002, and made sure he had the space, resources, and collaborators to do his best work.
To people who think I'm dogmatic: I famously disagreed with Vapnik about the best way forward in ML, but did what I could to promote his work, because research is about exploration and science is about results.
To people who think I'm dogmatic: I famously disagreed with Vapnik about the best way forward in ML, but did what I could to promote his work, because research is about exploration and science is about results.
18
30
982
74,880
A silent CV bug: swap RGB and BGR, and YOLO11s recall fell from 67.9% to 57.4% on COCO128. No crash.
Google says Gemini 4 Carbon can tackle long debugging tasks; we haven’t tested it. My checklist for any fix: reproduce, inspect the diff, retest on held-out data.
Google says Gemini 4 Carbon can tackle long debugging tasks; we haven’t tested it. My checklist for any fix: reproduce, inspect the diff, retest on held-out data.
0
0
4
288
Why Fewer CPU Threads Made YOLO 2x Faster (ONNX Runtime)
In Ultralytics’ benchmark, limiting PyTorch to two threads cut the time to process 1,000 images from 58.8 s to 27.1 s on an 8-vCPU machine. Same model, no retraining.
PyTorch handles preprocessing and postprocessing while O runs the model. Idle PyTorch workers can keep spinning on the cores the model needs. Ultralytics 8.4.193 caps those PyTorch threads at two for this CPU inference path.
Hardware matters: on 4 vCPUs the change made almost no difference, and one batch test was 5% slower. On Apple M5 Pro, the difference was within noise. These are the maintainers’ results, not our own benchmark.
Update, then measure the full image-to-result path on your deployment hardware. Two threads is not a universal setting.
Source: github.com/ultralytics/ul…
Images: COCO val2017 (cocodataset.org). No endorsement implied.
#ComputerVision #YOLO #ONNXRuntime #MLOps
In Ultralytics’ benchmark, limiting PyTorch to two threads cut the time to process 1,000 images from 58.8 s to 27.1 s on an 8-vCPU machine. Same model, no retraining.
PyTorch handles preprocessing and postprocessing while O runs the model. Idle PyTorch workers can keep spinning on the cores the model needs. Ultralytics 8.4.193 caps those PyTorch threads at two for this CPU inference path.
Hardware matters: on 4 vCPUs the change made almost no difference, and one batch test was 5% slower. On Apple M5 Pro, the difference was within noise. These are the maintainers’ results, not our own benchmark.
Update, then measure the full image-to-result path on your deployment hardware. Two threads is not a universal setting.
Source: github.com/ultralytics/ul…
Images: COCO val2017 (cocodataset.org). No endorsement implied.
#ComputerVision #YOLO #ONNXRuntime #MLOps
3
1
3
344
Satya Mallick retweeted
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
7,397
2,703
39,177
7,527,129
PyTorch 2.14.1 vs 2.15.0, same YOLO11n, same RTX 4060 laptop.
Speed: 12.15 vs 11.93 ms/image, runs overlap. No proven speedup. Detections: identical on 128 images. Peak tensor memory: 66 → 53 MiB (about 19% lower). Reserved stayed at 88.
Benchmark before you upgrade.
Speed: 12.15 vs 11.93 ms/image, runs overlap. No proven speedup. Detections: identical on 128 images. Peak tensor memory: 66 → 53 MiB (about 19% lower). Reserved stayed at 88.
Benchmark before you upgrade.
0
0
2
309
Can Meta's Muse Wonder spot a goal? Our case study: 26/40 goal and shot clips matched the dataset, versus our teammate's 34/40. All 12 tool calls worked, but selected match clips included 2/5 goals. Pair it with a tracker and human review.
Scope: Muse Wonder through OpenRouter, nue, one full-match run, no audio. Not a benchmark.
Footage: SoccerTrack v2 (atomscott/soccertrack-v2 on Hugging Face), CC BY 4.0. No endorsement by Meta, OpenRouter or the University of Tsukuba implied.
#ComputerVision #OpenCV
Scope: Muse Wonder through OpenRouter, nue, one full-match run, no audio. Not a benchmark.
Footage: SoccerTrack v2 (atomscott/soccertrack-v2 on Hugging Face), CC BY 4.0. No endorsement by Meta, OpenRouter or the University of Tsukuba implied.
#ComputerVision #OpenCV
0
0
0
248