hyperknot

Gemini 2.5 Pro unusable for coding?

Something really strange is going on with Gemini 2.5 Pro.

On one hand, it's supposedly the smartest coding model ever made. But on the other hand, I ask it to add one single parameter, and instead of a simple 2-line diff, it generates a 35-line one where it randomly changes logic, removes a time.sleep() from an API call pagination loop, and is generally just totally "drunk" about what I asked it to do. It's somehow both pedantic and drunk at the same time.

Every other model, even much smaller ones, can easily make the 2-line change and leave everything else alone.

I'm wondering how this thing beat the Aider leaderboard. Did something change since the launch?

hyperknot

#!/usr/bin/env python3

import pathlib
import subprocess
import sys
import shutil

root = pathlib.Path(__file__).parent.resolve()
target_root = root / 'dsd'

hyperknot

GitHub had problem showing the file, so I moved it here (you can right click and download it): https://raw.githubusercontent.com/hyperknot/country-levels/master/export/fips/counties_population.json

Parsed from:

• all geocodes XMLS: https://www2.census.gov/programs-surveys/popest/geographies/2018/all-geocodes-v2018.xlsx
• population CSV: https://www2.census.gov/programs-surveys/popest/datasets/2010-2019/counties/totals/co-est2019-alldata.csv

All counties except Puerto Rico has population data.