- figure out the best way to use multiple models in an agentic IDE. simple tasks should use cheaper models, and only complex ideas should use latest models. Basic changes like CSS should be done manually without the need to use AI at all. It all falls down to how you use your brain to reach efficiency. on which models to use and when.
- using AI is like using frameworks. you are calling some simple looking builtin functions to do monstrously complex stuff behind the scene.
- the money is not on what you can do but how more efficiently you can do what has already been done. Some frameworks do things better than others. and much of your value is knowing the hidden gems in the tech industry. Better tools sometimes can do more than simply knowing some syntaxes, in the long run.
- we should focus on the ideas, and much less on how to implement them since humans are meant to do creative stuff and machines, the repetative stuff. so we should use AI as tools to get past the tedious delays in implementing our ideas, which actually matter
- "Learning to do things manually", a friend of mine prefers, to learn to make components himself to understand how it is done. He wishes to learn DSA to empower his web dev career. I do agree that leaving it all to a blackbox may have harm in the long term.
# Useful Videos
- for Antigravity Antigravity Rate Limits Decoded - Beat Them Like This - Eric Michaud
- Antigravity does not limit on tokens, but on “work done” (??). bigger contexts (prompts) mean higher token costs.
- do the planning outside the agentic IDE. (use arena.ai and feed it to Gemini Flash, so it doesn’t have to think (which it is bad at)
- always to a fresh start; Antigravity checks through the entire chat every time
- keep a session_state.md for a comprehensive hand-off or “picking up where you left” instead of starting all over for each chat.
- match model to the task: Gemini Flash < Claude Sonnet < Gemini Pro < Claude Opus
- pinpoint the file (and exact range of lines?) that needs to be changed. You don’t want the agent to look through your entire project
- don’t “always allow” when running an action
- "The smaller the task, the better the results"
- "If you cannot break a big problem in small parts, you don't understand the problem well enough"
- fundamental engineering: "problem solving" and "critical thinking"
- plan out a solution
- breaking the problem down
- code it
- “(in its current state) Do not let AI do all the thinking for you. I am personally okay
- one of the must watch videos for the AI era I would argue
Comments
Post a Comment