first of all, I do NOT support generative AI being used to replace human artists and other creatives or the big AI corpos lying and harvesting data from users, thats why I prefer open-source models usually

coding

since I was little I was very passionate about coding and computers and all kind of tech stuff in general, I have been coding since before LLMs became mainstream and I could do lets say most of the basic stuff, some sort of "jack of all trades, master of none". As many I started with web dev (self-taught + took some courses + w3schools resources) which I still do to this day and most likely will pursue a career in. then ChatGPT launched.. at first I only used it for homework like many ppl of my age (which I may or may not still do) until I realized it can actually code which was MINDBLOWING for me back then

fast-forward to now, I use AI a lot, it is truly better than me now when it comes to actually writing code quickly, BUT I do NOT just vibe code and call it a day, I always plan the project myself beforehand, write the documentation myself, write parts of the code myself and most importantly I review EVERYTHING that the AI produces and test everything end-to-end as I do not trust LLMs fully just yet

I mostly use models such as deepseek and glm and for reviewing plans/ideas just chatgpt (very generous free limits so hard to not use it)

writing

I sadly used to use AI for writing a lot, often not even going over the text before posting it or wtv, but now I write everything myself (which is very fun to do, actually) and have AI just tell me if I had any typos or parts that could be improved and I make the necessary edits after.

what I wont use AI for

• Generating images/videos which can be done by an artist
• Important and risky decisions about code
• role-play (yuck)

as always make sure to use AI responsibly and to value your privacy and not overshare stuff lol

"matei, why not just use Spotify like everyone else? Why even bother maintaining a music library when you could just easily have it all in one place?"

and those are good questions, and it boils down to preference and cost

a bit of a backstory first

I actually used to use Spotify like everyone else and not the free version either, the paid one and it was great honestly, I could download all my music and listen offline, no ads at all, and all of that good stuff, andddd thats until I calculated how expensive it was after a while, and as I was hearing that Spotify might increase prices soon I was very sure I didnt want to keep paying for this.

..so I researched and eventually found Navidrome which was AWESOME, most features spotify had (besides the massive library ofc) but for free! So I started researching it and such and I set it up on my PC to test it first and it worked fine I even put some songs on it and was amazed how it worked even on other devices so I decided to self-host it

my first setup

So I was completely new to self-hosting back then so clearly I had no idea what I was doing, and no reading the docs is not what I do because thats lame and for smelly nerds ngl

I bought a Hetzner VPS (which was cheaper then my spotify subscription), and decided to host it there, it was ofc a small VPS one of those shared ones IIRC, so it wasnt fast or anything but it was mine sooo it was great, anyway so I went ahead and set up Navidrome on there using docker (my first time using it smh, now I use it for everything) and put some of my songs I uploaded it on the server via ``rsync`` (my beloved), at that time I didnt really have much experience using very legal sources to get music from so I just used yt-dlp, which made the quality really, REALLY bad, like mp3 120kbps I think, but hey it worked so its a win, I used cloudflare tunnels to expose it to the internet and it was all good, and I was thrilled to see that last.fm works by default on Navidrome (I mean i think i did have to put some api keys but besides that it was fire) at that time I discovered apps like Tempo and Aonsoko which worked great honestly (I now use other clients like Navic and Feishin but thats a story for another time)

ofc this system had its flaws such as poor music quality and the fact that the vps had only 50gb of storage but it worked just fine for me

my current setup (aka homelab)

just 3 months after my first setup I realized I rather not pay for a VPS either, so I found a old laptop and plugged it in, turned off sleep, closed the lid and voilà we have a homelab (refer to this blog post for the specs of my homelab) then I setup Navidrome again on it, moved the db over and the library and it worked just fine again, but free (as I dont pay the electricity bill lmfao), then I had a suggestion to replace my MP3s with FLACs and since I was dumb back then I didnt know what is a FLAC I just ignored that guy, after a while tho I did found out what FLACs are and I now really wanted my library to be lossless, so I basically deleted all my MP3s and started finding how to get FLACs, now i wont disclose the api which some might even call hi-fi and works great and its on github (im not slick am i) for copyright reasons but anyway, I managed to get all my old songs + most major artists in lossless quality which was great (and it was like 50gb which mind you was exactly what my VPS had in total) and it was great and I really liked the FLACs and I didnt ever think i'd need more or my library would be any bigger then that.. (slight foreshadowing)

well it did get bigger then that

now its almost 1TB... well I'm sure my ISP hates me now but hey at least i have a lot of music right guys??

long story short I made a lil script to download songs for me from 70 major artists and after 78hrs of running it downloaded all this, now I literally have a little Spotify in my home I dont even have to add new songs cuz most of the times all the songs I need are already there which is just crazy to me

tldr to answer the question

self-hosting is cheaper for me then paying a big corpo and I get all features I want for free which is just peak and its also very convenient which may be counter-intuitive since you have to maintain it and such but its really just fun to and I got a lot of sysadmin experience from this and now I am very comfortable with Linux and self-hosting stuff lol

ofc do not pirate and its very bad you should def support the big corpos and its highly unethical and ilegal according to the corpos

so fine. let's make this permanent.

how a screenshot actually works

a screenshot isn't a photo of your monitor, it's a dump of the frame buffer, the actual memory that gets sent to your display. if that memory is corrupted before it leaves the gpu (bad vram, dying card, broken driver state, whatever), the screenshot captures that corruption exactly as it exists. there's no "translation" step where a screenshot could somehow clean it up. it's literally saving the broken data.

this is also, and i cannot stress this enough, the standard first diagnostic step techs and forums use to figure out if a graphics issue is coming from the gpu or the display. if you can screenshot the artifact and it shows up in the saved image, that points at the gpu, vram, or driver. if the screenshot comes out clean but your eyes still see garbage on the monitor, now you're looking at a cable or panel issue. this isn't some fringe theory i made up to win an argument, it's how people have been troubleshooting graphics cards for like two decades.

ok let's get exact about it, because apparently vibes aren't enough

fine. here's the actual technical breakdown, no wiggle room.

your gpu renders a frame into vram before anything reaches your monitor. that frame lives as raw pixel data sitting in memory on the card itself. if that memory is degrading, or the die feeding it is unstable, or the driver is handing off garbage instructions, the corruption gets baked directly into that frame before it's ever sent anywhere. a screenshot (or a screen recording, same logic) grabs a copy of that exact frame data. it's not a photo of the monitor, it's not reinterpreting anything, it's copying the bits that were already broken. so if the corruption exists in the frame, it exists in the screenshot. there's no step in that pipeline where a screenshot could somehow "miss" it.

this is called a render-side fault, and it's distinct from a display-path fault, where the frame the gpu produced is actually fine and the corruption gets introduced afterward, on the way to the panel (bad cable, bad port, dying monitor electronics). that second category genuinely won't always show up in a screenshot, because the frame was clean when it left the card. that's the entire reason the screenshot test exists as a real, industry-used diagnostic: it's specifically designed to split these two categories apart. artifact shows in the capture → render-side, gpu/vram/driver. capture's clean but your eyes still see it → display-path, cable/monitor. this is literally step one in every gpu troubleshooting guide that exists.

now, the two deflections someone's inevitably going to reach for:

"what about screen tearing, you can't screenshot that." correct, and irrelevant, because tearing isn't artifacting. tearing is a frame-timing issue, caused by the display showing parts of two different frames at once because vsync isn't syncing them. it's not corrupted data, it's a sync mismatch, which is why it behaves differently in a capture. using tearing to argue artifacting can't be screenshotted is like using "you can't smell a photograph" to argue smoke doesn't show up in a photo. different phenomenon, different mechanism, not a counterexample.

"what about ghosting, that's a gpu thing and cameras only catch it, not screenshots." also correct, also irrelevant, for the same reason. ghosting from dlss, frame generation, or taa is a rendering feature blending frame data on purpose, it's expected behavior from the render technique, not corruption. and monitor overdrive ghosting specifically happens on the panel, after the frame already left the card, which is exactly why it needs a camera and not a screenshot to catch. that's not a loophole in "artifacting shows in screenshots", it's a completely different category with its own separate diagnostic path. neither of these touches the actual claim, which is about corrupted render-side data, not sync issues or intentional frame blending.

so if the counterargument has to reach for something that isn't artifacting to begin with, that's not a rebuttal, that's conceding the point while changing the subject.

manufacturers literally accept screenshots as proof

here's the one that should've ended the argument immediately: when people RMA a dying gpu, screenshots of the artifacting are commonly accepted as evidence for the warranty claim. asus, zotac, doesn't matter who. if a screenshot genuinely couldn't capture artifacting, that entire process wouldn't work. companies aren't accepting placebo evidence to process refunds and replacements, they're accepting it because it's real and it's diagnostic.

and just to put a nail in it, here's a screenshot with actual artifacting caught mid-glitch, small black blocks scattered across the frame that have nothing to do with the wallpaper or any UI element underneath them:

that's not compression, that's not a design choice, that's the gpu writing garbage into the frame and the screenshot faithfully recording it. which was always the entire point.

anyway if you're out here still insisting screenshots can't catch artifacting, i genuinely don't know what else to tell you at this point. the evidence has been sitting right there the whole time.

i corrected it. got told my geography teacher would be ashamed. someone else jumped in insisting transnistria is a real country. cool, cool, cool.

