AI can't reading code by itself

As the AI agent code 100% of the scripts, why still it ask me to provide code snippet?
i have put in Gemini key in .env.local file (and not .env), it still couldn’t figure out the same mistake where AI stuck at thinking stage

Jz my gut feel, it’s the third tic-tac-toe projects i vibe in Firebase Studio

  • First - mid June, very effective, smart. Firebase Auth success in first attempt (without much inform the AI). It even work well by storing game history, and the ultimate tic tac toe version without problem. but these days it failed to publish (then i started the second one)
  • Second - can feel obviously “dumb-er” , it gave lower quality UI, firebase auth failed, i can’t stand on it and started third vibe journey
  • Third (current version) - UI as good as first version, but the AI player stuck at “Thinking”

i started to have conspiracy theory that coz i burnt so much tokens for free and FB start assign not-so-smart AI model lol

Switch to code view and use a more advance AI model.

hi dms, i can (if i want to), but i am evaluating how efficient firebase studio could help my team to dev a working product with very lean developer team (like 1-3 developers).

I started to worry if such frequent “hallucination” required me to turn on 2.5 PRO (it’s not cheap). It cause around $0.5 in 2-4 human prompts (i can’t recall clearly)

and today it complained it can’t put the code in my project (which the studio is designed to)

and i saw the code broke by itself (yea i know i am the one commit the change)
the AI agent introduced the variable names in .env file, then later change it done two silly things

  1. Wipe out .env.local, keep only 1 variable (idk why)
  2. Introduce diff set of variable names in actual scripts and .env file

Other than the issue it can’t overwrite my code, i spotted it asked me to overwrite an incomplete code (as screenshot below)

It’s a fun experience to deal with the agent before i could recommend to the teams :smiley:

If you’re a startup working on AI, you can apply to the Google Startup Program to get up to $350,000 in free Google Cloud credits. That alone should help you reduce your costs.

As for the model getting confused, other IDE platforms have a method of “context” where you can select which files/folders the model should consisder when making a request. For Firebase Studio, try being extremely deliberate and name the files or folders you wish to adjust in your prompt. This will help guide the model and give it better direction.

When you have completed a task with the model, clear the chat or start a new one for best results.

Sometimes the prototyping AI makes mistakes, such as recreating a file that already exists, and I have to prompt it to review all files.

tht’s cool with $350k :slight_smile:

coz i m switching btwn prototyper and the “code” mode to input firebase project param, i noticed both agents in code/prototyper are two distinct agents. It could be the reason they contradict each other

1 Like