The AI twin gap, week 44: first-hand pieces won by 40 points, familiar ones by none

Human essays beat their AI twins in 7 of 10 pairs this week, by a mean of 23.0 points. Almost all of that margin came from five pieces that held information a model could not have had.

The most-read human essays on Hacker News this week beat their AI twins in 7 of 10 pairs, by a mean of 23.0 points: 58.3 against 35.3. That average hides a lopsided split. Five pairs contributed 221 of the 230 points of total gap, and the other five were close to level: two ties, one human win by a single point, one AI win and one modest human lead. What separated the two groups was less about the writing than about where the material came from.

The widest gap belonged to Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5, which scored 92 against its twin's 37, a gap of 55. Verdict called it "a substantial original build that compiles a working RV32IM CPU into Qwen3.5 weights" and noted that it "backs the claim with its own validation and benchmark numbers, though readers cannot check them here." A model given only the title can describe the idea. It cannot report the build.

Close behind came Phorion's Official Claude Code Extension Compromise: Security Advisory, at 87 against 46. The verdict line reads like a checklist of things that only exist because someone did the work: "a hands-on teardown of the backdoored Cursor Marketplace Claude Code extension, giving specific IOCs, the bugs that blunt the payload on each platform, and practical detection advice." Indicators of compromise and platform-specific bugs are exactly what a twin writing from the topic alone has no way to produce.

Three more large gaps came from primary sources rather than original research. DeepMind's Genie 4: A new generation of world simulation scored 69 against 28, as "the primary announcement" listing native audio, better human and physics simulation, Gemini-guided planning and a limited research preview. Mistral's Mistral Large 4 (Le Chonk) is now open weight scored 63 against 24 for concrete benchmark, safety, licence and pricing details, although Verdict flagged that "nearly all performance claims are self-reported and unverified." Most striking is the Imagination Library home page, which scored 70 against 25 despite "adding nothing beyond the program's own messaging." Being the source of current, accurate facts counts for a great deal when the alternative is a model guessing at them.

In the middle sits Apple's git still trusts the compromised eMudhra certificates, at 63 against 51. Verdict described it as "a focused, demonstrated report" that macOS's bundled git, curl and python3 still trusted the compromised certificates after 27.2. The human piece won clearly, but its twin scored 51, the highest of any twin this week. The brief does not say why, so I will not guess.

The small gaps are more instructive. A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis is original work, "an original benchmark by a dermatologist" in which a 20-minute fine-tuned Inception v3 beat two frontier LLMs, 62% to 42%, "though it still sent 18 of 30 melanomas home." It scored 47 against its twin's 46. Originality of method alone did not open a wide gap here.

Why Is This Fingerprinting Script Creating Bioweapons? tied at 40. Verdict credited "original crawl-based research showing an anti-bot script that plants safety-triggering text to stop LLMs deobfuscating it," then added that "its actual data collection is routine." The new finding is narrow, and much of the surrounding piece covers ground a model can already describe.

At the bottom are the pieces an AI could largely have written anyway. Sebastian Raschka's Build Your Own Reasoning Model: A Full Video Course tied at 32 as "a promotional course outline" with "a few specific design notes." And the only twin win of the week went against Orestis Ioannou's AI Promised Less Work. I Have Never Worked More., which scored 20 to the twin's 24. Verdict called it a "familiar AI-workload paradox made concrete by the author's own GitHub contribution statistics across roles and years." Personal data helped, but not enough to lift an argument most readers, and most models, have already met.

Each twin was written by a current model from the title and topic alone, and both sides were scored by Expulp Verdict the same way, claim by claim. Ten pairs is a small sample, and the full benchmark of 638 pieces is the better guide to how general this pattern is.

For a writer, the useful question is what in a draft a model could not have produced from the headline: a build, a teardown, a measurement, or a fact that only the source holds. If the honest answer is a familiar argument with a personal anecdote attached, the anecdote needs to carry far more of the piece, or the piece needs something new behind it.