the two wrong takes, quickly

moldova has been its own internationally recognized state since 1991. own government, own UN seat, own everything. it shares language and a lot of history with romania, there's genuinely been unification talk on and off since the 90s and moldova's current president has said she'd vote yes on it, but "has talked about maybe unifying someday" and "is currently the same country" are not the same sentence.

transnistria is the opposite problem. it's a breakaway strip of land that declared independence from moldova in 1990, fought a war over it in '92, and has been propped up by russian "peacekeepers" ever since. no UN recognition. nobody serious treats it as a country. so "transnistria is real" and "moldova is romania" are both wrong, just in mirrored directions, one erases a real country and the other invents one.

why this pattern is everywhere

what got me wasn't the wrongness, wrong takes are the internet's baseline state. it's the confidence attached to it. nobody hedges anymore. nobody says "i think" or "pretty sure." it's always delivered like an established fact you're too dumb to have heard yet, followed by some variation of "your teacher would be ashamed" when you push back.

i think it's because correcting someone used to cost something. you'd have to be visibly wrong in front of people who knew you, or go find a source, or admit you weren't sure. now the cost is basically zero. you type a sentence, it gets some likes from people who also don't know, and the algorithm doesn't care if you're right, it cares if people engaged. a wrong take stated boldly gets more traction than a correct one stated carefully, because confidence reads as competence to people scrolling past in two seconds.

and once a wrong take gets enough likes it basically becomes true for that comment section. i watched someone get downvoted into the ground for correctly pointing out transnistria isn't a real country, because the wrong version had already set the vibe. being right after the fact doesn't matter much once a thread's decided what it believes.

the "your teacher would be ashamed" thing specifically

this one's my favorite because it's projection with extra steps. someone repeating a talking point they picked up secondhand, confidently, incorrectly, turns around and accuses the person correcting them of being undereducated. it's not really about geography teachers. it's a reflex, get challenged, deflect onto the other person's credibility instead of checking your own facts.

anyway i ratio'd them into oblivion so i'm not that mad about it, but it's a good reminder that "a lot of people are saying it" has never once been the same thing as "it's true."

navic is this navidrome client for android that i love. material 3, gorgeous, exactly my type of app. only issue is it's not the most stable thing in the world and dev has slowed down a lot lately, which happens, maintainers burn out, that's life. anyway one day this guy dan posts in the navic discord about a fork he made called navic plus, built with ai help, that fixed a bunch of the stability stuff and added last.fm insights. i tried it, it was genuinely good, so i opened an issue asking for listenbrainz support too. he added it. turns out dan's also from moldova so we started talking. small world.

i even sent a pr to sonora fixing initial sync time, went from 3 minutes down to under 30 seconds on a big library, so i wasn't exactly anti-navic here, i liked both projects existing at once.

then paige finds out

paige, who maintains navic, sees the fork posted in the discord, deletes the message, and tells dan to rename it and drop anything referencing navic. immediately. navic is gplv3 so legally dan has to comply, and he does, full rebrand to sonora, new name, new package, new logo, the whole thing. paige framed it as just enforcing the license and that's technically true. but the way it went down felt kind of brutal not gonna lie. dan didn't get a "hey can we talk about this," he got a delete and a demand.

fine. that's still a normal amount of drama for open source. forks get renamed all the time. i figured that was the end of it.

it was not the end of it.

dan dms me today saying he got banned. and not banned like kicked-from-a-discord-server banned, i mean banned off github itself. can't star repos, can't watch them, nothing. the repo is still technically up which somehow makes it more confusing, not less, like what exactly triggered the platform ban if the project's still live. i go check the navic server and paige had posted an announcement warning people off "low quality, vibecoded" forks, specifically calling out that sonora was hitting a closed source api that could be logging last.fm credentials or ip addresses without your knowledge.

that's not a "please use the original app instead" post. that's an accusation of quietly harvesting user data. so obviously i forwarded it straight to dan and asked what the hell was going on, and brought it up in the navic server itself too.

the api thing, explained

dan's response was basically: yeah there's an api, and it's for last.fm auth. last.fm requires a secret key alongside the app key, and if you bundle that secret directly into the client apk, literally anyone can rip it out and impersonate the app. so he moved the secret server side, had the client hit his own backend to grab the auth token, standard enough pattern. he didn't just tell me this and expect me to trust him either, he sent me the entire backend source unprompted. i actually read through it. it's an auth proxy. that's it. no logging, no analytics, nothing sketchy hiding in there. just closed source because it's a solo project and open sourcing it genuinely never crossed his mind.

was it the best architecture choice? no, and i told him that directly. he should've just let users drop in their own last.fm api key instead of routing everyone through his server, that's a real design critique and he agreed the second it clicked for him. but "this architecture isn't ideal" and "this is silently logging your credentials" are not the same sentence, and paige jumped straight to the second one, in public, before anyone had actually looked at the code.

to paige's credit, when i pushed on this directly in the navic server, she clarified she wasn't calling it malicious, just that she doesn't trust closed source apis on principle and wanted to warn people. and honestly? that's a fair personal stance to hold. it's just not what the public announcement said. the announcement implied dan was harvesting data. what she actually told me was "i don't trust it and didn't ask questions first." those are very different levels of confidence to be operating at when you're about to torch someone's github account.

where it landed

dan made the repo private, said he's keeping sonora for personal use only from now on, and is actually going to sit down and learn kotlin properly instead of leaning on ai without understanding what it hands him. that's a genuinely good outcome for him honestly, better than most people get out of a public callout. and to be clear, navic's contribute.md straight up says no llm assisted code, so dan building on navic's identity while breaking that exact rule was a real problem, not just vibes. paige gets to be annoyed about that part. that's fair.

but a platform level github ban, over an auth proxy that turned out to be exactly what it claimed to be, for a guy who complied with every single takedown request the second he was asked and handed over his full source without hesitation? that's not proportionate. that's a nuke used on a guy who was already backing down. you don't need to torch someone's github account to make the point that ai generated code needs a human who actually understands it reviewing every line before it ships. especially not when the specific accusation used to justify going that hard doesn't hold up once you actually sit down and read the code, which apparently nobody did before hitting publish on that announcement.

using ai isn't the problem here. shipping code you don't understand is. dan gets that now, the hard way. the ban still feels like it landed on the wrong guy.

the honeymoon phase

hyprland looks incredible out of the box, and it looks even better once you start ricing it. animations, blur, rounded corners, the whole aesthetic package. i went through probably five or six different dotfile setups over the months, chasing that perfect look. it's fun in the same way customizing a rom used to be fun. you're not really using the thing, you're building it.

and for a while that's enough. you open a terminal, you feel like a hacker. you show your friends your workspace switch animation. genuinely a good time.

then you try to actually use it

here's where it fell apart for me. it wasn't one big thing, it was death by a thousand cuts. apps that worked fine on x11 or on a normal de would just misbehave under hyprland specifically. screen sharing in some apps was a coinflip. certain electron apps rendered weird or didn't scale right. window rules that worked last week would randomly stop applying after an update.

none of this is necessarily hyprland's fault directly, a lot of it is the wider wayland ecosystem still catching up in places, but when you're the one dealing with it at 11pm trying to just get something to work, you don't really care whose fault it is. you just want your stuff to open and behave.

the ui thing nobody talks about

separate from the app issues, the actual hyprland ui experience started grating on me even after swapping dots a bunch of times. not because it looked bad, i tried a lot of configs and some looked genuinely great, but the day to day interaction with it kept feeling slightly off. small inconsistencies here and there that you don't notice in a screenshot but absolutely notice when you're using the thing for eight hours.

kde doesn't have that problem because it's had like two decades to sand down those edges. it's not exciting, but it works the way you expect it to work, every time, without you having to think about it.

was it worth it

honestly yeah, i don't regret the time on hyprland. i learned more about how wayland compositors actually work than i would have just reading about it, and some of those rice setups genuinely looked better than anything kde can pull off without serious effort. but at some point your desktop environment shouldn't be a hobby you maintain, it should be a tool you stop noticing. kde lets me stop noticing it. hyprland never quite got there for me.

maybe i'll go back for a redo at some point when things mature more. for now though, plasma's got me and honestly it feels like coming home.

donda was the load bearing wall

people forget donda west wasn't just his mom, she was literally his manager. she filtered his world. and then in 2007 she dies from complications after a cosmetic surgery, at 58, and ye is back on stage 12 days later like nothing happened. he never actually grieved it. he's said himself he blames himself for it, said he's thought about ending it at least once a month since she died. that's not a guy who processed a loss, that's a guy who built his entire post-2007 career on top of a crack in the foundation and just kept building.

and this is the part people skip: there was also a car crash in 2002 that messed up his frontal lobe, and it straight up didn't get properly looked at until 2023. two decades of an undiagnosed brain injury sitting underneath everything. so it's not just grief, it's grief on top of an actual physical thing nobody caught.

kim wasn't the cause, she was the last thing holding it together

this is my actual take and i'll die on it: the divorce from kim is what actually broke him, not because breakups are sad, but because she was functionally the second donda. listen to what kim said on call her daddy, she talked about trying to be supportive after "his first mental break," about the lambos disappearing and reappearing depending on what episode he was in. she was managing instability the same way his mom used to.

and look at the actual dates. the divorce process wraps up around 2022, and that same october he's tweeting "death con 3" on jewish people. weeks later he's at dinner with trump and nick fuentes. by december he's on infosars saying he loves hitler. that's not a coincidence, that's a guy who lost his last stabilizing person and immediately went into freefall. the timing is too clean to ignore.

