RC RANDOM CHAOS

Gemini 4 Carbon Takes Hours on Tasks Competitors Solve in Minutes

· via Hacker News

Original source

Why is Gemini 4 Carbon taking so long?

Hacker News →

Google’s Gemini 4 Carbon model takes roughly an hour to debug code where competitors finish in minutes. Users report the slowness makes it nearly unusable despite the model’s strong technical capabilities.

Researchers traced the slowdown to three sources. Inference latency explains only 16–30% of the disadvantage against Astra; excessive work drives the rest. Carbon performs 7–11 times more operations per task—auditing code paths unrelated to the bug and defending against scenarios the problem already excluded. When researchers injected periodic check-in prompts asking if Carbon had enough information to fix the bug, one trial improved from 37 to 28 minutes, suggesting the model over-explores from habit rather than from information gaps.

Carbon also exhibits mechanical inefficiencies: it reads single files per step where competitors read several in parallel, and re-reads already-examined code (39% of lines versus 7% for competitors). These mechanical habits account for roughly 25% of excess runtime. The fundamental issue is behavioral—trained to prioritize correctness over efficiency, Carbon persists in thorough but unfocused debugging loops.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.