Kernel traced Gemini 4 Carbon through bug-fixing tasks to find why it sometimes took nearly an hour where other models took minutes. Slow responses explain part of the gap. The assistant also kept investigating after finding the cause.
- Extra work: Carbon averaged 133–140 steps per bug, against Astra’s 19–20. Kernel classified roughly half its steps in a separate trace analysis as investigation beyond the repair’s scope.
- Repeated reading: 39% of the code lines Carbon read had already been read, against Astra’s 7%. It also grouped fewer tool calls into each step.
- A check-in helped once: asking “do you already have enough to make the fix?” every five turns shortened one illustrated run from about 37 to 28 minutes.
The task was a bounded repair: fix the reported bug and leave unrelated behavior alone. Kernel also tried two other agent wrappers, the software running the model’s tools, and reported similar behavior.
Caveat: the tests are private, with no disclosed total task count or full repair-success comparison. One faster rerun suggests a stopping prompt worth testing; it cannot establish a general speedup or its effect on correctness.
The thread on Hacker News adds practical checks and a disagreement about when broader investigation pays off.
What the thread adds
- chriddyp brings a separate data-analysis test: “10.3 model calls per question vs 7.7 for Hydrogen.” Plotly’s November 18 log analysis explains that both models made about 15 tool calls per question. Carbon grouped fewer queries into each model response, paying its startup wait more often. Plotly also measured about 500 output tokens per request, including hidden thinking, against Hydrogen’s 780. In that workload, shorter responses still produced slower tasks. Plotly sells analytics software; its private exam and single run per model limit generalization to coding repairs.
- timtam — “I only got around the issue by strictly setting an effort budget or a stopping rule.”
- jedbrooke — their wrapper repeated “Be helpful, read in between the lines and proactively go beyond the user’s request” at each turn. “I immediately patched that out and saw decent improvement in that area.”
- bagacrap — “I’m much more likely to ask Carbon to review some work I’ve done than to help enact my ideas in the first place.” Broader exploration can serve a review even when it exceeds a narrowly specified fix.
Handles are pseudonymous, HN publishes no per-comment scores, and retrieval follows HN’s ranking; this is a slice of the thread, not a consensus.