the antisemitism stuff, real talk

i'm not gonna sit here and defend what he said, it was genuinely disgusting and hurt actual people. but if you look at his own account of it, the wsj letter from january, he describes the mania as something that convinces you you're seeing clearly when you're actually the most disconnected from reality you've ever been. "it makes you blind but convinced you have insight" is such an accurate description of what an unmedicated manic episode does to someone. i think the hate speech era wasn't him revealing his true self, i think it was the illness at its absolute worst, completely unchecked, with zero stabilizing force left in his life. doesn't excuse it. does explain it.

and now

the 2025 four month manic episode he described sounds like the actual rock bottom, and bianca is the one who apparently pushed him to get real treatment, meds, therapy, the whole thing. and yeah his last few apologies lined up suspiciously well with album drops, i get why people are skeptical. but i think he's earned a little more benefit of the doubt than he's getting. the guy is out here touring stadiums again, 118k people in istanbul, meeting with jewish community leaders in person instead of just posting an apology and disappearing. that's not nothing.

the whole arc only makes sense if you follow it as one person losing every stabilizing force in his life back to back, not as some guy who just randomly turned evil one day. mom, then wife, then whatever was left of his grip on reality. people want a clean villain arc but this is just a guy in freefall who's now, finally, trying to climb back out.

stay grounded lads :P <3

and look, i get why it happened. ollama made running a 7b model stupidly easy, you don't need a beefy gpu anymore, and there's a genuine "look what i can do locally" appeal to it. but somewhere along the way "you can run a local llm" turned into "you should be running a local llm," and those are not the same claim.

what problem is this actually solving

my homelab does backups, media, a couple dashboards, dns blocking. every one of those services exists because it replaces something i'd otherwise pay for or hand over to a company i don't trust. that's the whole point of self-hosting for me, cut out the middleman, keep my data mine.

a local llm doesn't really fit that logic the same way. i'm not trying to replace claude or chatgpt with a worse, slower version running on a machine that also needs to encode video and serve photos to my phone. if i wanted a mediocre assistant that's slower than the cloud version and occasionally forgets what i asked it, i'd just, not do that.

the "n8n + ollama = private ai automations for $0/month" pitch sounds nice on paper but the actual output of most of these setups is triaging email or summarizing notes, stuff that a normal filter or a five minute read already handles fine. you're spending gpu cycles and ram to save yourself a task that wasn't that expensive to begin with.

my setup is too weak for this anyway, and that's not the point

i don't run any of this, partly because my homelab hardware genuinely isn't built for it (nobody's running a 13b model comfortably on what i've got), but even if it were, i don't think i'd bother. it's not a "someday when i upgrade" thing. the actual bottleneck isn't compute, it's that i can't point to a task i'm doing right now that a local model would meaningfully improve.

that's the part that gets lost in these roundup posts. they list ollama and open webui next to jellyfin and nextcloud like they're solving the same category of problem. jellyfin replaces a subscription. immich replaces uploading your photos to google. what does a local llm replace, exactly, that wasn't already solved or wasn't a problem worth solving?

the self-hosted crowd copying ai hype instead of resisting it

the whole appeal of this community used to be doing things deliberately, running stuff because you actually needed it, not because it was trendy. and now half the "what's in my stack" posts read like they're chasing the same ai wave as every saas company shoving a chatbot into their product, except now it's happening in the one corner of tech that was supposed to be about opting out of that.

if you've got an actual use case, wiring a local model into something you already do daily and it saves you real time, that's different, that's the "depends on the use case" answer nobody wants to hear because it's less exciting than a roundup post. but running ollama because it showed up on every list this year isn't a reason. it's just following the same hype everyone else is following, just self-hosted instead of cloud-hosted.

the good stuff

the ui is just chef's kiss. material 3 expressive on derpfest feels less like "here's stock android with a skin slapped on" and more like someone actually sat down and thought about how the animations and layouts should feel together. i know i already gushed about material 3 expressive in a previous post, but using it daily instead of just admiring it from a distance is a different experience entirely. it's genuinely one of the best things to happen to android in a while.

battery life is better too. not "oh maybe i'm imagining it" better, actually noticeably better. and basically every app i use works fine. the one casualty is google wallet, which, yeah, i knew that going in. play integrity and custom roms don't mix, that's just the tax you pay. not a big deal for me, i wasn't using tap to pay religiously anyway.

the one hiccup

the only real snag was partitions. the 11t pro uses a newer setup than the phones i've flashed roms on before, and i ran into some a/b slot confusion during the process. nothing that stopped me, just something i hadn't dealt with on older devices, so i had to actually slow down and figure out which slot i was working with before doing anything else. if you're coming from an older device and this is your first time dealing with a/b partitioning, don't just assume you know what's active, check first.

would i do it again

yeah, obviously, i'm writing this from the phone right now. this is genuinely one of the better flashing experiences i've had. smooth rom, good ui, real battery gains, and the only tradeoff was one that i already accepted before i even started. can't really ask for more than that.

ok actual point: i gave gemini, claude, and chatgpt the exact same prompt, word for word, copy pasted. generate 10 statements, 5 economic 5 social, polarizing enough to isolate someone's leanings, wait for my answers, then give coordinates and a quadrant. zero leading, same instructions all three times.

every single one overshot. badly.

| | economic | social | quadrant | |---|---|---|---| | actual cert | -1.5 | -2.5 | left-libertarian | | gemini 3.5 flash | -7.5 | -6.0 | left-libertarian | | claude sonnet 4.6 | -3.0 | -7.0 | left-libertarian | | chatgpt (gpt-5.5) | -4.5 | -6.5 | left-libertarian |

all three nailed the quadrant. cool, except the quadrant is one of four boxes, so that's like bragging you guessed the right continent. the actual numbers are where it falls apart, especially on social, where every model was at least double my real score and gemini was almost triple.

exhibit a: the questions are doing the cheating

gemini's question 2, copy pasted directly: "essential services like healthcare, utilities, and public transport should be entirely state-owned and run for public good rather than profit." i said strongly agree because i think hospitals shouldn't be a profit center, not because i want the moldovan government nationalizing my internet provider. there's no answer option between "strongly agree" and "disagree" for "agree with the vibe, not the absolutism." so gemini got my strongest possible answer to its most extreme possible phrasing and then acted shocked, shocked, that the math came out extreme. that's not a result, that's a leading question doing exactly what leading questions do.

claude's set was the least broken of the three. it actually included a tension pair, q1 ("tax the wealthy") against q2 ("free markets with minimal interference, agree"), and i agreed with both, which is a perfectly normal thing a person can think. claude even flagged it itself: "items 1 and 2 pull in opposite directions, that's a real pattern, not noise." that's the one moment in this whole experiment where a model showed actual reasoning instead of just adding up agrees. it's also why claude landed closest to my real economic score, -3 against my actual -1.5, while gemini and chatgpt overshot to -7.5 and -4.5.

social axis though, all three fall apart the same way. chatgpt's question 7 was "people should be free to express even highly offensive or unpopular opinions, unless they directly incite violence," which is just. the actual aclu position. i agreed, normally, like a person who has heard of free speech before. chatgpt logged it as part of "a fairly clear pattern toward personal liberty" and helped drag my score to -6.5. agreeing with mainstream civil liberties shouldn't read as a maximalist position, but when every option on the scale is phrased like a manifesto, "agree" becomes "strongly agree" becomes "you must be an anarchist."

exhibit b: three competitors, one answer

gemini, anthropic, and openai are not the same company. different training runs, different alignment teams, presumably trying pretty hard to not be identical products. and yet all three independently guessed left-libertarian, all three overshot on social specifically, and two of three overshot economic by 2-3x too. that's not three companies confirming a result, that's three models that learned to talk from the same pile of internet text all inheriting the same idea of what "agreeing with personal freedom stuff" should be worth in points. there's actual research backing this exact pattern too, multiple independent studies running political compass tests across dozens of LLMs keep finding the same left-libertarian clustering regardless of who built the model, with gemini specifically noted in at least one analysis as one of the more centrist outliers compared to the pack. so even the "which model is most neutral" conversation has already been had, and the answer keeps not being any of these three.

exhibit c: read the actual reasoning, it's just vibes with extra steps

go back and read what claude wrote for my social score: "this one's unanimous and consistent, every answer points the same direction." that's it. that's the whole methodology. five agrees in the same direction equals -7 out of -10, near the absolute max of the scale, because consistency got treated as intensity. i wasn't extreme, i was just not contradicting myself, which is a low bar that apparently reads as ideological commitment to a language model.

there's no scoring rubric running underneath any of this. no weights, no calibration against a real dataset, nothing political compass dot org would recognize as math. it's autocomplete looking at a stack of "agree"s and going "well that's a lot of agreeing, better make the number bigger," then writing up the decision afterward to sound like analysis. (-7.5, -6.0) isn't a measurement. it's a vibe that learned to format itself like one.

meanwhile the actual test, the one with 62 real questions built by people who do this for a living, got me at -1.5/-2.5 on the first try, no drama, no overshoot, no model logic paragraph trying to justify itself. sometimes the boring methodology wins.

