artem

there are a lot of doomers on x and reddit and prettgy much everywhere on the internet saying it's impossible to get a job. idk bros 2+ weeks of focused learning studying and locking in makes you better prepared than most candidates on the market. i go to job interviews several times every month for senior java positions, and the questions are always the same and easily crackable if you just study the subject they're asking about. you don't need 52+ years of experience you just have to push yo ass and do a bit of focused learning. two weeks is more than enough. big corpos are still thirsty for developers, and this is the last place that'll be automated by ai agents given how inflexible most of them are and how they're scared to "leak" corpo data to llm providers (idk why this is the case I see in biggest banks of my country), so there's plenty of time if you're not ok with corpo, that's ok. you can be really helpful on x, in discord channels of projects you like. actually build something, and you'll get a list of problems you need to overcome. write a post/blog about how you solved those problems, leave meaningful comments on github issues. just show up every day, and people will notice. they'll see you're actually helpful and can deliver. dooming too much steers you away from just doing. consider that we all can make it brah.

@crypt0lake Finland is also a lot of russians with vpn

that tunisia color is literally just me

@kuberdenis use this next time pliz youtu.be/Xsrjy7140Lw

guys i learned how to do the funny boom effects

@pvncher @RepoPrompt im stupid lol. i thought something was updated but i was in "ide mode" instead of agent one and that's why i was confused

@pvncher @RepoPrompt I decided to try repoprompt again recently, got overwhelmed with new updated interface, will try again after reading the docs

I’ve been using 5.4 non stop in the @RepoPrompt agent since launch with none of these issues. It uses codex app server under the hood, but the tools and system prompt are different. It’s still my absolute favorite model right now

@DaLoneModerate @kuberdenis Americans are just retarded

@kuberdenis Dude do you see any tech Asian man here with a white gf? As you know most Asians in sf and NYC easily make 200-300k and probably cheat / drink / do other crazy stuff a lot less than most other race. Let that sink in.

I think i can read and pronounce chinese now I do not understand it but i am able to pronounce it correctly (fact checked through instagram reels) Go on write something in chinese phonetic in the comments and i’ll verbalise it writtenly

This codex issue is now fully resolved and stable for the last couple of hours. You have come to expect it, but yes, that means we will be reseting rate limits in a bit. Enjoy.

@maria_rcks @kitlangton finally t4code my beloved

Introducing... t4code (Tailscale+t3code) Code remotely, with the best interface... @kitlangton im sorry you got slopfork'd

is it possible to migrate my local codex sessions into T3 Code? @theo

@VictorTaelin @guidomodarelli this

@guidomodarelli the problem is that plan mode often takes as much time as just doing the thing, at least on my stuff. and then it gives me a 8-pages long wall of text that I'm supposed to read and approve it or not. at this point I'd rather just ask it to do the thing and look at the code

... IMO GPT-5.4 is kinda similar to GPT-5.3-codex, at least for my own stuff. I don't feel an intelligence bump as noticeable as what I felt from 5.2 to 5.3-codex, but it is definitely smarter. it still nails amazingly complex requests, and it did solve some problems that 5.3-codex failed, but still fails to understand intent in similarly silly ways. this makes it substantially less effective in practice than it could be, because, to use it well, you must be able to specify exactly what you want, precisely, with no room for misunderstanding, at the risk of losing 30 minutes of "work". I thought I was very good at this, until I met 5.3. there is always some valid interpretation to my literal words that I couldn't anticipate, and you can't even blame the thing, because it is always... technically correct. I think this is a serious problem that has to be fixed on their RL. if we keep going in this direction, then I can't help but wonder if GPT-6 will just be called GCC that said, this all comes from someone who's desensitized, fully immersed on the exponential. zooming out, the kind of work these things do now would be unthinkable, even last year. GPT-5.4 basically one-shot Bend2's effect system, including the IO type, dozens of handlers (print, read_file, etc., for each compilation target!), a monadic do-notation, and even a complete HTML5 compiler that allows us to write web apps in an Elm-like fashion. it did all that while I was grieving on my bed, I reviewed the code and it was very good. I also appreciate the 1.5x mode but it doesn't feel actually faster than 5.3-codex in practice, perhaps even a bit slower, but I'm not sure if that's the case. they mocked Anthropic for the price, but I'd take a more expensive 2.5x mode over that any day, specially since GPT-5.4 is much cheaper to begin with. also I may be wrong but does feel like its compression got a bit worse? it struggles to remember the first chat in a way that GPT-5.2 didn't, but perhaps I just got used to it. it is getting harder and harder for me feel how good a model it, specially when the bar is so high, and outputs are now massive PR's, instead of just a function or two so these are my thoughts, I think GPT-5.4 is the best model we have for complex tasks, but I'd still use Opus when I need something faster, with a better sense of style, and a bit less autistic ...

@kuberdenis @DeepLcom @Merantix waow codex is now in your workflow instead of claude?

i’d love to meet with moots in Berlin i’ll be hosting a short talk about my very boring ai-assisted software development workflow /w tmux & codex next month on a community event hosted by @DeepLcom at @Merantix :) it is April, the 16th and the conference is called >prompt

@kuberdenis @DeepLcom @Merantix dayuuum sad to skip

@floinkus > makes only mistakes like a boss

/loop make a $5M app that becomes more profitable every time you run this loop. make. no. mistakes.

@mikeysee @theo @OpenRouter @convex damn I hope openai will address this issue @thsottiaux

So just following up yesterdays discussions with @theo here are the results testing various reasoning efforts through @OpenRouter on the @convex evals. The GPT 5.4 xhigh result was the most surprising to me, so I re-ran it again to check and it got the same result which is inline with what Theo was saying that xhigh is worse than high.

@pvncher @RepoPrompt hmmm so it's useful for "I am interested what can I add to my app > run investigate > get a lot of ideas" or "I want to make my app more production-ready > run investigate > get a good bug report" ?

One of my favorite @RepoPrompt workflows is investigate. It generates a research doc and does deep research to generate hypotheses the agent can test. It lets the models work together to rule out scenarios and get to the root of an issue. Has solved many issues for me!

@Harsh_oxy @creatine_cycle I was his spotter actually

@creatine_cycle @ulanyde Ig he never went to a gym

>at the gym >some big guy: "hey bro looking huge" >me: "thanks, i am afraid of women" >him: "hell yeah me too bro" >me: "hell yeah"

@kuberdenis @thsottiaux @Yuchenj_UW bro unleashed his pic stash

@thsottiaux @Yuchenj_UW

Singularity. I’m sure Codex is 100% written by Codex too.

@thsottiaux @Yuchenj_UW lmao

@Yuchenj_UW Not yet, and that’s why it works extremely well

@thdxr idk I think this guy is just clout farming by posting fake news he did really stupid and embarrassing thing which is really hard to do honestly. I mean how you would allow claude code to delete all prod data this is so stupid

we'll adjust to this new world but this is why i was saying the people responsible for security/safety are freaked out your company maybe lax now to let everyone experiment but those times are probably ending