I was not sure if I wanted an iPhone Duo, but after experiencing it in real life for just 5 minutes, I’m 100% sure I don’t want it.

Same floor plan. Two 3D models. Which workflow was faster? We gave GPT-6.1 Terra and Claude Opus 5.5 the same dimensioned apartment drawing, then checked their exported meshes against the original CAD file that neither model saw. Both passed all 28 checks: 8 dimensions, 7 room connections and 13 openings. Neither needed a human fix. Small opening errors of 5–10 mm remained, within our preset 100 mm allowance. Terra's successful first run: 54.3 seconds. Opus's first run: 96.9 seconds. The catch: a 44-second Terra launch was blocked by our own tool-approval settings before it saw the drawing. Include it and Terra's initial total is 98.3 seconds, excluding setup-fix time. These are whole-agent workflow times, not pure model speed. One drawing and one run per model make this a case study, not a general ranking. Watch the full comparison: youtube.com/watch?v=BNjaA9K-… What would you count when comparing agent workflows: successful runs, or every attempt? Floor plan: Innerscene, "Three-bedroom apartment, 105 m²", CC0 1.0; rasterized from the source DXF. No endorsement by OpenAI, Anthropic or Innerscene implied. #AIAgents #ComputerVision #GPT61 #ClaudeAI

Nemotron 3.5 Super VL couldn't confirm step one. It still said "agrees." We showed NVIDIA's video model a robot assembling a lamp. It confirmed the bulb and shade, marked the base step as unobserved or uncertain, and wrote that the base never moved. Its verdict still agreed with the checklist. The robot actually completed all three steps. The verdict was right, but the model's own evidence did not support it. Then it counted 3 squats where our reference count is 2. Count them yourself. Ask for evidence. Check that the verdict follows from it. Then check the evidence against the video. Full test: youtube.com/watch?v=BbRxAsp6… Three case studies, not a benchmark. Reference answers were AI-assisted; human review is pending. Footage: FurnitureBench — Heo, Lee, Lee and Lim, CLVR Lab/KAIST (RSS 2023): clvrai.github.io/furniture-b… Squats: "Squat - exercise demonstration video" by FitnessScape, CC BY 3.0, via Wikimedia Commons: commons.wikimedia.org/wiki/F… License: creativecommons.org/licenses… — replay, half-speed playback, cropping and overlays added. Road-sign sample and model: NVIDIA. No endorsement implied. Cover is an illustration. #Nemotron #NVIDIA #ComputerVision #OpenCV

Every automated check passed. The evidence box still missed the bus. We gave Roboflow Agent one minute of drone footage and asked whether an intersection was blocked. The workflow ran, the output parsed, and the structural checks passed—but the box offered as evidence for the blue bus did not contain it. Three checks for your next vision workflow: Does it run? Does it parse? Does the evidence support the answer? Watch the full tutorial: youtu.be/27apodnar2M One feasibility test, not a benchmark or a measure of production accuracy. Footage: Songdo Traffic v2, Robert Fonod, Haechan Cho, Hwasoo Yeo and Nikolas Geroliminis. zenodo.org/records/17924857 — CC BY 4.0: creativecommons.org/licenses…. Resized and compressed; frames extracted and annotated. The dataset authors and Roboflow do not endorse this video. #ComputerVision #Roboflow #AIagents #OpenCV

Same face, different conditions: 0.95 in the studio, 0.78 against a car-video frame. Day, camera and setting changed together—we cannot blame lighting alone. Enroll across days and cameras. Similarity is not accuracy. #OpenCV #ComputerVision Lighting data: Collin Gros (2019), Facial Recognition Dataset VIDEO (part 1 of 2), Mendeley Data V1. doi.org/10.17632/xgg8xcscr5.… — CC BY 4.0 (creativecommons.org/licenses…). Used to derive scores and graphics; no dataset faces shown.

