Why Current AI Cannot Truly See the World
summary
Claims that frontier multimodal models systematically fail on grounded visual reasoning, with concrete examples: a railing image has 37 rails answered 2, a video of players has 16 passes vs model 6, and a robot-gripper question confuses a stove knob with a pan. Builds a vision-first model trained on human-labeled and synthetic egocentric/video data, using interleaved image+text visual thinking traces and image decoding as reasoning steps; reports state-of-the-art real-world counting and plans public benchmarks and early access.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 3:37Current AI vision paradigms
- 8:14Elorian AI approach and team
- 14:18Limitations of current benchmarks
- 17:31Visual thinking vs hallucination
- 20:00Future plans and Q&A
related talks
13mArtificial General Intelligence: Why aren't we there yet?
19mRedefining Intelligence: How Reasoning Is Re-Shaping AI in 2025
27mThe Cutting Edge of AI: From LLMs to Multimodal Intelligence
25mFrontier AI Models: A Review of Past Nine Months
24mMulti-Modal AI Agents Explained | Open-Vocabulary Alignment
17mVision AI in 2025
15mMulti-Modal AI Optimization: Combining Text, Vision, and Speech Efficiently
39mNew multimodal vision AI models and their practical applications