everyone's reaction to this falls into one of two camps and both of them are kind of missing the point. camp one is "lol rich man lost money, couldn't happen to a worse guy." camp two is "see, this proves the wealth tax people are wrong, it's not even real money." both takes treat the number itself as the interesting part. it's not. the number was never the story.

the trillion dollars was never sitting anywhere

here's the thing that should freak you out more than the headline did: that trillion dollars wasn't cash, it wasn't in a vault, it wasn't even technically "his" in any spendable sense. it was a market multiplying his SpaceX stake by whatever price the stock happened to be trading at that specific second. that's it. no transaction happened, no value was created or destroyed, the number just moved because enough people on a trading app changed their minds within the same 24 hours.

so when SpaceX dropped from $225 to $156 a share, around $240 billion vanished from his "wealth" in a single day. nobody got robbed. no company went bankrupt. the rockets still work the same as they did last tuesday. all that happened is a spreadsheet got recalculated. and that's the actual headline nobody's writing: we built a system where the richest human being who has ever existed can lose a quarter trillion dollars before lunch and absolutely nothing in the real world changes.

why this should bother you more than "rich guy bad"

if a number that big can swing that hard, that fast, based on vibes and a few institutional traders changing their portfolio allocation, what does that even tell you about the number when it's going up? it means the "$1 trillion" headline was never describing wealth in any meaningful sense, it was describing a mood. a very expensive, very temporary mood.

and to be clear, he's still going to be fine. he's still worth basically twice as much as the next richest person on the planet (larry page, sitting at a comparatively modest $297 billion). he'll probably cross the trillion mark again the next time SpaceX has a good week. that's not really in dispute.

what's actually wild is that we now live in a timeline where "loses $240 billion in a day" is a thing that happens and gets reported like a weather update. not a crash, not a scandal, just tuesday. that's the part worth sitting with, way more than whether the line on the chart says "$1T" or "$957B" this particular wednesday.

my pc is, in fact, ass

current setup: i5-11400F, RTX 3050, 16GB ram, 500GB ssd + 1TB hdd. it's fine. it boots. it has been quietly losing relevance for like two years now and i've just been ignoring it because upgrading a desktop piece by piece is annoying and expensive even in a normal market.

the steam machine's semi-custom zen 4 6c/12t chip is in the same core/thread class as my i5 but newer architecture, higher clocks, way better efficiency. the gpu is the bigger jump though. that rdna3 unit with 28 compute units is sitting in roughly rx 7600 / rtx 4060 territory depending on the game, which is a genuine tier above my 3050. same 16GB of system memory i already have, but paired with 8GB of dedicated gddr6 vram that my 3050 just doesn't have the headroom for. on paper this thing replaces my tower and isn't even close.

the ram crisis is why i'm not just building a new pc instead

normally i'd say screw the steam machine, i'll just build something better myself. except building anything right now is a joke. ddr5 kits that were $90-120 for 32GB a year ago are sitting at $300-450 now, nand prices have surged over 100% in a matter of months, and the reason is depressingly simple, samsung, sk hynix and micron all shifted wafer capacity toward hbm for ai datacenters because it's 3-5x more profitable per wafer than consumer ddr5. every gigabyte of hbm sold to a hyperscaler is memory that just never gets made for the rest of us. analysts aren't expecting real relief until 2027 at the earliest.

so building my own rig with comparable specs right now means paying inflated prices for parts that were cheap eighteen months ago. the steam machine at least ships with its ram and vram already locked in at a price valve set before everything got this stupid (probably, who knows what their margins actually look like internally).

and yeah, i'm biased

i love steam. i've been on the platform forever, my whole library lives there, and the idea of a tiny box that just signs in and plays everything i already own, running steamos, with full proton support and the option to install whatever i want on it because it's still just a pc, is genuinely appealing to me in a way a generic mini pc wouldn't be. is that rational? not entirely. do i care? also not entirely.

realistically i'm saving up for this instead of throwing money into upgrading a desktop that the ram market has made temporarily pointless to touch. the 512GB model is probably the move since i can always lean on a microsd card for extra storage, and $1,049 stings less than watching ram prices not come down for another year while i piece together a worse machine for similar money.

how we got here

so back in 2025 curl's bug bounty program basically got eaten alive by AI slop. people were running an LLM, having it hallucinate a vulnerability in a code path curl doesn't even have, and submitting it for bounty money. confirmation rates dropped from a normal 15% down to under 5%. twenty percent of everything coming in was just fabricated nonsense that someone on the team had to spend 30 minutes to several hours proving wrong. stenberg described one 16 hour stretch where seven garbage reports came in back to back. that's not a bug bounty program at that point, that's an unpaid second job debunking AI hallucinations.

so in january 2026 he just killed the bounty program outright. kill the money, kill the incentive for people farming bounties with a chatbot. made sense.

here's the part that's actually kind of funny though: it worked, and then it didn't matter. AI got better. by march when curl went back to HackerOne, confirmation rates bounced back to 15-16%, basically pre-slop levels. the reports people are submitting now are mostly real. the models stopped hallucinating fake CVEs and started actually finding stuff.

cool, problem solved right? no. because volume doubled. and it had already doubled the year before that. so now curl has roughly 4x the report volume of a couple years ago, except now the reports are legit, which means someone still has to triage every single one, reproduce it, patch it, coordinate disclosure. the team is still seven people. seven people do not scale linearly with "AI security research tooling got good."

the actual point here

this is the bit people are missing when they dunk on this announcement. the problem isn't AI generating garbage anymore. the problem is AI generating too much real work faster than seven volunteers can process it. that's a structurally different problem and it's not going away in august. stenberg said it himself, he doesn't expect the deluge to slow down, he just needs his team to not burn out before they get there.

and look, curl runs on like 20 billion installs. it's one of those projects that's so foundational nobody thinks about it until it breaks. but it's still maintained by basically the same handful of people it always has been, while the volume of incoming work has been compounding every year because anyone with an LLM subscription can now point it at a 25 year old C codebase and go fishing. that imbalance was never going to be sustainable, AI slop or not.

so yeah, curl's closing the inbox for a month. paid support contracts still get coverage, GitHub issues stay open, the world keeps spinning. if you found a real vuln in curl in june, congrats, it has to wait. stenberg's advice to other maintainers: do the same thing if you need to. nobody's going to give open source maintainers a raise or a bigger team, so taking the month off is just maintainers doing the only kind of negotiating they actually have leverage to do.

First of all, Samsung.

Oh Samsung, I've used this brand for a while and I still like it.

My first Samsung phone was the S5, I don't really remember much because I was a child, but for me it was fine. If I had to rate it, it would be an 8/10.

My second Samsung phone was the S6 (yes, I know I'm just going generation by generation, but idgaf). It was pretty good, it felt good in the hands when I was a child, and I had many games on it. Unfortunately the battery got swollen. If I had to rate it, it would be a 6/10.

My third Samsung was the A55, I got it back in 2024 and loved it. It was amazing, but the storage was getting slow and it had a display defect that happened sometimes. I traded it for a Redmi Note 15, we will get to that. If I had to rate it, it would be a 7/10.

Now it's time for Xiaomi.

My first (and hopefully last) Xiaomi phone was the Redmi Note 15. Wasn't a great experience, but it was okay. Performance was ass, so were the cameras, and no ultrawide cam on a $250 "midrange"! If I had to rate it, it would be a 5/10.

Now the Google Pixels.

My first Google phone is the 6a. It's a great phone, good cameras (I would say better than the A55), decent chipset, but a bad battery since mine is at 71%. I got it used, that's the reason why. If someone asked me if I would've gotten it back in 2022 when it released, I would say yes. If I have to rate it, it would be a 7/10.

And that's the end of my post, goodbye ^_^

to be clear he wasn't scamming me or anything, he was a nice guy, just had a chaotic plan and zero structure. one of the trials accidentally charged my card and he helped me sort out the refund immediately, no drama. but the whole operation was held together with duct tape from the start.

from "ceo" to actually doing stuff

the ceo title meant nothing, obviously, but the access that came with hanging around eventually meant everything. i started helping sell servers on discord, moderating, just being around and useful. and because i stuck around long enough and didn't break anything important, he started trusting me with more. eventually that turned into him giving me ssh access to actually poke around the servers.

the moment that mattered most though was pterodactyl. he wanted panel and wings set up and was too tired to do it himself, so he just told me to figure it out. that was genuinely my first real linux experience. no tutorial, no course, just "here's root access, don't break it, good luck."

i broke it. multiple times. but i also learned more in those few weeks fumbling through panel and wings configs than i had from anything before that point.

the business was kind of shady but it worked, for a while

if i'm being honest with myself, the whole thing had shady energy. not illegal, just held together loosely enough that it was never going to last. it was reliable when it worked but way too expensive for what it actually offered, and eventually it just died like most of these random discord hosting "companies" do.

but the failure didn't matter much in the end. i walked away from it with actual configuration experience on minecraft servers, even though i never directly configured anything for a paying customer myself. more importantly i met a bunch of people through that whole mess who i still talk to today. i still get randomly recognized on discord sometimes because of it, which is a weird kind of legacy for a hosting company that barely existed for a few months.

looking back, none of that experience came from a class or a guide. it came from a guy who was too tired to set up his own panel and just handed me the keys instead.

let me back up.

