Kacper Kazimierak (Casper)

Systems engineer & SWE working on distributed systems for the last 10y 🛠️
Fascinated by all kinds of engineering (HW & SW)

[Kacper Kazimierak (Casper) @TheByteRacoon]

let me flex a bit, because I'm amazed how it turned out.

I use to travel quite a lot, and have Tailscale set up on the Pi 4b, which runs my Home Assistant

but finally, I decided to move tailscale to a dedicated server in my homelab

speed improvement is... I guess let's call it noticeable 😂 I'm limited by the wifi speed now.

what's your internet speed guys?

[Kacper Kazimierak (Casper) @TheByteRacoon]

@riskirills @ThePrimeagen ai is not a human (I know, shocking right?) its not learning from mistakes, or improving after the training phase

so if it doesn’t write good code from the start - it’s crap

[Kacper Kazimierak (Casper) @TheByteRacoon]

@ryancarson @thecsguy No it’s not. Unsubsidised ai cost is so huge that more and more often it doesn’t make financial sense to use agents for everything.

Its circling back to an assistant for skilled people making them even more skilled

[Kacper Kazimierak (Casper) @TheByteRacoon]

@sporadica maybe, but I moved all my money and crypto from Coinbase

you know… just in case

[Kacper Kazimierak (Casper) @TheByteRacoon]

@ThePrimeagen it’s not reasonable

we are literally talking about crypto 😅

[Kacper Kazimierak (Casper) @TheByteRacoon]

@deedydas interesting, because models should be trained on those codebases

problem lies with the context window, performance grafually fail as the context is filled

[Kacper Kazimierak (Casper) @TheByteRacoon]

did you know that in Go you can declare an int value with underscores? 🤯

compiler will just ignore them

it's something I just learned...
after a few years of coding in Go

[Kacper Kazimierak (Casper) @TheByteRacoon]

it may explain some of their incodents, (but still the load increase is not so much given the 2 years timeframe

every major company launching big product knows how to cope with that

what those numbers don’t justify are the non load related incident like the last merge queue issue

[Kacper Kazimierak (Casper) @TheByteRacoon]

@CDay152 @AdamSinger yeah, well... don't start a new war every now and then and the defense spending will drop

we learned that in Europe almost 80 years ago

[Kacper Kazimierak (Casper) @TheByteRacoon]

@big_duca this post is soo old, and you guys stil copyt paste is all the time

it was funny and smart when read first few times

[Kacper Kazimierak (Casper) @TheByteRacoon]

@techbromemes for real bro 😅

[Kacper Kazimierak (Casper) @TheByteRacoon]

I don't know how AI is supposed to get so advanced

we are at the limits of the transformers architecture (that is behind current LLMs), and no clear solution on the horizon

new models cost billions to train already, for few % gains at most

it might become cheaper, but still there is a limit of “intelligence”

[Kacper Kazimierak (Casper) @TheByteRacoon]

@signulll only in the b2c space, in b2b and large enterprise in general mcp are on top now

those companies cannot afford to let LLM loose and hope it’ll be ok

to much is at stake there

i don't talk about slop companies like MS, but the ones outside of pure software business

[Kacper Kazimierak (Casper) @TheByteRacoon]

@Lintilla369 @asaio87 I completely disagree, not unless there is a huge leap in current models

[Kacper Kazimierak (Casper) @TheByteRacoon]

@ItsCuthulhu @asaio87 you can't do it properly in 10 prompts, not even close in like 95+% of cases

maybe only if you are talking about some tiny app with 1-2 basic functions

but even then, it takes much more than some js to run reliably and securely

[Kacper Kazimierak (Casper) @TheByteRacoon]

@Lintilla369 @asaio87 but, devs can learn this too, and with much more head start (and solid foundations bc of experience) I’m not sure devs will be outpaced...

[Kacper Kazimierak (Casper) @TheByteRacoon]

@Surendar__05 price

[Kacper Kazimierak (Casper) @TheByteRacoon]

@hiarun02 I did, and so far I’m happy

cc limits are becoming ridiculous eecently, compared to codex where I still didn't managed to hit limits on plus plan

I know that openai heavily subsidies costs, which anthropic is slowly doing less and less... but as a customer should I care?

[Kacper Kazimierak (Casper) @TheByteRacoon]

@S_N_SH_E_ simple thing, if the basic tools are the same, then any additional knowledge amd experience becomes valuable asses

those guys yapping that “it's better to have no knowledge to AI” are just bunch of nonsense trying hard to be important

[Kacper Kazimierak (Casper) @TheByteRacoon]

even tho most of my posts and comments are AI skeptic, I still don’t get this guy

choosing career path around tech, and then complaining it changes fast is a Karen move... just nitpicking

it can be demotivating when AI can write code, even sometimes with acceptable quality - yes, but on the other hand, it’s like writing in Python vs Assembly

you can do in 5 minutes something that used to take weeks to learn just to get started

the wrong problem to focus on is fighting with AI tools on every line of code, it's literally like becoming micromanaging boss that most of us made fun of at one point or another

better problem to focus on is to treat AI like a bunch of juniors, they need supervision, examples, explaining things, etc.

and those years of experience are huge boost here, everyone can vibe code a product now, but to build one that is scalable, don't burn a pile of money on infra, so it doesn't have a single point of failur, or dependant on a single api that can change overnight - this still takes experience and knowledge, AI is terrible at it