Astra-generated code now unreadable to humans; corrigibility becomes a matter of faith

On September 16, multiple community authors converged on the same observation: if you actually stop and read the code produced by agents like Astra (and presumably Fable), you find heavy use of wild metaprogramming and obscure primitives in exchange for highly optimized output. Humans can no longer meaningfully review or correct such code, making corrigibility a matter of faith.

Confirmed

tszzl and aryaman2020 independently noted that Astra's (and likely Fable's) code relies on aggressive metaprogramming and obscure primitives for top-tier performance.
tszzl argued the only remaining option is to send another Astra to read and use the first Astra's output.
tszzl stressed this shift is very recent: roughly one to one-and-a-half model generations ago people still read generated code because much of it was unusable out of the box; now quality is high enough that line-by-line review has stopped.
aryaman2020 added 'works though' — the code is hard to read but genuinely runs and performs.

Why it matters

The discussion cuts to a core AI-safety concept: when AI-generated code is incomprehensible to humans, review and correction become impossible, and safety guarantees degrade from engineering to trust.
The transition happened within one to one-and-a-half model generations, far faster than anticipated, suggesting traditional code-review processes may already be obsolete.
The coexistence of 'it works' and 'it's unreadable' highlights a structural trade-off between performance and readability/maintainability: engineering has no incentive to give up performance, while safety loses its grip.

2026-09-16 ~ 2026-09-16 · 5 related posts

Primary sources

Frontier coding agents now write code no human can read, warns tszzl — tszzl · 2026-09-16 12:14 (UTC+08:00)
tszzl: we now need another Astra to read Astra's output — a shift within 1-1.5 model generations — tszzl · 2026-09-16 12:14 (UTC+08:00)
"Works though": the dilemma of unreadable but functional Astra-generated code — aryaman2020 · 2026-09-16 12:28 (UTC+08:00)

[source] Frontier coding agents now write code no human can read, warns tszzl — tszzl · 2026-09-16
[source] tszzl: we now need another Astra to read Astra's output — a shift within 1-1.5 model generations — tszzl · 2026-09-16
RL-trained Astra writes uniquely unreadable code as corrigibility becomes a matter of faith — aryaman2020 · 2026-09-16
[source] "Works though": the dilemma of unreadable but functional Astra-generated code — aryaman2020 · 2026-09-16
Reading agent-written code: 'corrigibility' has become a matter of faith — daniel_mac8 · 2026-09-16