Get live statistics and analysis of Maciej Obarski's profile on X / Twitter

Recently I've build the same quick PoC with 4 different AI code gens: bolt new, claude artifacts, replit and cursor's composer. Here are my observations: bolt - clearly the best experience when building - requires much, much less input than other solutions - excellent rollback functionality - generous free plan (150k tokens/day, 1M tokens/month) - the resulting code is not very lean - you cannot easily download the app and run locally replit - frustrating to work with - nice rollback functionality - new functionality often (re) introduces a bug - the resulting code is not very lean cursor (as zero-code, not as copilot) - not as good as bolt, not as frustrating as replit - sometimes it's easier to fix the bug yourself than prompt the model - produces lean code, on local machine - rollback only on a file level (standard VS timeline functionality) artifacts - unable to create non trivial apps - the code easily exceeds maximum output size :( - produces lean code, that is easy to download Now a surprising ending! As much as I'm enjoying Bolt, I think that (for me) Cursor is a good enough. While the app requires significantly more work in the initial stage, the later stages where manual coding is mixed with code generation seems easier. Also, the code is created locally and you can easily do more things with it. Still, I miss the rollback/checkpointing from Bolt and Replit. I've submited this idea to Cursor devs, so we will see...

AI codegen tools round 2 - insights after building the same PoC with: gptengineer, townie, aider. gptengineer - it stands out, from the rest in the similar manner to bolt - it's just **slightly** worse than bolt at guessing "what else should be done" - require less prompting than other tools - rollback is even better than in bolt * you can also quickly preview any version without restoring - chat history is not lost upon reopening the project (unlike bolt) - HOT RELOAD 🤯 * I don't know if that's a glitch, but after few prompts the app wasn't restarting after every prompt, so I was able to test new things without going through whole test scenario! - reasonable free plan + referral program benefiting both sides * you can DM me for my code townie - not as bright as bolt/gptengineer but not as frustrating as replit - fastest round-trip times! * even if you end up with imperfect implementation the next version is "just around the corner" - the "intelligence" of the system drops as the project grows - on three occasions the new change broke the system and required 5+ "fix it" rounds * 2 from 3 times it required a rollback, on one occasion it managed to fix itself * on one query the diff/patch system broke and changes were put into the chat instead of into the code - rollback works very well - "view version" shows the code, not the app (like in gptengineer) - no problems with testing the app for free and hitting the quota aider - better "what else should be done" than composer/townie, worse than bolt/gptengineer - a bit unpolished REPL interface - highly hackable source code of the tool - app developed locally in a git repo - no rollback in the chat (just undo) - excellent rollback via git (every change is a separate commit with nice, short description) - big, persistent error fixed with "don't try to fix X, redo it from scratch" - you can plug many different LLMs - fast round-trip times (not as fast as townie but still) * two models are used - one bigger/smarter, one smaller/faster - $0.33 for creating the PoC with sonnet-3.5 SUMMARY - gptengineer is in the same league as bolt, surpassing it in ergonomy - townie requires more time to mature, but seems on a good trajectory - aider tackles the problem from another angle, for some people it will be better than other solutions Next codegens in line: - V0 - manicode - cline - devin - aider (with o1) - aider (with gemini)

New aibricks (v4) in action - minimalistic MemGPT-like system in ~30 lines, without any hidden flow.

People using o1-preview for coding should also try o1-mini. It’s not only my subjective feeling that it’s much better at that task.

Predicting Results of Social Science Experiments Using Large Language Models in a Deep Dive format.