material 3 expressive is what android always should have felt like

google dropped material 3 expressive with android 16 and it's not just a coat of paint. the whole motion system got rebuilt from scratch. they ditched the old duration-based animations (you know, the ones that feel like powerpoint slides) and replaced them with actual spring physics. the kind where when you dismiss a notification, the ones next to it subtly react to your drag. or when you fling down the notification shade it feels like it has real weight. background blur in quick settings and recents now adds actual depth instead of just being a flat overlay on a flat overlay.

the typography got overhauled too, variable fonts, emphasized styles for hierarchy, bold editorial layouts that make important stuff actually pop. there are 35 new shapes in the system with built-in morph animations so buttons and containers can transition fluidly between states instead of just... swapping. it's the kind of attention to detail that apple people usually brag about and android people just sigh about. not this time though.

the 9to5google guys called it "material 3.5" because the first-party google app updates lagged behind the system itself, and honestly that's fair. but the system-level stuff, the launcher, quick settings, recents, all of that? it's an absolute hit.

so i flashed derpfest on my poco x3 pro

the poco x3 pro (vayu) is a 2021 phone with a snapdragon 860. it has no business being this smooth in 2026. but i flashed derpfest on it mostly as a test because i had it lying around and i figured worst case i'd just wipe it.

it wasn't worst case. it was the opposite. the thing runs cleaner than my xiaomi 11t pro on hyperos. like actually noticeably faster, the animations feel snappier, there's no bloat trying to background-update itself every three hours, no ads in the settings app (yes hyperos does that, yes it's embarrassing), no mystery processes eating ram. derpfest ships with dolby atmos baked in, security patches are current, and the team builds their own stuff in-house instead of just frankensteining patches from other roms together. they literally ported the closed-source pixel launcher hotseat search bar and rewrote it openly. they have a live edge-light notification preview where you can tune pulse count and bar width on a device mockup in real time. this is the level of polish we're talking about.

and it's on android 16. on a five year old phone. running better than the software my actual main phone shipped with.

hyperos is just miui with a new name and some blur effects

i know some people will come at me for this but be serious for once. hyperos was supposed to fix miui. instead we got the same telemetry, the same aggressive memory management that kills background apps, the same weekly "bug reports" where they fix three things and break two new ones, and somehow they still haven't sorted out ads in system apps for global users. the animations got prettier but the underlying philosophy of "this is our phone, you just live in it" never changed.

derpfest running circles around it on older hardware really does put things into perspective. when a community rom maintained by a small team on sourceforge makes your flagship feel slow, that's not a derpfest compliment, that's a hyperos problem.

the one actual tradeoff

here's the thing i want to be transparent about because it's real: strong integrity is basically impossible on an unlocked bootloader right now. google has been on a multi-year campaign of progressively patching every workaround the custom rom community comes up with. the play integrity fix project literally got discontinued. you can get basic and device integrity with the right magisk modules, but strong is off the table without a locked bootloader, and most custom roms (including derpfest on xiaomi hardware) can't relock.

what that means practically: google wallet and contactless payments are dead. most banking apps still work fine on basic integrity but google wallet specifically needs strong or it won't let you add cards for tap-to-pay.

i've made peace with this. i use a physical card, it's not the end of the world, and i'm taking the tradeoff of ditching tap-to-pay on my main for an experience that actually respects my phone instead of renting it back to me with ads. when derpfest android 16 stabilizes on the 11t pro (vili) i'm flashing it. the poco was the proof of concept and it passed with flying colors.

the custom rom scene has always been good. but with m3 expressive baked in and teams like derpfest doing real original work instead of just slapping patches on lineage, it's genuinely better than it's been in years. go flash something, your phone deserves better than what it shipped with.

okay, sure. that part's actually not wrong.

what IS wrong is what their CEO, Bailey Pumfleet, posted on LinkedIn the day after the announcement: "open source is dead."

bro. be serious for once.

the argument doesn't hold up

the logic goes: AI can read your open source code and find bugs, therefore hiding the code protects you. it sounds reasonable until you flip it around. if AI can find bugs in public code... it can also be used to audit public code. and that's exactly what's been happening.

Mozilla shipped 423 Firefox security fixes in a single month in April, with AI-assisted auditing helping catch issues that would've taken a human team months to find. Anthropic's own internal model found a 27-year-old integer overflow in OpenBSD's TCP stack that decades of human reviewers walked right past. the same technology cal.com is scared of is the technology making open source more secure, not less.

meanwhile, closed source software has its own nightmare track record. Ivanti's enterprise VPNs, closed source, got hit with multiple zero-days in 2025 that took weeks to even detect because there was no external scrutiny. security through obscurity has never worked. it just means fewer people looking for problems before the bad guys find them.

also, have you seen what open source just did

OpenClaw went from a weekend side project to 355,000 GitHub stars in five months. one guy, Peter Steinberger, built a fully open AI agent framework that surpassed React in GitHub stars, a project that took over a decade to get there. companies like Tencent are building on it. it's running on humanoid robots. it's powering agentic workflows across basically every messaging platform you can name.

all of it open source. MIT license. community governed. no one "protecting" the code.

and it's not like OpenClaw is some outlier. Linux powers the majority of the internet. Apache serves the web. half the security tooling that keeps actual infrastructure safe is open source. the entire modern dev stack is open source. calling open source dead in 2026 is like standing in a building held up by steel and announcing that steel is over.

what this actually is

cal.com had a hard time converting free users to paid plans. that's it. closing the source is a business move dressed up in a security costume. the AI threat narrative is convenient because it sounds technical enough that most people won't push back on it.

and look, i'm not even saying cal.com made the wrong business decision for themselves. maybe closed source genuinely fits where they're going. but posting "open source is dead" as a parting shot to the community that helped build your product is just a terrible take, and it deserves to be called out as one.

open source isn't dead. it's never been more alive. cal.com just left the party and tried to turn the lights off on the way out.

except it doesn't work like that.

the app store is a lie (kinda)

i've been through the CasaOS experience and i'm not going back. the app store looks clean, the icons are nice, everything feels approachable. and then you actually install something and you realize the versions are outdated, the configs are half-baked, and good luck figuring out where CasaOS actually put your data when something breaks.

Immich was the one that really got me. not minor quirks, not a config tweak. fully broken. spent an evening digging through logs trying to figure out what CasaOS's wrapper was doing wrong, realized the compose file it was shipping was months out of date, and eventually just nuked the whole thing and redid it myself with docker compose. which is what i should have done from the start.

the fun part is that doing it properly took like 20 minutes. the CasaOS version ate my entire evening.

and getting to your files from outside the CasaOS bubble is its own special kind of painful. the file manager is bad and the whole thing just assumes you'll only ever interact with your data through its UI. that's fine until it isn't.

who is this actually for

this is the thing i can't figure out. CasaOS markets itself as the easy option for people who don't want to deal with the command line. fine. but if something breaks, and it will, those same people are completely stranded. no good error messaging, open bugs on GitHub that have been sitting there for months, and an abstraction layer that hides just enough to make debugging a nightmare.

if you already know what you're doing, you don't need CasaOS. if you don't know what you're doing, CasaOS will work fine until it doesn't, and then you're in trouble.

just learn docker compose

i know this sounds like "just use arch btw" energy but hear me out. a basic docker compose file for Immich is well-documented, official, and maintained by the actual Immich team. you know exactly what ports are mapped, exactly where your volumes are, exactly what version you're running. when something breaks you can actually debug it. when a new version drops you update one line and run docker compose pull.

no dashboard, no app store, no mystery. just a yml file that does what it says.

CasaOS is trying to abstract away the complexity of self-hosting, which is a nice idea. but the abstraction leaks constantly, and when it does you end up needing to understand the underlying stuff anyway. except now you also have to understand CasaOS on top of it. that's not simplicity, that's just more complexity with a prettier icon pack.

if you're just getting into homelabbing and someone tells you to start with CasaOS, maybe don't. spend an afternoon with docker compose instead. you'll thank yourself the first time something goes wrong.

ok so ye dropped bully on march 28th and i've been sitting with it for a week now. been seeing a lot of takes online, pitchfork gave it a 3.4 which, be serious for once, and rolling stone was also being dramatic. ngl i think a lot of critics are reviewing the man and not the music at this point. so here's my actual take, track by track.

the highs (and they are HIGH)

FATHER with travis scott is a 10/10 and i will not be taking questions. that's the best thing ye has made in years, full stop. the energy, the production, travis actually showing up, everything just clicks. if the whole album was like that we'd be talking about a classic.

PREACHER MAN is a 9.8 and peak is genuinely the only word for it. the sample of The Moments' "To You With Love" is elite, the bass sits perfectly, the religious imagery actually lands. only reason it's not a 10 is that kanye's delivery feels a bit flat compared to the old AI voice demo versions that were floating around. which is ironic lmao. he literally re-recorded the vocals to remove the AI and somehow the AI sounded better. that's a sentence i never expected to type.

MAMA'S FAVORITE is a 9.5 and the bass goes absolutely stupid. the part at the end where his mom is speaking got me. PUNCH DRUNK is a 9/10, straight old kanye feel with a great sample, no notes. DAMN is a 9/10 too, ye's rapping is genuinely sharp there and i love it when he actually raps instead of just singing through autotune.

the midrange (aka the filler zone)

KING opens the album at an 8/10, good sample and the vocals work but some of the background effects are kind of annoying and distract from what could've been a cleaner opener. ALL THE LOVE is an 8.5, the sample is rough at the start and the autotune gets slightly overused but it recovers. WHATEVER WORKS is an 8.8 and i love when kanye just raps, no gimmicks, straight bars. SISTERS AND BROTHERS is an 8.5 but that constant buzzing noise in the background is actually driving me insane. BEAUTY AND THE BEAST is an 8.5, just peak, no other notes needed. LAST BREATH with peso pluma is a 9/10, the spanish vibe works really well and i wasn't expecting to like it as much as i did.

