Why is Gemini 4 Carbon taking so long?

Hacker News by 2 min read 171x views
Why is Gemini 4 Carbon taking so long?

Share Post

Sample trial: period to discover the logic vs. period to archetypal edit

Minutes from the commencement of all run. GPT-6 Astra at max effort, Gemini 4 Carbon at high.

In this example trial, the two models identified the logic early, Astra inside a infinitesimal and Carbon inside concerning six. Astra had its fix written three minutes later; Carbon didn't compose any code for another 25 minutes.

Carbon power not have been certain it had the logic (a cognition gap), or it power have known and kept going anyhow (a broad habit, learned from training that rewards correct answers additional than it penalises additional steps). To inform them apart, we re-ran among the trials in a modified harness that automatically injects the identical check-in all five turns: "do you already have adequate to create the fix?"

Sample test alongside check-ins: period to discover the logic vs. period to archetypal edit

The identical bug and measure as above. Gemini 4 Carbon at high, alongside the check-in injected all five turns (circles).

With check-ins, Carbon completed in 28 minutes alternatively of 37, and the delay between naming the logic and editing shrank from 25 minutes to concerning 10. This implies that the slowness seems to arrive from how Carbon behaves, not what it knows.

Other Article Hacker News
↑
Close Right Ads
Close Left Ads