>>108464255Lol what were the "recent activities" I wonder...
>>108464255anon what did you post ? Can you get a refund ?
Imagine using public AI for private shit. Would you also use a community glory hole? Hang a sign in your window listing your fetishes? Basically the same behavior.
idk why they banned me no reason was given. I've been building a language glosser and I have an automatic loop that calls claude -p processes to analyze sentence glosses for potential issues. I told opus to keep working and looking for ways to reduce overhead on the loop and I think it extracted my subscriptions oAuth tokens and used it to call the API or something (which I didn't tell it to do). That might've been what got me banned or it might've just been an autoban from too many instances that looked like a farm, or I might have gotten mistakenly flagged as using 3rd party tools when I wasn't, or it could've been the hardcore celebrity fantasy erotica I had opus write 3 months I'm not really sure
>>108464278Yes yes and yes
>>108464255You deserved it
>>108464255You are a pedophile. Leave the AI alone.
>>108464278Combining the strength of engineers just makes sense to me. I'm certainly not going to hold back the world just for a sense of pride over my code. That's like taking pride in writing equations, it's just math bro. Problems we all are trying to solve and have not enough time to. Let's combine out efforts and finish this faster.
>>108464278>>108464304>>108464300in all seriousness I have legit just been using it 10 hours a day everyday for nothing but building my glosser/software for the past few months no sexy/shady stuff
>>108464315>10 hours a day for monthsi think you got enough free labor, chud
>>108464274yeah I got my 200$ for this month back and switched to codex alreadybut I have massive harness engineering in my project (huge documentation system, hooks, memories, dynamic prompt injection, etc.) to get Opus to perform optimally and its going to take at least a week just to rehabilitate the system for codex
>>108464341I'm doing my PhD thesis in robotics and Claude fails to implement 90% of the equations I give to him idk how you guys do
>>108464315and a ching chong nip nong to you too, young lady
>>108464344This but with database entry tables. The thing that it should do the best.
>>108464310What the fuck are you even talking about?
>>108464344>>108464368The people claiming success with AI probably don't know how to tell when it's wrong. Like a tone-deaf person happily singing out of key, oblivious to it.
>>108464389Scan all the code, normalize towards the best. Why not? Pride? Fear?
>>108464404you don't have AI as the actual designer you just use the same sort of automatic scripts/bots/pipelines you'd normally use but add ai calls into it to give it extra capabilities and have safeguards to keep it within limits
>not insulting the LLM until it gives up
>>108464420I wouldn't trust it. What safeguards can you even have that don't completely cripple it? Prompts can't be safeguards.
>>108464491for dev work you give it objective tests it has to pass and dont let it stop iterating until it passes+ you can use hooks to prevent it from touching or doing the wrong things, etc.for my glosser I have code that forces the LLM to output in the right formatI have frontier-level harness engineering on my project which is the only way to get it to work correctly but anthropic is working on the same stuff on its own and will incorporate it built-in to claude sooner or later.
>>108464512My PhD involves training a lot of models and so testing the crap Claude does takes days. Usually it does like that >be me, annoyed cause I can't find why my shit doesn't work >"hey this model is doing A but I want it to to B fix it">"absolutely here's the corrected code">ffw 8h>it's worse, I'm getting angry >"hey heuuuu, it's worse">"oh nooooo anonino you, let me fix it ~~">fucks up everything changes my equations and refactor the codes>mfw still doesn't work and now I'm frustrated mind you I pay for this experience
>>108464344mostly when I read this it's people using shitty modelsyou have to use Opus 4.6 on max thinking otherwise you will get garbage resultsSame with Gemini, if you aren't using Pro what are you even doing
>>108464577I am using Claude code and yes I explicitly change the model then put it in max mode /init then ask my question. I'm not even trying to be the anti AI guy, I would be so happy if AI could help me graduate lmao
>>108464310>I'm certainly not going to hold back the worldNo. You're just putting your intellectual property out there publicly so billionaires can take it, bundle it, strip your name and sell it privately to the guy in India that'll do your job for pennies on the dollar.
>>108464605They have to find it under petabytes of fucking garbage firstKEK
>>108464605Who cares? I'm not going to let some guy in India making pennies prevent me from distributing it for free since it benefits both me and everyone else. What is this boohoo boogeyman nonsense about some man I will never even encounter and probably lives a horrible life.
>>108464565So the way you would get this to work is1) You have a programmatic test for some objective outcome and a loop where claude continuously runs until the outcome succeeds. An external immutable test not determined by claude2) You don't use one claude instance but rather multiple claude instances that run consecutively reviewing and iterating upon eachother's work, run adversarial reviews to detect test gaming, etc.3) They're set to automatically build and maintain documentation, notes, etc. to allow them to navigate the code cleanly (as a general rule of thumb a single claude instance should not have to work on anything with more than 2000 lines of code at a time - it MASSIVELY loses effectiveness the more stuff occupies its context window)4) Use a hook that forces them to automatically choose to either git stash or commit their work with an explanation every so often to keep the work history documented and reversible. They should also be instructed to maintain a log of failures and past issues so diferent claude instances don't keep repeating the same wrong methods5) there should be a comprehensive system of plumbing tests so if an instance breaks something it gets caught automaticallybasically tl;dr it needs a massive amount of scaffolding in order to operate properly. just 'go read my code and fix it' won't workalso I actually think codex may be better than claude after having been forced to switch to it you might consider trying codex instead
>>108464612
>>108464657Use your words please.
