A correct answer can still be the wrong reason to trust an AI. We tested NVIDIA Nemotron 3.5 Super VL on a road sign, a robot assembling a lamp, and a squat-counting clip. In the robot test, it could not confirm one step, yet agreed with the checklist. The twist: the robot actually completed the steps. The gap was in the model’s evidence. A consistency check can flag a verdict that does not follow from its observations—but cannot catch every confidently wrong observation. These are three case studies, not a benchmark. Reference answers were AI-assisted; human review is pending. Watch “Nemotron 3.5 Super VL Tested: A Correct Answer for the Wrong Reason” on LearnOpenCV’s YouTube channel: youtube.com/channel/UCeoXXZ2… Video sources: NVIDIA (model and road-sign sample); FurnitureBench, CLVR Lab/KAIST (robot footage); FitnessScape, CC BY 3.0 (squat footage). Full source and license credits in the video description. Thumbnail is an illustration. What evidence would you require before letting a video model sign off on a task? #NVIDIA #Nemotron #ComputerVision #AI

A robotics concept worth learning: end-effector targets. Dream-1 can specify where each gripper should go while the robot's own inverse-kinematics solver handles the joints. That shared interface helps transfer a policy between robot bodies. Dream Labs also trains video and action prediction together. Scope matters: two arms, parallel grippers, and an IK solver that can reach the targets. These demos don't establish reliability. Watch our explainer. Footage and research: Dream Labs — dreamlabs.ai/research/dream-… (simulation 3×; real hardware 4×). We haven't tested the model. AI voice and avatar. #Robotics #ComputerVision #RobotLearning

Google says Gemini 4 Argon watched one video of a person playing cornhole. Then Apptronik's Apollo robot sank 7 shots in a row. The pipeline: video → map the scene in 3D → plan whole-body joints → write sim code → real robot. Google calls it Video-to-Sim-to-Real.

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…

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.

YOLO looked at a cricket match and scored the helmet higher than the ball. Real ball: "sports ball" 0.612 Helmet: "sports ball" 0.723 Zero errors in the run. Still wrong. Confidence ≠ correctness. 🏏

Google’s new EmbeddingGemma 2 can find videos from a single sentence. We tested it: “A cat in a hat” → nailed #1 of 40. Then we flipped the challenge: 4 clips reversed. Forward vs backward looked 99.2–99.96% identical to the model. Result? Only 1/4 guesses correct. What’s the first thing you’d search your camera roll for? 👇 #AI #ComputerVision #MachineLearning #Gemma #VideoSearch #Python #LearnOpenCV

Detection, Decoded 1/ Your object detector draws 5 boxes around 1 person. A cleanup step called NMS deletes 4 of them before you ever see it. NMS-free detectors skip that step entirely. Here's why 🧵 2/ NMS keeps the most confident box and deletes anything that overlaps it "too much". But how much is "too much"? That one number is a guess, and it's different for every camera and every scene. 3/ Make it strict, and two friends walking together become one person. Make it loose, and one person counts as three. And in a crowd, the cleanup has thousands of boxes to compare, so it slows down exactly when it matters most. 4/ Here's the surprise: the duplicates aren't a bug. Detectors are trained to make them. 5/ NMS-free training (YOLOv10, YOLO26) matches each object to exactly one prediction. The model learns to answer once. There's no cleanup step, because there's nothing to clean up. 6/ The honest part: in our CPU test, speed was almost the same. The win is a simpler pipeline: one less step, and one less number to guess. Full video 👇

Your AI Agent Failed. Do You Need a Better Model, or a Better Harness? When an agent fails, the first instinct is to upgrade the model. Often the real problem is the harness: vague tools, no write limits, "failed" errors with no detail, no record of completed work, no trace. Fix those and read the trace before you blame the model.

YOLO26 or RF-DETR? YOLO26: shortlist for edge constraints and broad task support. RF-DETR: include in custom detection/segmentation tests. Compare errors, latency and memory on your hardware. Benchmarks build the shortlist. Your data decides. #ComputerVision

OpenCV 5 Calibration Migration: The Compiler Passing Isn't Done OpenCV 5 splits calib3d into geometry, calibration, and stereo (plus point clouds), with a header compatibility path and new multi-view rig calibration. But it compiling is not "done." Save a baseline, run the same images through both versions, compare reprojection error and parameters, then test known distances. opencv.org/opencv-5/

An Agent Says It Reviewed Your Images. Did It Actually Open Them? Built a small vision agent: review 4 images (one unreadable), suggest examples, save a report. It viewed the real pixels, invented nothing, kept the failure visible. Then an independent validator checked the file list, dimensions, and original hashes. Passed. Before trusting "done," open the deliverable.

I Gave an OpenAI Dot a Hiring Task, Then Tried to Trip It Up Gave an OpenAI dot 4 fictional applications, then changed the priority, added 2 more, and slipped in a duplicate. It returned 6 unique applicants (not 7), kept claimed vs demonstrated skills separate, marked gaps unknown, and left the interview invite a draft. One staged session, no proven edge over chat yet.

Workflow or Agent? The One Question to Ask First A detector finds the car. But "is this driver allowed to park here" needs more. If records are complete, a fixed workflow gives a yes/no. If the path branches on what each step returns, an agent chooses the next step. Before adding an agent, ask: would a workflow do?

ChatGPT vs Claude. Same video. Different computer vision pipelines. We asked each to write detection + segmentation code, then hid the names. Pick A or B before the reveal. Which result would you trust—and why?

Your Agent Says the Dataset Is Ready. Would You Trust It? Resize a 1600x900 image to 640x640 without stretching: it becomes 640x360, so you pad 140px top and bottom. A broken script scales the boxes but forgets to shift them down. Dimensions look perfect, every box sits 140px too high. Overlay boxes on output, and test wide and tall images.

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.