Watch the full Muse Wonder football test here: youtube.com/watch?v=gZHAZtzs…

We tested Meta’s Muse Wonder on football—and checked what it actually saw. On 40 goal/shot clips, its best guess was correct on 26 (65%); our teammate got 34 (85%). A red marker hidden in one frame went unnoticed. Held for one second, it was spotted. Then we asked it to cut highlights from a 92-minute match. All 12 tool calls worked, but only 2 of the 5 goals landed in the selected clips. A useful coordinator needs reliable eyes. This is one controlled case study, not an overall benchmark. Watch on LearnOpenCV: “We Tested Meta's Muse Wonder on Football. It Missed Goals a Person Could See.” youtube.com/@LearnOpenCV Data: SoccerTrack v2 (atomscott/soccertrack-v2, CC BY 4.0). #ComputerVision #MultimodalAI #ModelEvaluation #OpenCV

Datalab’s Indra OCR runs on CPUs, with no GPU. Its recognizer has about one million parameters. Datalab reports 49 pages/sec on a 32-core server; on 4 cores, about 7 pages/sec. Hardware matters when reading benchmark headlines. In our upside-down scan demo, it rotated the page, detected lines, determined reading order and wrote Markdown in 344 ms on 2 vCPUs. Built for clean printed documents; handwriting, faded scans and complex tables remain limitations. What document type should we test against Tesseract and PaddleOCR next? Sources: datalab.to/blog/indra-ocr github.com/datalab-to/indra Demo page: Attention Is All You Need — Vaswani et al. (2017). #OCR #ComputerVision #DocumentAI #OpenCV #Python

Would I let a video model sign off on completed work? Only after checking how it reached the verdict. Our Nemotron 3.5 Super VL test is a useful example: the robot completed the checklist, but the model could not support one step and still agreed. A correct verdict alone would hide that failure. We also tested a road sign and squat counting. Three case studies—not a benchmark. Human review of references is pending. Full video, “Nemotron 3.5 Super VL Tested: A Correct Answer for the Wrong Reason,” on LearnOpenCV: youtube.com/watch?v=BbRxAsp6… Video sources: NVIDIA; FurnitureBench (CLVR Lab/KAIST); FitnessScape (CC BY 3.0). Full credits in the description. Thumbnail is an illustration. What would you check beyond the final answer? #AI #ComputerVision

What interests me about Dream-1 isn't just a new robot demo. It's the possibility of needing less training data every time the hardware changes. Dream Labs trained it on the YAM and showed zero-shot transfer to 27 unseen robots in simulation, plus real arms. I'd want to know where that transfer breaks: unfamiliar workspaces, poor IK, and new gripper designs. Today the scope is bimanual robots with parallel grippers. Selected, sped-up demos aren't a reliability test. I haven't run Dream-1 myself. Footage: Dream Labs (3× simulation, 4× real robots). Research: dreamlabs.ai/research/dream-… AI voice and avatar. What would you test first on your own robot? #Robotics #AI

A crowd is gathering at Omarchy installation party in Shenzhen. Kudos to @dhh for creating this excitement around open source.

A new open-source OCR model turned our chart into a clean table. One column was wrong in all four runs, about half a point low. It read the line against the wrong axis. The table looked fine. OpenCV caught it.

Meta's new Muse Wonder can take in video and call your tools. Its own model card: "Visual and spatial reasoning, especially 3D tasks, remains a known limitation." Meta's reported benchmarks: agents and coding. No video. Use it as the planner, then test it on clips where you know the answers.

EmbeddingGemma 2 Makes Your Camera Roll Searchable, But It Can’t Tell Time’s Direction Google’s new EmbeddingGemma 2 finds a clip among 40 videos from one sentence (13 of 14 searches hit #1). But play the same frames backwards and it barely notices: a video and its reverse are 99%+ similar. It sees what’s in a scene beautifully, and barely notices when. See exactly where EmbeddingGemma 2 shines and where it fails, in the full video: xitter.cf/LearnOpenCV/status/210… blog.google/innovation-and-a…

A 0.5B Model Turned a Room Scan Into a Structured Layout in 16 Seconds SpatialLM turned a bedroom video into structured text in 16s on a laptop GPU, with a half-billion-parameter LLM. Point encoder → adapter → LLM, one record per object. Bed size within 5 cm, 5 of 6 walls and the window line up. A few misses too. One room, one run. github.com/manycore-research…

10-yr-old showed me the snake he saw at the beach. It was only 6 inches long. Scale ambiguity fools humans too.

OpenAI's new Decisions API: one fast, typed answer per request. → probability a condition is true → pick from your list → score against your levels Text + images, GPT-6 Luna, $0.10/M input tokens, no output charges. "Up to 10x faster" is OpenAI's figure. Test on labelled data first.