Hope you weren't waiting for Ladybird to save you from Chrome and Firefox.
>>109875064I still just manually manage 1-3 (mostly 1) agent(s) at a time. Is everyone catching AI psychosis or am I just not getting with the program. I can't tell. Usually I just ignore this kind of noise but I respect Andreas, or at least used to idk what he's up to nowadays.
>>109875064All 3 use AI, what’s your point? Mozilla earlier this year said they used Claude to find security flaws to patch
>>109875076It's just very weird to me how people can use AI and not see how shit it is at actually writing code. As a rubber duck and/or partner to design stuff? Sure. Code review? Absolutely. Debugging? Usually great. But for writing actual code? How can anyone let this thing code on its own, let alone run an entire team of AI "programmers" when even the top models routinely fuck up basic shit and get confused? I just have to assume that anyone who uses AI for this and finds the output acceptable either has no standards or didn't actually look at what's being output beyond "does it compile and does it run?".>b-but you haven't tried DA REAL AII get a few thousand dollars worth of GPT and Claude usage at work, I'm talking about Fable 5.1 here, it just gets worse as you scale down.
>>109875142My experience is that AI generates decent but not perfect code when you give it a well-defined shape for what it should do (either reusing some existing pattern in the codebase, or asking it to write something in a very particular way)If you let it off the hook then it will slowly start introducing more and more code that is frankly just toxic for maintainability, at least from a human perspective.Absolutely incredible for finding bugs though, both in terms of code review and "help me debug this thing". Like seriously game-changing.
>starts project where everything is written from scratch>abandons project to start another project where he will never write code againAndreas is a fucking weird dude
>>109875064this faggot:>has infinite tokens>runs unimportant projects which result doesn't need to be audited
>>109875095>Hey word-guessing machine, find me somethingis a very different kind of gay from >Hey word-guessing machine, write me something
>>109875402what kind of gay are posters in this thread
>>109875532standard /g/ays
>>109875532Lot's of reddit spacing, so gay just for the updoots
>>109875064I had hope for Ladybird but I'm really counting on Pale Moon to keep up, and it is for the most part. Depending your preferred sites you rarely have to open a fallback browser. And usually the offending sites are bloated webapps that are better used with "ungoogled-chromium --app=https://x.com"A lot of the common complaints of PM come down to people using it wrong.
>>109875064I already rejected ladybird when they did the swift migration, then abandoned swift, then did some rust migration, now they're going full "agentic coding" mode.
>>109875064Based vibeGOD creating stuffLuddite seethe incoming
>>109875064I see infinite token farmer says, tokens are good
Ladybird will still mog firefox in performance
>>109875064Thats what it takes to build a browser anon
>>109876596Same, I remember seeing Swift and thinking "at least it's not rust" lol
>>109876605Can't wait the propiertary stuff to handle their handsLibtards Going to bed incoming.
>>109875064The only reason Ladybird has been able to make such rapid progress in the first place is thinks to all of the AI he's using to develop it.It would have taken so long to build a browser from scratch otherwise.
>>109876596>>109876919Constantly jumping onto new stuff is never a good sign, you cannot be that kind of person if you want to build a browser. >just one more tech stack and I'll finally make it bro please just one more memelang bro this one will change everythingIf you have the right mindset for this kind of work, you'd be with your head down writing code in C/C++ with little knowledge of what some tech grifter on Twitter is shilling this month.
>>109877954Well he did have that mindset for years though while building serenity and ladybird, looks like he drank the swift/rust/ai agents kool aid after that and it's not a good sign. Maybe get the full browser out and rewrite it in rust if you want to.
AI saved FOSS.
>>109878212Was there ever a point in history when the bottleneck was code quantity rather than quality?
>>109878217every foss project that went dormant due to a lack of contribution. ai agents will never tired up.
>>109875064You guys have agents? I just give the entire project zip to chatgpt and tell it to find bugs
This guy was a severe drug addict and then started using steroids, there were too many red flags from the start.
Complaining about AI being used in software in 2026 is like complaining about Fortran in 1960.
>>109876287Pale Moon just backports Mozilla code and force it to run in the obsolete XUL code. Its been nearly 10 years since XUL died, modern extension developers don't even know how XUL works anymore.
