Limits and Possibilities of AI Coding Tools: From Model Anomalies to AI-Driven Development
코딩/개발 | Sun Jul 05 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources
Real-world coding cases including GPT-5.5 Codex token clustering anomalies, Claude tool call errors, and AI-driven PHP engine development.
Analysis
[OpenAI Codex] reported reasoning token clustering anomaly in GPT-5.5 [1]
- reasoning_output_tokens concentrated exactly at 516
- fixed boundary spikes also observed at 1034 and 1552
- GPT-5.5 accounts for 82% of exact-516 events
- possible correlation with degraded performance on complex Codex tasks
[Anthropic Claude] discovered worsening tool call schema errors in latest models [2]
- Opus 4.8 and Sonnet 5 insert non-existent fields into edit tools
- occurs more frequently in SOTA models than in older models
- tool calls processed as XML-like in-band signalling
- errors particularly occur in nested array structures
[Phargo] released Rust-based PHP interpreter written by AI [3]
- rewritten from scratch with 0 lines of PHP original source code
- passes 17.4% of 22
- 037 official PHP test suite tests
- successfully renders WordPress front page
- developer does not know Rust and only directs the AI
[Rust Compiler] encountered code optimization bug in JavaScript engine parser [4]
- parser error occurred in bool as u32 casting version
- explicit if-else version worked correctly
- branch elimination optimization produced asm a few bytes shorter
- different results produced from identical logic