Subject: Feedback on Gemini 3.6 Flash Model Performance
Hi Antigravity Team,
I would like to report a significant decline in the response quality of the Gemini 3.6 Flash model over the past few days. The model now exhibits severe “dumbing down” behavior, particularly when tasked with modifying or debugging existing code.
Problem Description:
When I ask the model to fix bugs or refactor code, it frequently fails to resolve the issue, even after multiple attempts. In many cases, the model introduces new errors or breaks previously correct sections of the code. For relatively simple syntax errors, it sometimes takes 3-5 rounds of back-and-forth without reaching a working solution, whereas previously it would resolve such issues in the first attempt.
Steps to Reproduce:
- Provide a block of code with a known, simple bug (e.g., a missing parenthesis or a type mismatch in Python/JavaScript).
- Ask the model: “Please fix the bug and output the corrected code.”
- Observe the response: The model either outputs the same incorrect code, modifies the wrong part, or adds new syntax errors.
Example (Illustrative):
(Note: I have omitted the actual code for brevity, but can provide it upon request. The issue is consistent across multiple languages)
Expected vs. Actual Behavior:
- Expected: The model should accurately identify the bug and provide a clean, corrected version of the code.
- Actual: The model provides an “answer” that still contains the original bug or introduces regressions.
Environment:
- Antigravity Version: (You can fill this in if you know it, e.g., v1.2.3 or Browser version)
- Model: Gemini 3.6 Flash (may also affect other variants)
This is severely impacting productivity, as the model is no longer reliable for iterative development tasks. I kindly request the team to investigate any recent changes or rollbacks on the backend that might have caused this regression.
Thank you for your attention. Please let me know if you need any logs or specific code examples to help debug this issue.