Your YouTube playlists, rewritten.

Lex Fridman · 2h 14m · 3 min read

Lex Fridman Podcast: The future of small reasoning models

A two-hour conversation on whether sub-10B models can match GPT-4 on reasoning. The key claim: recent gains come from better synthetic chain-of-thought data, not bigger architectures, and letting a small model think longer closes much of the gap on math and code, at the cost of far more tokens per answer.

The standout example is a 7B model fine-tuned on 100K filtered traces that beats its own teacher on GSM8K, a result with real implications for hobbyist fine-tuning. The host and guest spar over timelines throughout, but the closing AGI-timeline section is safely skippable.

Veritasium · 23m 41s · 2 min read

Veritasium: Why bridges are starting to vibrate themselves apart

Why slim modern pedestrian bridges are starting to wobble apart: when enough people unconsciously sync their steps to the bridge's sway, a feedback loop amplifies the motion fast, and just 5% of walkers locking in is enough to trigger the cascade. It explains why these failures appear suddenly rather than building up gradually.

Tuned mass dampers and stiffer cross-bracing both fix it, though each trades off cost or elegance. Expect more of these failures, not fewer, as computational design keeps pushing slenderness ratios every decade.