>>109875142>how shit it is at actually writing codeSkill issue
>>109875064>Hope you weren't waiting for LadybirdRustshit? Nah, not reallyYou don't have to wait anyway compiling is still a thing
>>109875142I think the point is that it doesn't matter how bad the code is because no human is going to look at it. I think we're going to find out in year or two if that holds... AI tends to make lots of unnecessary code and safeguarding against unnecessary things, so the question is how big of a mess can the AI work with? but in general I think the big realization here was that agents bruteforcing through the code quality problem will actually produce something that works (and burn through immense amount of tokens). We'll see
>>109878781>AI tends to make lots of unnecessary code and safeguarding against unnecessary thingsIronically that is to safeguard against itself because the machine knows that its slop generator is garbage so it will add code to check for edge cases that absolutely should never exist ever just in case it fucks up somehow and that code handles it gracefully instead of blowing up
>>109875142People are going to call this a skill issue but its actually more about building an intuition for navigating the jagged capabilities of models. Hand it a task with a well defined end state and it will do really good work, vice versa not so much and you'll need to do a lot of guiding. Sometimes you can still get good results from poorly defined tasks if those tasks are highly represented in the training data, for example most SOTA models will easily knock out a react webapp for some really common shit. But as soon as you move away from CRUD programs and lack a good quality oracle for the model to iterate against quality of model work is going to tank. Once you understand this you can actually get value from the models since they can save you a lot of time if you outsource the stuff theyre good at to them but they will quickly eat up that time and energy savings if you try to force the square peg through a round hole.
>>109875402They arent word guessing machines anymore they are task-reward maximizing machines. Most of the improvements we've seen in the past year are from RL post training, and what you're describing is the pretraining loss minimization process. If you have a task with a clear goal and ways to catch reward hacking they are very capable. Security is one of these, "find the bugs then prove it with a poc" is pretty easy to implement and verify.
>>109878841there's a big chance I have completely misunderstood how these work but:how do these agent orchestrator systems make sure the "well defined end state" is split into well defined small pieces so the sub-agents don't fuck up the implementation details? I haven't used any agents but whenever I ask just chat with an LLM I know I have to be very specific about my usecases and what sort of data layouts I expect, otherwise it will just put in some generic containers that will not be up to the task. so this gives me the feeling that it's not an easy task for the automation pipeline to give me the actual result I want.
>>109875186>shape>>109878841>oraclei guess it's nice that AI learned how to COPE like low-end humans
>>109875064>Hope you weren't waiting for Ladybird to save you from Chrome and Firefox.Chrome and Firefox devs are doing the exact same thing btw.
>>109875064Is he retarded? The llms will be trained on the chrome and firefox codebases. So now he's just doing a really expensive compilation of chrome
It reads a bit like that "day in the life of a woman in tech" videoor that "activated almonds" snippetby which I mean, hella gaybut I wish him the best, at the end of the day he has accomplished far more than I ever will, and I'm just some faggot
>>109875064it's so weird to see this guy go from carefully crafting software and waxing lyrical about the value of the process vs the end result and doing things yourself and understanding everything you do to full on AI psychosisI understand that was for a little hobby os and now he's trying to crate a real alternative browser for real world use, but just looking at this post, taken at face value, why isn't ladybird done? a handful of these guys managing frontier models should have gotten the browser done from scratch in a couple of weeks
>>109875064Ladyboybros... not like this...
>>109878730Retards always say this but when you ask them to present their vibeslopped software they never do. You're a 10x engineer with this shit, right? Maybe even 100x? So even if we only take the last year, that's 10 years (or 100) worth of development. Please show your amazing creations. >>109880708I genuinely think it's a drug addiction relapse. Something that was hard and time consuming before is now done for him in a tenth of the time with no real effort on his side. I know for a fact it gives a dopamine high at the start, it did for me too, but that falls apart after a little while when you start noticing just how quickly it becomes a bloated unmaintainable mess.He was a drug addict, not only does the subconscious memory of those highs never truly go away but you have to be predisposed to it to fall into it in the first place. I think he's getting a little bit of that high back with this and for him it's a stronger pull than for most others. We'll see in a few months if he manages to pull out of it.
>ask AI to write code for a simple NPC that follows player in Godot>says something about splitting code in 3 files, ends up dumping everything in 1 file>shit doesn’t work>reads random binary files for no reason, tells me there is no navmesh generated (it is generated)>turns out little retard used velocity_computed() but didn’t toggle the property to actually use it; tell it what’s wrong>it tells me I’m wrong and keeps looping about navmesh>I was correctAll these xitter retards have to be paid off, right? There is no way they use this shit every day and their experience is THIS different.
>>109878871That happens with all overfitted reward mechanisms, the specific problem with LLMs is that they just guess words but the meaning behind them is a complete mystery to them even if they were fed the definition.
>>109876605>creating stuffasking for stuff
