TJ Robertson
Region: US
Tuesday 28 April 2026 03:14:40 GMT
6887
291
24
29
Music
Download
Comments
BeeCee :
I dunno sometimes I feel like they spit out cheap answers or even lies just so they dont have to do the work.
2026-04-28 15:50:33
7
Aaron Hall, Attorney :
I’ve noticed: the best models do best when you tell them the problem only. They figure out solutions you might not have considered.
2026-05-12 20:53:08
0
Pmakk :
I believe your responses from ai is a direct reflection of your own intelligence.
2026-04-28 13:04:18
0
🌙SOC_MA_2026💿🌙 :
I work on the same problem across multiple models. I will spend a good amount of time developing the prompt on one model, having it co-author it with me, and then deploy that across the others.
2026-04-28 11:26:01
0
Eli Peters :
the problem with models is that they dont evolve... many wasted tokens running the same failed tasks... add it to memory but its ignored... its an llm trained on old code... old code was anything over 2 years... post AI that is likely lower now
2026-05-01 05:33:43
0
Lorris :
But is it really better this way, I was asking myself the same question. I was designing spec for a software and opus language is now doctorate level or like a Latin codex book. It makes it hard to understand how to steer it correctly. Maybe they need to tweak it a little bit, I think it's maybe over complicating things but maybe it's the only path to make it better, a bit like windows vista to windows 7, we are in-between. I went back to opus 4.6 today, much easier to work with. With 4.7 I was walking half blind, it feels like lower model like GLM which some technical people says it's as good as Claude, I don't think it is. But people who loves complicated output are going to love it, like doctors 😁😏 Also current server load makes it very hard to use Opus 4.7, almost a week past and still having errors even when the USA is asleep.
2026-04-28 14:06:34
2
JohannesDB🇩🇰 :
That is spot on!!!
2026-04-28 15:24:27
0
Star Nomad :
I’m reading soooo many reports of regression, including an analysis of thousands of lines by Claude of Claude and they’re finding it’s definitely slacking to save tokens
2026-04-28 15:43:31
1
bdamonvideo :
Prompt templates? I use an AI to write prompts for another AI
2026-04-28 15:57:05
1
RoboLawyer-tm :
there is something called architecture and something tells me that you're not actually doing that
2026-04-28 12:32:05
0
MasterEmusic :
Loosen your lasso
2026-04-28 03:44:15
1
Eli Peters :
yes! - have to always validate my prompts with "will this hamstring" a better result
2026-04-29 01:07:08
2
BraveRogue :
Nope. This is clearly model regression. I was getting amazing results until the new models. It clearly doesn’t hold onto context as well, doesn’t think as deeply consistently, requiring even more hand holding through prompts. I don’t trust it anymore. Maybe it’s a problem with it writing more memories and I should clean the slate? I’m so frustrated with the new models. Same thing happened when they got rid of GPT 4o. It felt dumbed down and everyone could few jt.
2026-04-28 15:19:10
1
Goeroe Bert :
if you need to be so very specific. why not do the work by yourself?
2026-04-28 10:39:15
0
reddirt4 :
If you have to “engineer” the prompt and constantly develop some arcane, dense technique of how you explain everything then how “intelligent” can the AI be? If you talk to a capable person, speaking in plain straightforward language typically suffices.
2026-05-01 01:51:08
0
To see more videos from user @tjrobertson52, please go to the Tikwm
homepage.
