Descriptions:
Matthew Berman, AI commentator with 600,000 YouTube subscribers, hosts a casual live stream probing the boundaries of current AI models through adversarial and edge-case prompts — what he calls “baiting” AI. The session tests Gemini on spatial reasoning tasks and classic logic puzzles, exploring where frontier models still stumble and why certain capability gaps persist despite rapid overall progress.
A central thread is Berman’s explanation of why AI models have become so strong at coding: the feedback loop is short, errors produce clear messages, results are easily verified, and the commercial incentive is enormous — he cites Anthropic reaching $44 billion in revenue driven primarily by coding token sales. He also references Andrej Karpathy’s observation that labs “patch” famous failure cases like the strawberry letter-counting problem by injecting targeted training examples, raising questions about what genuine generalization looks like versus surface-level fixes.
The stream includes a live discussion of Gemini 3’s performance on common-sense spatial prompts, the jagged nature of AI intelligence (strong on some dimensions, brittle on others), and the challenge of evaluating creative writing quality without human feedback loops. The opening minutes are occupied by audio troubleshooting, but viewers who stick with it get a frank, technically grounded conversation about model capability frontiers from a creator who has been closely tracking this space for years.
📺 Source: Matthew Berman · Published May 01, 2026
🏷️ Format: Livestream
