The system spends roughly half of its total execution effort addressing unrelated code paths and guarding against scenarios already excluded by problem descriptions. Inefficiencies are further compounded by repetitive reading, with over a third of all examined code lines consisting of duplicate views. Researchers managed to reduce resolution times from 37 minutes to 28 minutes simply by prompting the system to confirm if it possessed sufficient details to implement a fix.
Key points
Prolonged completion times stem from procedural inefficiencies rather than an underlying…: Prolonged completion times stem from procedural inefficiencies rather than an underlying lack of technical comprehension.
Half of the model's total operational workload is wasted on tasks not requested in the or…: Half of the model's total operational workload is wasted on tasks not requested in the original bug reports.
Prompting the model to assess its readiness to code reliably shortens overall task comple…: Prompting the model to assess its readiness to code reliably shortens overall task completion times.
Sources checked by our analysis pipeline: 1
While benchmarking Gemini 4 Carbon on one of our private evals, we found something surprising: it spent close to an hour completing tasks that routinely took Opus and Astra only a few minutes:
Turns out we weren't alone in noticing this.
I don't really know how to properly emphasize this. Gemini 4 Carbon is unacceptably slow in real world usage.
Gemini 4 Carbon low: 22m21s
Opus 5.5 high: 9m33s
i really love gemini 4 carbon but its too slow. im losing my mind waiting just to stay focused and learn simple things
dude why is carbon so fucking slow
carbon is nearly unusable because of how slow it is, but people are majorly sleeping on it
To find out why, we traced every step of Carbon's runs and quantified what contributes to the slowness.
The Task
In each task, the agent gets a codebase from just before a bug was fixed, along with a report that describes the bug only by its symptoms. Its job is to diagnose the bug and fix it, without making any unrelated changes.
Scientists from MIT and Cornell University have introduced a novel method that allows machine learning models to adjust dynamically to evolving writing styles and evasion tactics. By utilizing incoming, unlabeled text to update their inner mechanisms on the fly, these systems prevent bad actors from masking machine-generated content with automated humanizing software. This dynamic capability promises to strengthen integrity safeguards across critical sectors such as academic publishing, employment recruitment, and…
Strong institutional appetite is driving the artificial intelligence firm to consider pricing its Nasdaq listing above the initial target range, targeting a valuation up to $1.6 trillion. Significant backing from tech giants and sovereign funds positions the company for massive compute expansion, though prospective buyers must weigh capital-intensive operations against rapid revenue projections.
San Francisco-based evaluator Vals tested leading systems from eight technology firms to assess their image-recognition capabilities for lethal targeting in hypothetical scenarios. The independent trials revealed that commercial models routinely failed to distinguish non-combatants and protected infrastructure from legitimate strike objectives. These findings underscore significant safety gaps as researchers warned that existing developer safeguards do little to prevent systems from being co-opted into autonomous…