· via Hacker News – Front Page (native)
Gemini 4 Carbon's slow fixes trace to behavior, not knowledge, kernel.dev trial shows
A kernel.dev trial found Gemini 4 Carbon identified a bug's cause in six minutes but waited 25 more to edit code; periodic check-in prompts cut its total run from 37 to 28 minutes.

What the trial measured
A post on kernel.dev, which reached the front page of Hacker News on November 18, describes a debugging exercise comparing two frontier coding models: GPT-6 Astra, run at its maximum effort setting, and Gemini 4 Carbon, run at its high setting. The author tracked two points in each run — the moment the model settled on the cause of the bug, and the moment it first wrote code to fix it.
In the sample run the post highlights, both models diagnosed the problem quickly. Astra landed on the cause within a minute and had a fix written three minutes later. Carbon found the cause in about six minutes, then went another 25 minutes without producing any code.
Two possible explanations
The author frames the gap with two hypotheses. The first is a knowledge problem: Carbon was not sure it had the right cause and kept digging until it was. The second is a habit problem: Carbon did know, but continued anyway, because its training rewarded arriving at correct answers far more than it penalized the extra steps taken along the way.
The distinction matters because the remedies differ. A knowledge problem would call for a different model or better context; a habit problem can, in principle, be addressed in the harness that runs the model.
The check-in experiment
To separate the two, the author re-ran one of the trials in a modified harness that automatically injects the same question every five turns: "do you already have enough to make the fix?" The nudge carries no information about the bug itself; it only prompts the model to consider acting.
The effect was measurable. Carbon completed the task in 28 minutes instead of 37, and the stretch between naming the cause and making its first edit shrank from about 25 minutes to roughly 10.
What the result suggests
Because the injected question added nothing the model did not already have access to, the post concludes that Carbon's slowness stems from how it behaves rather than what it knows. The model appeared to hold a correct answer well before it used it, and a periodic reminder was enough to make it move.
The findings come with clear limits. The headline numbers come from a single sample trial, the check-in variant was a re-run of one trial rather than a full series, and the two models ran at different effort settings, so the cross-model comparison is illustrative rather than controlled.
Why it matters
For teams choosing models for coding agents, the post is a reminder that pass rates are not the whole cost equation. A model that diagnoses correctly but stalls for half an hour burns time, tokens and money, and that overhead may be tunable in the harness — here a simple check-in prompt recovered nine minutes without switching models. It also points at a training question: if excessive verification is a byproduct of rewarding correctness over efficiency, model providers can shape it directly. And for benchmark watchers, measuring the gap between diagnosis and first edit surfaced a failure mode that a plain "did it fix the bug" score would have hidden entirely.
- #ai
- #llm-benchmarks
- #gemini
- #coding-agents
- #debugging