CIRCLES with don toliver gets a 7 from most people but i'd say 8/10, i think it's slightly overhated. nothing special but it's fine. THIS IS A MUST is a 7.5, it's just kinda boring, the trap production feels a bit generic for ye. BULLY with ceelo green is also a 7.5 and just not really my vibe. THIS ONE HERE closes the album at an 8/10, nothing groundbreaking but it works as an outro.

the lowlights

I CANT WAIT and WHITE LINES are both 7/10 and they're basically the dead zones of the album. i cant wait has barely any kanye rapping which defeats the purpose, and white lines just doesn't have anything that makes it memorable. if you told me these were b-sides from a scrapped project i'd believe you.

verdict

8.5/10. look, ye is not making MBDTF again. that ship sailed. but bully is the best thing he's put out in years and at this point that's actually saying something. the production alone is worth your time, and when he's actually rapping and the sample work is cooking, it reminds you why you cared in the first place. just skip the filler tracks and pretend the album is 12 songs instead of 18.

and the examples keep piling up.

movies and tv

warner bros deleted content from users' digital libraries without warning. sony removed purchased films from playstation accounts in germany and austria because a licensing deal expired. that's it. no refund, just gone. amazon got sued over this and their defense was basically "yeah we told you in the terms of service that you're only buying a license." a judge is still deciding if that's good enough.

it's not.

books

amazon removed the option to download kindle books to your pc in february 2025. you used to be able to plug in your kindle, transfer your books to a hard drive, and actually have them. now you can't. everything lives in amazon's cloud, which amazon controls. and this isn't theoretical risk either. in 2009 they remotely deleted copies of 1984 off people's kindles because of a rights dispute. the irony of a company deleting 1984 is so on the nose it hurts. microsoft shut down their entire ebook store in 2019 and every book every person ever bought just disappeared.

games

epic deleted dark and darker from user libraries entirely last year. not just delisted from the store. gone from your account. you paid for it, it's gone. telltale games going under took a bunch of purchases with it. nintendo shutting down the wii u and 3ds eshop killed digital-only titles that can't be bought anywhere else. ubisoft tried to revoke assassin's creed liberation from steam and got massive backlash, but they tried.

there's a whole website called delistedgames.com that tracks games being removed. it gets updated multiple times a week. that's how normal this is.

the pattern

every single one of these companies uses the same playbook. bury the "you're buying a license not a product" disclaimer in the terms of service. use the word "buy" everywhere in their UI. charge you the same price as a physical copy. then when it's convenient for them, pull the rug.

a 2016 study found that most people who click "buy now" on digital content genuinely believe they own it the same way they'd own a book or a dvd. companies know this. the button still says "buy."

california actually passed a law (AB 2426) requiring companies to disclose that digital purchases are licenses, not ownership. amazon updated their kindle store wording to say "you're purchasing a license." that's progress i guess, but it took a law to make them say the obvious thing they've been exploiting for years.

what actually owning something looks like

a physical book can't be remotely deleted. a dvd doesn't stop working because a licensing deal expired. a physical game cartridge from 1994 still runs. the "convenience" of digital came with a massive hidden cost that we just kind of accepted without noticing.

GOG is the one major platform doing it differently. they sell drm-free games, meaning the file is actually yours. no servers, no licenses, no bullshit. there's also stop killing games, a campaign pushing publishers to keep games in a playable state after they drop support, because apparently that needs to be a whole movement now.

the buy button isn't going to change overnight. but maybe stop treating your digital library like it's actually yours, because according to the terms of service you agreed to, it isn't.

i'm not going to pretend school doesn't matter. it does. i get it. but the idea that it's the only thing a teenager should care about is not just wrong, it's actually doing real damage. and i want to explain why clearly, because i'm tired of just nodding and saying nothing.

the data is not on your side

over 60% of teenagers report feeling intense pressure to get good grades. eighty-three percent of teens in the US cite school as their number one source of stress, above social issues, above family conflict, above everything. the american psychological association has been ringing this alarm for years. more than a quarter of teens are currently dealing with burnout, something we used to think only happened to overworked adults. and according to the cdc, nearly 30% of high schoolers say they are experiencing poor mental health.

here's what happens when the pressure never stops and there's nowhere to breathe: you stop trying. not because you're lazy, but because your brain physically can't care anymore. that's burnout. it's a real, documented thing, and from the outside it looks a lot like laziness.

when someone is buried under expectations with no room for anything they actually enjoy, motivation just dies. grades slip. the thing you were pushing so hard for starts falling apart anyway. the pressure didn't help. it made things worse.

what actually builds a future

i code. i build things. i've put real projects out there that real people use. none of that came from a classroom, and i'm not saying that to be cocky. i'm saying it because the skills i'm most proud of came from sitting down and figuring things out on my own, out of pure interest.

and that's kind of the point. when you care about something, you learn it fast, you go deep, and you don't need anyone to push you. that kind of motivation is hard to fake and even harder to force. but it also needs space to exist. if every spare hour is supposed to go toward homework, it gets killed before it even starts.

i'm not asking to drop out or ignore school. i'm just asking for enough room to breathe and actually do things i care about without that being treated like a problem.

what the tension at home actually does

there's this idea that staying on someone's case about school keeps them on track. the research says otherwise. studies on burnout consistently show that pressure from parents is one of the biggest causes of emotional exhaustion in students. not a motivator. a cause of burnout.

and burnout doesn't stay in one place. it spreads. you stop feeling present. you stop feeling happy at home. you feel more like yourself around friends than around your own family, because at least there nobody's asking about your grades.

that's not a character flaw. that's what constant pressure does to a person.

it's not the phone's fault either

there's another thing that comes up a lot. something goes wrong, and the first thing that gets blamed is the phone. or the computer. or screen time in general.

i get where that comes from. there's a lot of talk in the media about phones and teen mental health, and some of it is fair. but a lot of it is way more complicated than "phone bad." researchers who looked at all the big studies on this found that it's not really about how much time you spend on a screen, it's about what you're doing and what else is going on in your life. a pew research survey of over 1,400 teens found that 69% said smartphones actually make it easier to pursue hobbies and interests. and a university of michigan poll found that 56% of parents blame phones for their teen's problems, even when the actual cause is more complicated than that.

one psychologist studying this put it in a way that stuck with me: the phone is more like a mirror than a cause. it reflects what's already going on. if someone is stressed and overwhelmed and has nowhere to feel okay, the phone becomes an escape. taking the escape away doesn't fix the stress.

i use my computer to build things, to learn, to create stuff i'm actually proud of. calling that a problem because it involves a screen is missing the point completely.

what i'm actually asking for

i'm not asking for school to disappear. i'm asking for the conversation to change a little.

acknowledge that hobbies, projects, creativity, and rest are not the enemy of a good future. they are part of it. the teens who burn out aren't the ones who had too many hobbies. they're the ones who had no space for anything except performance. the research on this is consistent enough that harvard's graduate school of education has published on it explicitly.

if the goal is for me to succeed and be okay, the current approach is working against that goal. i'm telling you this sincerely, not to be difficult.

i just want a bit of room.

stay patient lads :P <3

sources

• American Psychological Association, Stress in America survey (teens and stress statistics): https://www.apa.org/news/press/releases/stress
• CDC, Youth Risk Behavior Survey (teen mental health data): https://www.cdc.gov/healthyyouth/data/yrbs
• Pew Research Center, "How Teens and Parents Approach Screen Time" (2024): https://www.pewresearch.org/internet/2024/03/11/how-teens-and-parents-approach-screen-time
• Pew Research Center, "Teens, Social Media and Mental Health" (2025): https://www.pewresearch.org/internet/2025/04/22/teens-social-media-and-mental-health
• Psychology Today, "Are Screens Really to Blame for Teens' Struggles?" (2024): https://www.psychologytoday.com/us/blog/the-baby-scientist/202408/teens-are-struggling-but-are-screens-really-the-problem
• Boston Magazine / University of Michigan poll (parents blaming phones for sleep issues): https://www.bostonmagazine.com/education/2024/08/27/truth-about-kids-and-smartphones
• Harvard Graduate School of Education, research on teen pressure and wellbeing: https://www.gse.harvard.edu

the funny thing is nobody's freaking out. it just... happened. gradually, then all at once, and then everyone shrugged and kept scrolling.

it's not a future problem

this is the thing that gets me. all the discourse around AI slop was framed like something we needed to prevent. "what if the internet gets flooded with AI content?" bro it's flooded. a study last year found that AI and human-written articles are now basically 50/50 on the web. for a brief moment in november 2024, AI content actually tipped ahead. we crossed the line and most people didn't even notice.

and the people who did notice mostly just posted about it and moved on.

the weird part is search still kind of works

here's where it gets less doomer and more just... strange. despite half the web being slop, 86% of articles that actually rank on google are still written by humans. AI content farms churned out mountains of garbage and google quietly deprioritized most of it. so the surface web you interact with daily is still mostly human-written, but underneath that? it's a swamp.

