Petru - Tech Driven

@adhamuxi Confidence is some internal state, not anchored in reality.

@adhamuxi Or to brainwash yourself to the point where you are delusional 😂

The fastest way to build confidence is to build skill.

@adhamuxi More like Argentina should be banned from international football for a couple of years for their attitude.

Spain 1–0 Argentina or Spain 3–0 Argentina — which one is correct according to you?

@adhamuxi Actually should’ve said “you’re absolutely right”

@techdrivenpetru At this point, half the timeline is AI talking to AI anyway

What's one way ChatGPT has helped you do something you didn't think you could before?

@adhamuxi Hahaha true

Hmmm

@Krishna_agg_ @Im_IrushiK right lol

@Im_IrushiK Yes, I don’t mind using Fable. It’s building my app without any bugs.

Either this is the best gift ever… or Claude is fighting hard to stay relevant with their nerfed models, right??

@_mattwelter How did you get there in the first place lol

1 year ago i made $48,000 in a single month but now with the product being 10x more developed and better in every way, i'm not hitting that $48k anymore i'm at $30k a month how do i get back to $48k months plz

@rubenssoto_ai Because that's what the influencer said.

Why do people use Supabase for small SaaS projects? Why overcomplicate things and sometimes spend money when SQLite is good enough for most use cases?

@13_narcissus But AI is dope, just don't outsource the thinking and software logic, cause then you're toast. Other than that, it can write way faster than you.

Every senior dev I know quietly uses AI all day. Every loud AI hater I know hasn't shipped anything in years. Draw your own conclusion.

@Oukly_ @JamesWill932 @psomkarbuilds Never trust benchmarks. You know how it goes. When a metric becomes a north star, it's no longer a good metric. But I'm glad you find it better, maybe it's just in my head, maybe i'm being routed to some cluster with crappy configs, dunno. Maybe EU infra sucks

@techdrivenpetru @JamesWill932 @psomkarbuilds Yes I meant in quality. Also there are plenty of benchmarks no need to guess

At this point if AI writes 90% of code, who even survives in tech?

@13_narcissus That's a yes/no question. Technically, a sentence is unsolicited advice :)

A junior dev asks you: "Should I still learn to code in 2026?" You get one sentence. Go.

For me at least lately at least Claude, keeps failing at most basic instructions, to the point where I moved to Codex, like... I tell it to use some MCP, it starts looking in folders, I tell it to save some config file, it puts it into the wrong folder, I tell it to revert, it does the same thing again, tell it to move it, it moves something else or in the wrong dir... like ... for some operations i'm way faster and better to just do it manually.

@techdrivenpetru @JamesWill932 @psomkarbuilds Yes huge difference, like night and day.

100% comparable yea. They had the same amount of programming data to train on. Nothing more has been produced ever since that would make any difference. As an exampple: I designed a recommender system that runs on its own ML model -> 100% vibe coded it with GPT 3.5 and Claude web a few months before Claude code was released, just by copy pasting in the browser. The models were good from the start, just too expensive to make sense economically, just like they are now. They still don't have any intelligence, just like they didn't had before. They still hallucinate and fail, just like they did before, they just have tool calling now and more fake reasoning, which isn't reasoning, it's just multiple runs for the same output. And don't mix all the harnesses and tools around AI with the actual model output quality :)

@techdrivenpetru @psomkarbuilds If you think the code quality of GPT-5.6 Sol or Opus 4.8 / Fable is the same as GPT-3.5, I honestly don't know what else to say.

@jturner_dev @psomkarbuilds No way, I'm using Claude on a max plan with settings to the max. You just can't deny hallucinations lmao.

I use don't code much anymore, a few solid skills and memories created and this isn't a problem much anymore. If you are using a free plan, or weaker self hosted LLM model its a problem. Use claude code on a max plan no hallucinations. I think alot of that came from the beginning or using cursor when it would switch models on you half way through dev.

@pcshipp 😇

@techdrivenpetru Those number stats not showing

I researched 20+ app paywalls. 99% of them look almost identical. But one app stood out. Their pricing strategy is a little different, and they're making 4x more revenue than the others. They don't offer a monthly plan. Weekly: $12 with a 7-day free trial Yearly: $59 with no trial It's a paywall designed around psychology. If users choose the weekly plan, they end up paying much more over time. If they choose the yearly plan, the app gets the money upfront and let's be honest, most people won't use the app for an entire year. They'll probably use it for 20–30 days. What do you think about this strategy?

@thatguybg 99% of software may use the same elements. 99% of the bugs live in the interactions between them. How does the finite-element approach handle retries, partial failures, permissions, and state transitions without rebuilding the hard part?

my hot take is that 99% of software can be built with the same set of finite elements spending time/tokens building any of this from scratch over and over again is peak AI psychosis