Dev · Monday, 24 August 2026 · Issue 86 · 10 stories

LLMs that verify, not just generate; AI code's hidden debt

Two stories that look unrelated until you see the same lesson: the output is never the answer. A 27B model cracked a license check in 30 minutes, but the real news is how it caught its own errors. And AI code passes tests while rotting your architecture. The skill isn't generating, it's verifying.

Languages & Frameworks

Akka-style actors for Rust: tellus 0.1.0 adds supervision trees and death watch

Nine years after writing about Akka's actor model, Heiko Seeberger has ported the idea to Rust. tellus 0.1.0 runs actors on Tokio with typed messages, supervision trees, and a death watch whose termination signal is ordered after every message the dying actor sent, so receiving it proves you saw them all. The break with kameo and ractor: receive() is fully synchronous, with no per-message async futures, boxes, or Send-bounds strapped to local state.

heikoseeberger.de ↗7 min read at source

Tools & Platforms

LLM-as-a-Verifier: an open engine that scores answers, not trust-me-bro outputs

A general-purpose, open-source framework that stops treating an LLM's word as final: it runs fine-grained verification over candidates, scoring each answer on criteria rather than accepting the model's assertion. The trick is averaging over the full logprob distribution of score tokens plus scaled repeated evaluation. Selection beats Pass@1 even when the verifier judges its own model's rollouts.

github.com ↗10 min read at source

ChatGPT search just started leaning hard on the site: operator

Promptwatch's tracking shows the share of ChatGPT Search queries using the site: operator jumped from ~0.4% to 16-17% overnight, aligned with the GPT-5.6 rollout. The invisible change has real SEO consequences: site owners can suddenly steer generative answers by domain in a way they couldn't before. It also signals a shift toward more restricted, provenance-aware retrieval, the market OpenAI then quietly reduced Reddit sourcing too. For anyone doing "GEO", this is ground truth from outside OpenAI's opaque system prompt.

simonwillison.net ↗2 min read at source

Crankwave: bake engine audio once, replay it with zero simulation

Crankwave is a deterministic C++ engine-simulator that treats the full simulation as a build-time step, not a runtime dependency. Define engines in JSON, audition or bake audio through the same native and WebAssembly core, then ship a hash-bound .crankwave package that game clients replay from just RPM, throttle, and load. It's an MIT-licensed source rewrite of AngeTheGreat's engine-sim without the upstream GUI, scripting runtime, or .mr input path.

github.com ↗3 min read at source

AI-Assisted Development

A 27B local model cracked a license check in 30 minutes

Qwen 3.8 27B, running on a single compact workstation, reverse-engineered a commercial app's license check in 30 minutes and caught and fixed its own errors along the way. It refused the jailbreak prompt, correctly called out that the author wasn't the developer, then produced a full audit up to the bypass line. A sharp demonstration that frontier-scale models may not be required for serious security work.

xda-developers.com ↗9 min read at source

AI code looks clean and is architecturally rotten: the data proves it

AI-generated code arrives syntactically perfect and structurally incoherent: models can't see your codebase's conventions, so they re-implement what utils already do and ignore bounded contexts you spent weeks establishing. The numbers make it concrete, GitClear shows 4× more duplication, refactoring collapsing from 25% to under 10%, and 7.9% churn within two weeks. Passing tests and clean lint are the trap; the debt surfaces in review, months later.

xylos.com ↗14 min read at source

Coding agents demand a new skill: verifying output, not just reviewing it

The skill that separates productive agent users is the confidence to direct a change and then confidently confirm it landed correctly. Willison argues that eyeballing every generated line has never been the best check on a software change. Real verification means knowing what else you can inspect, tests, behavior, diffs, when line-by-line review isn't the practical move. It reframes code review as a verification mindset, not an exhaustive read.

simonwillison.net ↗1 min read at source

Engineering Practice

Chmod +x a SQLite database: a working ELF replacement

What if the binary you run were a SQLite file you could query? SELF, a fleshed-out prototype from a PhD thesis, makes chmod +x mean "run a database" and lets you `SELECT soname FROM ldd` instead of parsing ELF by hand. The killer argument: ELF already is a database, just one that hand-implements string interning, indexes, and foreign keys. Turning executables into real databases kills a generation of bespoke parsers.

fzakaria.com ↗14 min read at source

Redox OS swaps its scheduler for EEVDF, with huge fairness wins

Redox OS replaced its Deficit Weighted Round Robin scheduler with an EEVDF-based one, and the gains are dramatic: a 782x improvement in fairness, an 82% cut in context-switch time, and 2.6x more throughput. The write-up walks through the lag-and-deadline mechanics in an unusually readable beer-bar analogy while crediting Linux's own proven EEVDF switch. A compact case study in why scheduler choice still moves the needle on a modern OS.

redox-os.org ↗14 min read at source

Other

Write Postgres functions in TypeScript, Ruby, or SQL dialects, compiled to safe plpgsql

plx is a PostgreSQL extension that compiles Ruby, PHP, JS, TypeScript, Python, Go, COBOL, Oracle PL/SQL, and T-SQL bodies into plain plpgsql at CREATE FUNCTION time. Unlike native PL/* extensions, no interpreter loads into the backend, the catalog stores real plpgsql, so every plpgsql performance and trust guarantee applies. Write in the syntax your team knows; run under PostgreSQL's own interpreter.

commandprompt.github.io ↗2 min read at source

You’re caught up.

That was all 10 stories in the 24 August issue.