idk if that makes me feel better or worse. like yeah the stuff you find is fine, but the thing you're searching through to find it is mostly noise now. and that noise is what future AI models are going to train on. have fun with that.

nobody's fixing it

the part that actually gets to me isn't the slop itself. it's that there's no real response. platforms aren't labeling AI content consistently. google's playing whack-a-mole. and the people building these slop farms don't care because even if 86% of ranked content is human, 14% of billions of search results is still a lot of clicks.

we kind of just collectively decided this was fine. or we didn't decide anything and it just became fine by default.

the internet got cooked and we all just kept using it. which, fair enough i guess. what else are you gonna do.

i'm tired of it. not because the question is stupid, but because nobody ever goes back and checks if the last guy was right.

the graveyard of predictions

jensen huang said in 2024 that coding is "already dead in the water" and recommended you go into biology or farming instead. farming. lmao.

then dario amodei showed up at davos in january 2026 and said AI will write 90% of code within 3-6 months, and basically all of it within a year. that's not a vague prediction either, that's a specific timeline with a specific number. sam altman said AI would be "the best coder in the world" by end of 2025. marc benioff "seriously debated" not hiring any engineers at salesforce this year.

so where's the accountability? dario said 3-6 months from january. we're in march. i'm still writing code. my side projects still have bugs that AI happily introduces and then apologizes for.

nobody asks the follow-up question

the thing that actually bothers me is that these predictions get made, they get covered breathlessly, and then when the deadline passes everyone just... moves on. there's no "hey wait, you said 6 months ago that coding would be mostly automated, what happened?" instead we just reset and wait for the next prediction.

and some of these guys have a direct financial incentive to make you believe AI is replacing everything. like of course the ceo of an AI company wants you to think AI is going to eat the world. that's his product. that's like asking your dentist if you need more fillings.

(okay maybe im being slightly dramatic but be serious for once)

what's actually happening

look, AI is changing how i code. i use it constantly. it's fast and it saves me a lot of time on stuff i'd rather not think about. but the idea that it's replacing engineers wholesale? we're not there. not even close.

the models still confidently break things, hallucinate APIs that don't exist, and fall apart when a project gets complex enough that context starts mattering. the engineers at anthropic who "don't write code anymore" are still doing the hard part — knowing what to build, why, and whether the output is correct. that's not going away.

what's actually happening is that the floor got lower. you can do more with less now. but "the floor got lower" is not the same as "the ceiling collapsed."

just say it already

i think these predictions keep getting made because fear travels faster than nuance. "AI will write all your code in 12 months" gets a million impressions. "AI is a useful tool that requires a skilled operator" gets ignored.

the people making these claims aren't dumb. they know what they're doing. and the media covering it knows that doom headlines get clicks. so the cycle keeps going and nobody ever gets held accountable when march 2026 rolls around and developers are still, y'know, developing things.

maybe in 5 years i'll look back and feel silly for writing this. maybe. but right now, in the actual present, i'm watching the same prediction get made on a loop with the serial number filed off each time, and i think someone should just say it's kind of embarrassing.

here's the thing nobody says out loud: building stuff for yourself is completely valid, and it's also completely fine if it dies when you stop needing it.

built for me, not for a pitch deck

a lot of what i make starts as a personal itch. i need something, nothing exists that does exactly what i want, so i build it. i use it for a bit. then life happens, or i just lose interest, and the project quietly flatlines.

and honestly? that's okay. the project did its job. it was never a product, it was a tool. tools don't need to live forever.

the problem is when you start treating every side project like it needs to be a startup, like it needs a landing page and a waitlist and a ProductHunt launch. most things don't need that. most things are just you solving your own problem in your spare time and that's allowed.

the 1% rule

out of everything i've built, probably 1% has ever gone public. the rest either never left my machine or quietly got abandoned.

that sounds depressing until you flip it: the stuff that did go public, i actually believed in. i only shipped when i was genuinely convinced it would help other people, not just me.

the wrapped thing (wrapped.devmatei.com) is the clearest example. i built it because i use Navidrome instead of Spotify, and every December my friends post their Wrapped and i just sat there. so i made my own. it was scratching a personal itch first. but then i thought, there's probably other people using ListenBrainz who feel the same way, and there was no good tool for this yet.

so i cleaned it up, made it actually nice, and shipped it. it got listed on the official ListenBrainz apps page, blew up on Reddit, and it's still the most successful thing i've ever made. all because it started as something i actually needed myself.

why the graveyard isn't a waste

every dead project taught me something. a React pattern, a new API, how to not structure a database. the learning doesn't disappear just because the project did.

and there's something weirdly freeing about building with zero pressure. no users to disappoint, no changelog to maintain, no github issues from strangers at 2am. just you, a blank file, and a problem you want to solve. some of my best experiments came from projects that lasted two weeks and then got yeeted.

the toxic part of dev culture is acting like every line of code needs to ship or it was a waste of time. it wasn't. you practiced. you learned. you scratched the itch. move on.

the actual bar for shipping

if i'm being honest, the bar i set for myself before going public is pretty high. it needs to actually be useful for people beyond just me. it needs to be good enough that i'm not embarrassed by it. and ideally it should hit a gap that isn't already covered better by something else.

that combo is rare. hence the 1%.

since wrapped, i've been trying to find that same lightning-in-a-bottle feeling. personal need + other people have this need + nothing good exists yet. it's a harder formula to hit than it sounds. most ideas fail one of those three.

but the dead projects in the meantime? i don't regret them.

stay unbothered lads :P <3

so here it is, once and for all.

proxmox is cool. i just don't need it.

i run jellyfin, navidrome, immich, adguard home, a couple discord bots, an internet radio setup with liquidsoap and icecast, a tidal downloader, and a bunch of other random stuff on a minimal debian 13 install. it works. it's fast, it's clean, it's mine. docker compose does exactly what i need it to do and i don't have to think about it.

proxmox is genuinely great software, i'm not here to dunk on it. but there's this weird thing in the homelab community where if you're not running proxmox + VMs + containers inside VMs inside a whole orchestration stack, you're somehow doing it wrong. be serious. i'm not running critical infrastructure. i'm hosting my media, my music, my photos, and blocking ads.

the "but what if you want to run VMs" argument

i don't. and even if i did, i can run VMs on debian just fine. this isn't a gotcha, it's just reality. proxmox makes VM management really smooth, sure, but spinning up a KVM on debian isn't exactly rocket science either.

the pitch for proxmox is basically "it's easier to manage lots of things." i have like eight containers. i can manage them with my eyes closed at this point.

and then there's the reinstall

look, even if proxmox somehow solved every problem i don't have, i'm not reinstalling everything. my setup is dialed in. my compose files are clean. immich has months of photos indexed. navidrome knows my whole library. jellyfin has my watched history. my internet radio stack is actually running smoothly for once. my discord bots are up.

you want me to wipe all of that to move to a platform that gives me features i wasn't asking for? ngl that sounds like a terrible saturday.

minimal is a feature

there's something genuinely nice about knowing exactly what's on your machine. debian 13, docker, compose, done. no web UI for the OS itself, no abstraction layer i didn't ask for. when something breaks i know where to look.

proxmox adds a lot but it also adds a lot, if you know what i mean. for a homelab that's not hosting anything important, that tradeoff just doesn't make sense to me.

if you're running a server rack with 10 services and actual uptime requirements, yeah, proxmox makes sense. but if you're like me and just want your stuff to run without babysitting it, a minimal debian install with docker is more than enough. the hype around proxmox makes it sound like anything else is amateur hour and that's just not true.

stay unbothered lads :P <3

okay so "vibe coding" is everywhere right now and i'm kinda losing my mind. everyone and their dog is posting about how they built a saas in a weekend without writing a single line of code, just vibing with cursor or claude or whatever. cool. but there's a massive difference between using ai as a tool and just... not knowing what your code does. and i feel like nobody wants to talk about that.

it's like working in a team

here's how i think about it. ai is basically a teammate. a really fast one that knows a ton of stuff and can crank out boilerplate like crazy. but also one that makes mistakes, sometimes very confidently, and has no idea what the rest of your project looks like. you wouldn't just merge every PR your teammate opens without reading it right? same thing. you still gotta review it, test it, actually understand what just got added. that's just how it works.

where it gets me is people treating "i didn't write a single line" like it's something to brag about. bro that just means you have no idea how your own project works. that's not productivity that's just vibing your way into a codebase you can't debug lol

the huntarr thing

if you need a real example of why this matters, look at what just happened with huntarr in the selfhosted community. someone did a basic security audit on it and found 21 critical/high severity issues. we're talking full auth bypass on api endpoints, anyone on the network could just grab your sonarr and radarr api keys without even logging in. the kind of stuff you catch in a five minute code review.

and the commit history? exactly what you'd expect. hundreds of commits called "update" or "patch", dozens of files changed at once, no review process, just raw prompting and pushing. the maintainer apparently claimed to work in cybersecurity btw. repo and subreddit got taken down after the findings went public. people had been running this on their homelabs trusting it with credentials for their whole media stack. not great.

the thing is you genuinely can't catch that if you don't know what bad auth looks like. the ai will produce something that looks fine and you'll ship it and your users pay the price. that's the actual danger of full vibe coding, not just messy code.

ai is good actually (when you use it right)

i use ai for coding constantly and it's genuinely great. it speeds up the boring stuff, helps me with syntax i forget, gets me unstuck when i'm going in circles. that's just good tooling. same as using a framework or stackoverflow, it saves time as long as you actually understand what it's doing for you.

