Bilingual

We charge $10k a week to delete AI-generated code
15 days ago
- We analyze AI-generated codebases that have become bloated and difficult to maintain.
- We offer a free analysis and commit to a specific code reduction target (e.g., from 100,000 lines to 35,000).
- The process involves a week of focused work with three senior engineers, starting with a detailed checklist of app functionality.
- We refactor code by consolidating duplicates, replacing custom frameworks with libraries, and rebuilding unsalvageable parts.
- You receive the reduced codebase, QA checklist, guardrails (like CLAUDE.md and lint rules), and a two-week warranty.
- Pricing is $10,000, with payment proportional to the achieved reduction target; only pay for results delivered.
- Lines are counted fairly (non-blank, non-comment), and we avoid code golf to ensure readability.
- We are real engineers (not agents) using Claude Code under strict oversight, based on decades of experience.
- Contact maciej@odra.dev for a free analysis of your repository if this resonates with your project challenges.
Fable 5 wrote a Windows kernel in 38 minutes
15 days ago
- The ntoskrnl-rs project successfully rewrote the Windows NT kernel in Rust, booting in QEMU and passing self-tests.
- Claude Fable 5 generated the core kernel (about 5,100 lines) in 38 minutes, including scheduling, memory management, and I/O subsystems.
- The kernel loads unmodified Windows drivers and binaries, providing a sandbox for analysis and debugging.
- Verification remains a bottleneck, as AI-authored code requires extensive testing for trust and correctness.
- Model choice matters: Fable excelled at greenfield generation, while Opus handled iterative debugging and security work.
- This demonstrates AI's potential to rewrite legacy critical infrastructure, reducing costs and memory-safety bugs.
A Large-Scale Empirical Study of AI-Generated Code in Real-World Repositories
9 days ago
- The study conducts a large-scale measurement comparing AI-generated and human-written code in real-world repositories.
- It addresses limitations of prior lab-based studies with synthetic benchmarks and small-scale tasks.
- Metrics cover code-level aspects like complexity, style, security, and commit-level traits such as size, frequency, and stability.
- Findings reveal small differences in code-level metrics between AI and human code in real-world settings.
- Insights include variance in security quality across programming languages and new data on code duplication and commit characteristics.
- The results provide implications for AI-assisted programming practices.
Reviewing code you didn't write
10 hours ago
- Reviewing code is a separate skill from writing code, focused on finding critical issues that determine whether the work should ship.
- Reconstruct the change's context by checking out the branch, reading tests, and understanding the intent from tickets or design notes.
- Review the core path first: evaluate the approach and check for duplication before diving into details.
- Order review comments by priority: approach, then correctness, failure, security, and evidence; block merges only for critical issues and use 'nit:' for preferences.
- When reviewing AI-generated code, be skeptical of invented APIs, unnecessary abstractions, and overly complete test suites; verify against the task and project rules.
- Improve review skills by studying merged PRs and seeking feedback from other engineers on your own review comments.

Hasty Briefsbeta

#ai-generated code

We charge $10k a week to delete AI-generated code

Fable 5 wrote a Windows kernel in 38 minutes

A Large-Scale Empirical Study of AI-Generated Code in Real-World Repositories

Reviewing code you didn't write