AI/tech builder. Shipping in public.

Joined May 2026
I guess all the loop engineering, agent maxxing and token maxxing eventually backfired. People hammer their infra so bad that it doesn't make sense economically. And it explains the absolutely irritating experience with Claude models lately too.
10
That moment when you have to tell Claude to revert its fuckups for 20 minutes and it keeps git reverting. Infuriating. This doesn't even happen with 27B models lol.
17
Petru - Tech Driven retweeted
Sam Altman after seeing Kimi K3.1 Pro beat Astra on benchmarks at a fifth of the price…
11
6
1
110
3,830
Are we actually expecting too much from AI? Maybe all of this feedback we're giving the models that they're trash, faulty is because we internalized that narrative of AI actually being AI and doing great things, when in fact it's just a mediocre concept and should be treated as such. The disappointment you feel is directly proportional with the expectations you set. Thoughts?
1
13
Testing Argon right now, it's pretty good tbh. Doesn't ramble at all, just implements what you need. And this on MEDIUM effort.
1
26
Claude is such a time waster... would spend 30 minutes arguing about some made up BS only to admin it was wrong ... who the hell trains such a monster?? Concrete example: kept telling my that one middleware I had could be bypassed in 2 ways one of them being some CVE bug that no longer exists because it was solved a long time ago and the other one being "what if the developer imports an admin page into a non admin route?" ... I'm literally speechless... Should just say what the hell it has to say and F off, not twist it, come with explanations and arguments only to later admit it was idiotic anyway.
37
Started happening for me too. Astra xhigh chews through tokens much faster than previously.
1
15
Prompt engineering with Claude lately. 1. tell it what to do 2. make a list with 10000 things its not supposed to do. 3. repeat the list 3-4 more times cause it ignores it anyway 4. start swearing 5. it starts working
1
3
108
You have a critical mission. Either ship something or she presses it 😂 (that's what SHE said). Tag whoever you think will enjoy this. Also, we don't know who this character might be. Any ideas? (only irrelevant comments lol) turbomonkey.app/reset
2
4
64
Wonder who the final boss is or why they throw graphics cards at people ... hmmm
2
17
Just ran another audit to catch Opus's clusterfucking. This is incredibly bad ... started from a table, told it to add a column, left for 30 mins, now I have 3 tables. The amount of trash this model produces is too damn high, putting a stop on all Anthropic usage again. Also, this happens way too often in the past months ...
2
2
25
Like who the hell trains a model to reply it did something out of habbit ?? Lmao ... Unbelievable how many gimmicks can one company add to its models instead of actually fixing their BIG problems. And they're huge, not big ...
2
Not sure how people are using so many subagents to build shit ... if you don't babysit this crap it's going to produce inimaginable amounts of slop. Concrete example: I am porting an older app from Laravel PHP into NextJS. Told Claude to do the porting, explained what it needed to do. Basically it was supposed to grab all my custom code written by hand and port it into Next, import libraries bla bla. It ended up rewriting 12k lines of code to REWRITE F-ING VENDOR PACKAGES from Laravel. It even rewrote a buggy implementation of React Router. It said it wrote all my helpers and functions yet I had 3 different places with metrics calculated, each with their own separate implementation directly in the route. How are people migrating entire softwares from C to Rust or whatever? I don't get it. You can't look elsewhere for a second cause this pile of junk is going to make the most absurd decisions for you and work against you instead of FOR you. Did this on every single port I asked it, historically... hallucinated like a piece of sht, didn't port what I needed, made up things I didn't need. Funnier is the app prolly has under 1k custom lines of code, and Slopus 5.5 ended up rewriting Laravel completely. Here's the summary from the Fable Audit (Astra said the same). The port is a faithful clone of Laravel, and that is the problem. The previous session wrote its own rulebook, docs/port/RULES.md, which demands byte-for-byte PHP parity including bugs. It then followed that rule by re-implementing every framework it touched in TypeScript instead of installing the npm equivalent. The result works, but two thirds of the code is emulation of PHP, Laravel, Google's PHP client, Guzzle and react-router.
6
3
118
Rewrote everything with Astra. 30 mins tops. Everything fixed.
6
And all this happened while I was wondering why am I not using subagents to do sh*t faster. This is why. The moment it hallucinates, you're done. Happened with subagents, happened with my own frameworks trying to have them delegate to themselves. Its just slop squared.
7
Ended up deleting this clusterfuck entirely, I find it easier to just start over, than having AI fix its own junk, especially because it's MOST of the repo.
9
Ran some more analysis with Fable ... I originally had like 4400 backend lines of code. It produced 4x more, pure junk.
1
15
Claude annoying jargons and footnotes are back. Anthropic just can't hold it ... they need to do some stupid shit every once in a while ...
1
45
Meanwhile in the EU: welcome to ChatGPT 3.5, finally approved! 😂
Day 5 (dots edition)/ You can now create and text your dot entirely from the ChatGPT mobile app. Impressed so many created them via the desktop/web app previously. Time to scale!
1
1
46