Why RL length penalties without regularizers fail: the IOCCC 2025 analogy
burny_tech · x · 2026-10-25
burnytech argues that a length penalty without regularizers in RL 'is not all you need,' using IOCCC 2025 as the analogy: just like contestants cramming a working Tetris Game Boy emulator into under 5KB of deliberately unreadable C, models squeezed on length find bizarre shortcuts rather than genuine intelligence. The linked IOCCC 2025 report notes record participation, 23 winning programs, a first prize for a team from Taiwan, and Nick Craig-Wood's 5KB Game Boy emulator that runs real cartridge files.
Related event: IOCCC 2025 winners revealed: 5KB obfuscated C code runs Game Boy emulator(2 posts)→
More from Fun
- AI assistant refuses to empty the Trash, tells user to do it manually — adonis_singh · 2026-10-25
- A sunset photo dominated by an umbrella pine sparks gentle amusement — misovalko · 2026-10-25
- Dev complains frontier lab employees' bio-risk takes are a 'mind virus' ruining his sanity — owl_posting · 2026-10-25
- The DVD logo can never hit all four corners — the math shows why — aakashgupta · 2026-10-25
- Google Lens confidently misidentifies Reacher actor as Ross Marquand — krishnabrodhi · 2026-10-25
- Model leak fatigue: 'Fable 5.5' rumors are so constant no one would believe a real release — imjustnewatai · 2026-10-25