Descriptions:
AI Explained delivers a technically dense analysis of Gemini 3.1 Pro following its February 2026 release, framing the broader argument that we have entered a “vibe era” of AI where benchmark scores are increasingly unreliable guides to real-world model quality. The root cause, the video argues, is structural: post-training now accounts for roughly 80% of total compute in modern LLMs, and domain-specific optimization means that a model excelling in one area can genuinely underperform in another — a property that simply did not hold in earlier generalist training regimes.
The analysis is anchored in specific data. Gemini 3.1 Pro scores 77.1% on ARC-AGI 2 versus Claude Opus 4.6’s 69%, yet falls behind on GDP Vow and other expert-task benchmarks. Epoch AI’s chess puzzle benchmark shows Claude Opus 4.6 scoring 10% — marginally below an older Claude Sonnet model’s 12% — illustrating that capability gains are no longer uniform. AI researcher Melanie Mitchell’s finding that changing color encoding in ARC-AGI inputs reduces accuracy adds a further methodological caution about what benchmarks actually measure.
On hallucinations, the video surfaces a telling comparison: Gemini 3.1 Pro hallucinates on 50% of its incorrect answers, versus 38% for Claude Sonnet 4.6 and 34% for Chinese model GLM 5. The video also highlights an admission buried in Gemini 3.1’s nine-page model card: deep think mode performs “considerably worse” than standard mode even at high inference budgets — a candid disclosure that sharply contradicts the launch messaging.
📺 Source: AI Explained · Published February 20, 2026
🏷️ Format: Deep Dive
