// channel ▣
Deep Dives
Long-form technical investigations. We take one idea apart, bolt by bolt, and put it back together in front of you.
3 dispatches in this channel
Deep Dive
Three Kinds of Smart: Why 'Best Model' Is a Category Error
Raw intelligence, SWE effectiveness, and flash-tier quality for real-world work are three different competitions — and the leaderboard keeps conflating them. A map of who actually holds what.
3 min read
Deep Dive
The Context Window Quality Problem: Bigger Isn't Always Better
A 10M-token context window sounds impressive. But attention quality degrades across long contexts, and most public evaluations don't test beyond 128K. Here's what actually matters.
3 min read
Deep Dive
Inside DeepSeek V4's Reasoning Stack
DeepSeek V4 posts frontier-class math and code scores at a fraction of the inference cost. We dissect the publicly documented pieces of its reasoning pipeline.
1 min read