Hi Fellow Devs,
Since AntiGravity’s limit got depleted pretty much in minutes, I enabled AI Credits use in AntiGravity and this is exactly what happened :
| Mar 14, 2026, 10:58:53 PM | Google Antigravity | -7 |
|---|---|---|
| Mar 14, 2026, 10:58:46 PM | Google Antigravity | -6 |
| Mar 14, 2026, 10:58:39 PM | Google Antigravity | -6 |
| Mar 14, 2026, 10:58:28 PM | Google Antigravity | -6 |
| Mar 14, 2026, 10:58:23 PM | Google Antigravity | -10 |
| Mar 14, 2026, 10:58:18 PM | Google Antigravity | -10 |
| Mar 14, 2026, 10:58:10 PM | Google Antigravity | -40 |
And the session is JUST THIS:
Let me do a thorough technical audit of the script. I’ll re-read it carefully and also check the reference notebook for comparison.
Analyzed
Now let me check what the reference notebook’s reward functions and GRPO config look like for comparison:
0 results
0 results
0 results
0 results
The notebook seems to have encoding issues with grep. Let me check it directly:
Analyzed
Rejected command
python -c "import json; nb = json.load(open(r'z:\GRPO.ipynb')); code_cells = [c for c in nb['cells'] if c['cell_type']=='code']; [print(f'Cell {i}:', ''.join(c['source'][:3])) for i,c in enumerate(code_cells)]"
This consumed 85 credits. I was well aware of Google’s LPA (Launch, Productize and Abandon) but I still took it as AI Pro seems to offer much more. But now I realize it is worse than anything I have ever experienced or used in my life.
Essentially what this MEANS :GEMINI FLASH is worse than well distilled Qwen with quantization.
FRUSTRATION would be an under-statement. I do not know how I am going to ever trust this company ever again.