AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

SAM 3: The Eyes for AI — Meta's Unified Concept Segmentation Model

· Latent Space

Meta's Nikhila Ravi and Pengchuan Zhang join Roboflow co-founder Joseph Nelson to discuss the launch of SAM 3, which they describe as more than a version bump — it's an entirely new interface for segmentation. SAM 3 unifies interactive segmentation, text prompting, open-vocabulary detection, and tracking into a single model using "concept prompts" (short text phrases like "yellow school bus" or "watering can"), eliminating the need to click every instance. The launch actually included three separate models: SAM 3 (image/video), SAM 3D objects, and SAM 3D body. The team detailed the architectural decisions behind SAM 3, including decoupling the detector from the tracker to resolve identity vs. identity-agnostic representation conflicts, introducing a "presence token" that explicitly separates recognition from localization, and building on Meta's Perception Encoder. The new SA-Co benchmark contains over 200,000 unique concepts (vs. 1,200 in prior benchmarks like ELVIS), and a novel data engine reduced per-example annotation time from ~2 minutes to ~25 seconds through AI-assisted verification steps. SAM 3 runs at ~30ms per image with 100 objects on an H200 and scales linearly with object count on video (10/28/64 objects on 2/4/8 H200s). A major theme was positioning SAM 3 as "the eye" for LLMs and frontier multimodal models. The team demonstrated SAM 3 Agents, where LLMs (Gemini 2.5, Llama) call SAM 3 to ground complex visual reasoning tasks SAM 3 alone can't handle. They explicitly benchmarked against Gemini 3 Pro and Florence 2, showing SAM 3 was faster, more accurate, and produced segmentation masks rather than just bounding boxes. The philosophical framing: simple visual tasks (counting <20 objects) should be native to frontier models, while complex tasks warrant hybrid tool-call approaches. Pengchuan predicted SAM 3-native integration into multimodal models is more likely than a pure tool-call paradigm.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.