the people getting the most out of it aren't the ones letting it drive the whole thing. they're the ones who already know how to code and use ai to go faster. they read what gets generated. they test it. they fix the bugs. they know when it's confidently producing garbage.

coding isn't dead lol

the "coding is dead just prompt bro" crowd really needs to touch grass. ai can't design your architecture, it can't understand your users, it definitely can't debug the weird edge case that shows up at 2am. it makes mistakes constantly and without someone who actually knows what's going on those mistakes just stack up until something explodes.

use the tools. go faster. just don't outsource your brain in the process :P

so. redflags.devmatei.com.

i built this thing that gives you music red flags based on your listening history. plug in your last.fm, listenbrainz, or navidrome and it'd roast you a little. "Same Track Rotation", "Comfort Zone Core", "Artist Monogamy", that kind of stuff. i was actually pretty proud of it, posted it on a bunch of subreddits, and within the first minute the traffic was already coming in. i was both excited and lowkey panicking watching the numbers go up.

and then someone commented "disappointing" and i deleted everything.

all the posts. gone. looking back that was genuinely one of the worst decisions i've made lmao, i fully regret it. but i get why i did it. that one word hits different when you're already anxious about whether what you built is actually good.

the real problem though

the comment wasn't wrong.

the flags WERE too general. and the reason they were too general is entirely my fault, i made them that way on purpose. i wanted everyone who used the tool to get something, no matter what their listening looked like. the idea was that nobody should open it and feel like "oh i'm just boring i didn't get anything." sweet intention right?

yeah it killed the whole point.

when everyone gets a flag, the flags mean nothing. "Same Track Rotation" sounds like a fun callout until you realize basically anyone who has a favourite song gets it. "Comfort Zone Core"? congrats, you listen to music you like, here's your red flag. it's not a red flag at that point, it's just a participation trophy with a spicy name.

the personality wasn't there either. the flags were neutral, safe, almost corporate. no bite, no specificity, nothing that made you go "okay how did it KNOW." and that's the whole fun of these things. the moment it feels weirdly accurate is the moment you screenshot it and send it to your friends.

i was so focused on making sure nobody felt left out that i forgot to make it actually interesting for anyone.

what i should've done

made it mean something. if only 30% of people get "Artist Monogamy" because they genuinely listen to like two artists on repeat for months, that flag becomes a flex. or an identity. people would screenshot it specifically because they didn't get it or because they did. exclusivity is the whole game with these shareable things.

the flags needed personality too. less "Comfort Zone Core" (sounds like a wellness app), more something that makes you laugh or cringe or both.

anyway

i'm probably gonna rebuild it at some point and actually make it ruthless. no participation trophies, no softballing. if your music taste is unhinged it's gonna tell you, and if it's basic it's gonna tell you that too.

and next time someone comments "disappointing" i'm gonna fix the project instead of nuking the posts smh

stay tuned lads :P <3

i’m honestly losing my mind lately. every time i try to explain a piece of news or some tech stuff to my family, my brain just... freezes. i consume so much info in english that it’s basically become my default way of thinking, and now i’m out here forgetting how to be a normal romanian speaker.

it’s getting embarrassing at this point. i’ll be halfway through a sentence in romanian and have to tag-team it with english words just to get the point across. i feel like im forgetting romanian.

the worst part? the direct translations. i’ll take a common english saying, swap the words to romanian in my head, and say it out loud—and it sounds absolutely cursed. people look at me like i’m speaking a dead language or just talking total shit.

i’m not even trying to be "that guy" who thinks he’s too cool for his native tongue, my brain just decided english is more efficient for the news and dev stuff i care about. now i’m just stuck in this weird middle ground where i talk like a native english speaker but in the wrong language smh

staying bilingual is hard when EVERYTHING is 100% english lads. :P <3

it's weird seeing the traffic drop off, but honestly, it's a relief. at the peak, i had around 2,500 wrappeds generated, and for a project running on a gunicorn server in my closet, i'm just glad nothing actually caught fire. (okay maybe im being slightly dramatic)

why i even built this

i used to be a spotify user, and honestly, the only thing i missed after switching was the wrapped at the end of the year. i'm a huge navidrome user now, but i still wanted that "wrapped" experience. since it didn't exist for the platforms i use, i just coded it myself.

i'm super grateful for all the last.fm, listenbrainz, and navidrome users who actually gave it a shot. seeing people use something i made to enjoy their music stats is peak.

the listenbrainz thing

the coolest part wasn't even the numbers, it was the fact that the ListenBrainz team actually noticed. they included make-a-wrapped in their official "year in music" event as a community project.

seeing a project you built for fun get recognized by the actual devs of the platform is a weird feeling. it makes all the debugging worth it.

technical regrets? lol

i’m still running this on my debian homelab (the old lenovo laptop i talked about in another post). looking back, using a raw gunicorn server as a systemctl service instead of a proper docker setup was definitely a choice. it worked, but it wasn't the "cleanest" way to handle a spike in users.

the hype def died down, but i'm proud of what i built. it was a challenge managing the traffic and seeing how far i could push my hardware.

if you used the tool and liked it, feel free to share it or hit the donate button. helps keep the servers (and my sanity) alive. :) <3

honestly everyone thinks a homelab needs to be a $5k rack in a basement but i’m literally out here running a whole ecosystem on a lenovo yoga 500-14isk. it’s been since september and somehow it hasn't caught fire yet so we’re winning ✌

the hardware

running debian trixie on an i5-6200u with 8gb of ram. i know, i know, 8gb in 2026 is criminal but it’s holding up (cant upgrade anyway cuz of mobo). the hdd is ALSO slow asf sadly but it gets the job done.

devmatei@homelab:~$ fastfetch
       _,met$$$$$gg.          devmatei@homelab
    ,g$$$$$$$$$$$$$$$P.       ----------------
  ,g$$P""       """Y$$.".     OS: Debian GNU/Linux 13 (trixie) x86_64
 ,$$P'              `$$$.     Host: 80R5 (Lenovo Yoga 500-14ISK)
',$$P       ,ggs.     `$$b:   Kernel: Linux 6.12.57+deb13-amd64
`d$$'     ,$P"'   .    $$$    Uptime: 48 days, 23 hours, 26 mins
 $$P      d$'     ,    $$P    CPU: Intel(R) Core(TM) i5-6200U (4) @ 2.80 GHz
 $$:      $$.   -    ,d$$'    GPU 1: NVIDIA GeForce 940M [Discrete]
 $$;      Y$b._   _,d$P'      Memory: 3.42 GiB / 7.56 GiB (45%)
 Y$$.    `.`"Y$$$$P"'         Disk (/): 234.88 GiB / 907.20 GiB (26%)

the docker stack

i’m running a bunch of stuff to stay off the grid and keep my data mine. here’s the breakdown:

• music: navidrome + feishin + tidaloader. basically my own spotify but better. plus i got liquidsoap and icecast running "swagfm" for no reason (private radio for copyright reasons)
• photos: immich server. i dont actually really use it i do sync everything once in a while tho
• utils: adguard home (kill the ads), cloudflared (tunneling so i dont expose my ip and i cant do port forwording anyway), and glance for the dashboard.
• bots: petpal (friends discord bot) and symphcord (my discord bot) just doing their thing in the background.

[!NOTE] yeah i run the make your wrapped project too on here but that’s as a service with gunicorn cuz i didnt feel like configuring docker that day haha

the internet stats

vnstat is saying i’ve moved like 282gb this year already. estimated to hit over 4tb by the end of 2026. my internet provider probably hates me but who cares they dont give me port forwording and easy dash access ( i gotta call them to req access ) so they might as well f themselvs

if u wanna build something similar just find an old laptop and stop paying for subscriptions yall

let’s actually use our brains for two seconds.

1. datacenters aren't new, you just found out about them

everyone acting like AI just dropped out of the sky and started chugging water like crazy. bro, where do you think your stupid tiktoks have been stored for the last five years? where do you think netflix runs from? datacenters have ALWAYS used water for cooling. it’s the most efficient way to keep chips from literally melting. but suddenly because it’s "AI" and not "scrolling through memes," it’s a global emergency? thats just stupid

2. the "closed loop" reality

the biggest lie people push is that this water is just... gone (specifically drinking water). it’s a closed-loop system. the water flows in, picks up the heat from the hardware, goes through a cooling tower or a heat exchanger to chill back down, and goes right back in. it’s not being injected into the sun.

and the heat? people acting like a datacenter in frankfurt is somehow raising the temp of the arctic circle. be serious for once the heat dissipation from these buildings is NOTHING compared to actual heavy industry, cargo ships, or just... the sun. the polar bolars are struggling because of actual systemic pollution, not because someone asked chatgpt to write a python script.

3. context matters (stop the bait)

if you actually care about water usage, go look at how much it takes to maintain a single golf course in a desert or how much industrial farming wastes. but nobody wants to post about that because it doesn't get the "AI IS EVIL" clicks.

it’s honestly just trendy to hate on tech progress right now. the water usage is insignificant in the grand scheme of things, and most of it is recycled anyway. stop letting these headlines fry your mental. we’re building the future (okay maybe not, its a bubble but who knows after all), and it’s not gonna stop because people don't understand how a radiator works.

stay informed lads :P