Ben in NYC
Region: US
Thursday 17 July 2025 22:36:07 GMT
1937
48
55
1
Music
Download
Comments
giorggiak :
you are expecting the code to look as if you would have written it. try doing that with another person first :) find your patterns and then do the same with the LLM.
2025-07-17 23:44:18
1
shimaru :
Try another agent. Things like Junie or Claude code or Trae can write tests, run tests, and autonomously fix the code until it works. Or until you get tired of it. Or until you run out of tokens. If a problem can't be solved in 30 minutes, it probably won't ever be solved. You should discard all and start from scratch with more detailed instructions or smaller iterations
2025-07-19 11:14:12
1
Kindragonrc :
It bogs down in complicated scripts. The bigger the project, the longer you take to vet the code. The LLM is pretty much a super snippet generator, but you didn’t write the snippet, so….
2025-07-19 00:12:07
1
Dr.Frank :
You didn't by any chance got the feeling it is testing you, to study your behaviour, and even builds purposely mistakes in, to route you to install other components, where it is than sort of testing them out...???
2025-07-18 20:00:48
1
kipsthorne :
I've only just started using AI. It has worked great for simple tasks, when I tell it exactly what to do, and it's probably taken half the time or less. Tests have been great too. But give it higher level problems, the reasoning isn't there yet.
2025-07-18 04:42:23
1
Genoma Labs :
You need to tell as example to chat gpt that grock checked the code and found errors and suggestions and you edge them, they take it personally
2025-07-18 10:14:31
1
fr3klz :
start with the tests… TDD is incredibly keeping AI agent in check
2025-07-18 13:19:50
2
Lakefront Capital LLC :
This is so true!
2025-07-18 09:56:26
1
Mountain man :
Your videos have great information. I’ve noticed something similar to this, but I think I’ve tied it to context window/drift. Even now if you get above 30 to 40,000 tokens, it will start making mistakes, big mistake mistakes, small mistakes, it will completely change what it’s doing sometimes. I tried reinforcing, anchoring, identity and purpose, writing prompts in multiple layers, but the only thing I found to fix those drift laziness is to have it summarize its duties position in those duties and what’s left to do and then open a new thread and copy and paste those duties into the new thread
2025-07-25 14:00:11
1
Chris Barker :
I found this with replit. took me from 0 to 80%, but from there anything took 10x as long as it should. then I switch to Claude code and yhoooooooo....any developer who says they are safe from AI has their head in the sand!
2025-07-21 15:08:29
1
To see more videos from user @ben.in.nyc, please go to the Tikwm
homepage.
