NateWiki — Open Brain corpus
Home · The Doctrine · All topics · Search
priority 2 35 posts

Model Reviews, Benchmarks & Picking

The playbook

Nate's core claim, repeated and sharpened across two years of releases, is that "which model is best" stopped being the right question once every frontier lab cleared the floor for easy tasks. What's left is a fragmenting capability surface, and the real skill is routing: matching a specific job to the model whose optimization target fits it, then verifying the output yourself.

His method evolved from ad hoc side-by-sides (Claude 3.7 vs. six rivals on a Monopoly-simulator prompt, o3 vs. Gemini 2.5 on invented "resonance" tests) into a repeatable stress-test doctrine: give models deliberately hostile, real-world-shaped work — poisoned CSVs with duplicates and SQL injection, a 465-file small-business "shoebox" migration seeded with canary records (Mickey Mouse, a fake $25,000 payment), a handwritten Christmas-tree tally sheet — because clean benchmarks reward answer-key memorization (Grok 4's #1 leaderboard score collapsed to real-world garbage on identical tasks) while messy tests expose a model's actual stance toward ambiguity. He names three recurring stances: Claude reconstructs faithfully and flags uncertainty; GPT imposes structure and will fabricate coherence when reality won't cooperate; Gemini interprets for insight but lets its own outputs contradict each other.

The throughline that hardened by 2026: capability gains no longer show up as "smarter," they show up as narrower and more literal. Newer models (GPT-5.x, Opus 4.6+) increasingly do exactly what you ask and nothing more — the inference that used to paper over sloppy prompts is gone, so prompts must become specs (role, phases, deliverables, constraints, checkability) rather than wishes. Self-grading is unreliable in both directions (Opus oversells itself, GPT undersells), so peer review — one model checking another's work — is the only pattern he trusts for anything with stakes. Compute availability and harness (Codex, Claude Code, Cowork) now matter as much as raw weights; a slightly weaker model you can actually run beats a stronger one throttled by caps. And with Chinese open-weight models (DeepSeek, Kimi, GLM) closing the gap on price and matching mid-tier benchmarks, the operative question shifted again: not "is it good enough" in the abstract, but "what does an accepted result cost end-to-end" — tokens plus retries plus review time plus what breaks if you have to switch models later.

Key moves

The posts

Related