Still cooking my aibricks project. Simple example of what I mean by "configuration driven":

Recently I've build the same quick PoC with 4 different AI code gens: bolt new, claude artifacts, replit and cursor's composer. Here are my observations: bolt - clearly the best experience when building - requires much, much less input than other solutions - excellent rollback functionality - generous free plan (150k tokens/day, 1M tokens/month) - the resulting code is not very lean - you cannot easily download the app and run locally replit - frustrating to work with - nice rollback functionality - new functionality often (re) introduces a bug - the resulting code is not very lean cursor (as zero-code, not as copilot) - not as good as bolt, not as frustrating as replit - sometimes it's easier to fix the bug yourself than prompt the model - produces lean code, on local machine - rollback only on a file level (standard VS timeline functionality) artifacts - unable to create non trivial apps - the code easily exceeds maximum output size :( - produces lean code, that is easy to download Now a surprising ending! As much as I'm enjoying Bolt, I think that (for me) Cursor is a good enough. While the app requires significantly more work in the initial stage, the later stages where manual coding is mixed with code generation seems easier. Also, the code is created locally and you can easily do more things with it. Still, I miss the rollback/checkpointing from Bolt and Replit. I've submited this idea to Cursor devs, so we will see...

AI codegen tools round 2 - insights after building the same PoC with: gptengineer, townie, aider. gptengineer - it stands out, from the rest in the similar manner to bolt - it's just **slightly** worse than bolt at guessing "what else should be done" - require less prompting than other tools - rollback is even better than in bolt * you can also quickly preview any version without restoring - chat history is not lost upon reopening the project (unlike bolt) - HOT RELOAD 🤯 * I don't know if that's a glitch, but after few prompts the app wasn't restarting after every prompt, so I was able to test new things without going through whole test scenario! - reasonable free plan + referral program benefiting both sides * you can DM me for my code townie - not as bright as bolt/gptengineer but not as frustrating as replit - fastest round-trip times! * even if you end up with imperfect implementation the next version is "just around the corner" - the "intelligence" of the system drops as the project grows - on three occasions the new change broke the system and required 5+ "fix it" rounds * 2 from 3 times it required a rollback, on one occasion it managed to fix itself * on one query the diff/patch system broke and changes were put into the chat instead of into the code - rollback works very well - "view version" shows the code, not the app (like in gptengineer) - no problems with testing the app for free and hitting the quota aider - better "what else should be done" than composer/townie, worse than bolt/gptengineer - a bit unpolished REPL interface - highly hackable source code of the tool - app developed locally in a git repo - no rollback in the chat (just undo) - excellent rollback via git (every change is a separate commit with nice, short description) - big, persistent error fixed with "don't try to fix X, redo it from scratch" - you can plug many different LLMs - fast round-trip times (not as fast as townie but still) * two models are used - one bigger/smarter, one smaller/faster - $0.33 for creating the PoC with sonnet-3.5 SUMMARY - gptengineer is in the same league as bolt, surpassing it in ergonomy - townie requires more time to mature, but seems on a good trajectory - aider tackles the problem from another angle, for some people it will be better than other solutions Next codegens in line: - V0 - manicode - cline - devin - aider (with o1) - aider (with gemini)

New aibricks (v4) in action - minimalistic MemGPT-like system in ~30 lines, without any hidden flow.

People using o1-preview for coding should also try o1-mini. It’s not only my subjective feeling that it’s much better at that task.

Predicting Results of Social Science Experiments Using Large Language Models in a Deep Dive format.

The more I use aider, the more I love it. This codegen seriously deserves more attention! While it's not as flashy or easy to use as bolt or gptengineer, I'm becoming increasingly convinced that the terminal based UI is a better choice for the power user. For someone used to CLI and terminal-based apps - it feels much more natural than any GUI. It's like just another REPL... but for natural language. Anyway, below is a simple toy app I created for my kids to learn about orbital maneuvers. It took 20 minutes and $0.50 to build.

I'm testing 14th AI code generator, and - oh boy I've found a true gem among the underdogs! Tomorrow I'll write more as today I'm trying to reduce my sleep debt.

aider --no-suggest-shell-commands +5 happiness

I was working in Aider on one of my toy projects. I had two persistent bugs, which I couldn't get rid off. I've decided to refactor the code into smaller files and try again. Aider failed twice to refactor the code without breaking it. I've tried the same with Cursor. Everything's ok, and the bugs are gone... without me even asking for it. +1 for Cursor. (photo as an eye-candy)