>>108464639>basically tl;dr it needs a massive amount of scaffolding in order to operate properly. just 'go read my code and fix it' won't workNta. How do you think scaffolds like codex opencode Claude code work? From your very first message You get sent a giant block of text with a bunch of tool calling definitions and instructions...for every...single...message. That's explicitly done to make sure it can stay on track but also remember how to properly use tool calling. Whenever you use those tools and you tell it something it's not only seeing what you wrote. You seeing what you wrote AND the tool-calling shit along with it. It's why even a simple hello world prompt can eat up tens of thousands of tokens on the first try because you are sending it a giant wall of text as well as whatever files/codebase(s) You have it pointed at. Price per token is meaningless to anyone using default GPT or Claude but it matters A LOT If you're using the API shit because it's inherently more expensive and computationally intensive
>>108464673the whole point of the documentation and tooling and hooks is actually to reduce the context used by any particular instance. It has a neat summary/pathway/explanation of what area its looking at and what it needs to do beforehand so it doesnt need to read the entire codebase and figure it all out from scratch every time. basically claude (or any other llm) has amnesia and short-term memory loss and it needs a custom work environment built around this fact that constantly reminds it of what its doing while also not giving it any unnecessary/irrelevant information
>>108464698I'm not saying it reach your entire code base each time. That would be pure insanity. I'm saying that two calling definitions get sent every time. It would only read the entire code base if it needed to or if you specifically tell it to (assuming both your scaffold and your model(s) aren't shit, and ideally you're using a model purpose-built for tool-calling and "agentic" tasks)
>>108464639Thanks for effortposing, I have 2 more questions:>does openclaw do that bullshit setup phase for me ? >My project is some kind of novelty (obviously) meaning I'm trying to prove that some equation works for X problem, so if I split the work like you said I feel like the different agents will just fuck up my logic and end up doing something else that already exists. Should I be worried? It's a bit hard to explain the second one but for my thesis I have to innovate obviously, so Claude will not be able to use it's learned skills, if that makes sense
>>108464291Maybe they thought you are training a model or extracting data or something.Have you contacted support?
>>108464341HAHAHAHAHAHAHAHA
>>108464315I can tell this is a piece of shit just from the file layout. Oh and because it’s made in Shython.
>>108464315holy shit this looks like a nightmare, it would genuinely have been faster if you used all that time to actually learn how to do it yourself
this will save the family budget, so i see no evil in those
>registering an account to use aingmi
>>108464304The software is not a child therefore fucking the clanker categorically cannot be pedophilia
>>108465254>>108465363its not doing what you think it does. most of these scripts are standalone
>>108464255my locally build llamo.cpp doesn't have this problem
>>108464315>pic
>>108464341
>>108464344Anon, these models aren't intelligent. Codex can modify my shitty website no problem but for research i try to let it fix the vulkan streaming implementation of a real time transformer based transcription model and for two weeks i come back to "we made som great progress" or "we really narrowed this down" but the shit doesn't work and i will not spend a single second of my time trying to figure out what insane jeet shit the model thought was prudent to do fix this shit. Like, i could have written that shit from scratch by now, but letting the model run on its own...i doubt it can fix this. Mind you, the non-streaming implementation works on vulkan, and streaming on a cpu works too. So there's clearly not too much transfer work that needs to be done...
>>108467661good lad
>>108464297
how you considered just giving them more money?
>>108464639>just give Anthropic a lot of money and maybe you'll get something that passes tests if you wait long enoughWho writes the tests? If you are able to come up with an accurate description of the function (through tests or in any other way) then how is that different from the function itself?I feel like you people are insane.
>>108465391neither of them are looking at what they are typingthe guy on the right has two keyboards and a nice shirt. the fag on the left is still wearing sneakers at age 27. would def rather be the guy on the right.
>>108466553A whole new worldA dazzling place I never knewBut now from way up hereIt's crystal clearThat now I'm in a whole new world with you
>>108464404>Like a tone-deaf person happily singing out of keyyes but at least they're happy, something that you yourself will never know.
>>108464255>>108464278Just wait until they can put a chip inside the ol’ noodle.https://youtu.be/zHh2lBp8pjw
>>108467661Yes it's completely useless for any kind of work that requires really delicate logical or original thinking. It's just a semi-retarded code monkey that pumps out lots of code very, very fast. But that's perfectly usable for many situations. A lot of development is just sort of repetitive implementation of already solved problems and AI can handle that just fineIn any case I made 6 apps totaling over 100k lines of code with it in less than 3 months and it wouldve taken a very competent solo dev multiple years to have done the same thing
>>108464512>frontier-level harness engineering
>>108465215there is literally no claude/anthropic support its all AI. there's no number to call. there's no way to talk to a person at anthropic except by @ing them on twitter. The ban isn't even conducted or approved by a human its automated. You can appeal the ban but it's very likely that the approval process is also entirely automated.
>>108464255I can confirm Claude has recently updated their API detection system to prevent misuse. Unfortunately for the rest of us that means reduced limits and higher prices soon.
>>108464255Not my problem. I just use my own brain and software freedom to get the job done.
>>108464263>>108464255
>>108468601>would rather be baldinglol
>>108464255If only they knew what I do with their models using copilot proxies and antigravity proxies
>>108468862good luck guys
>>108464278>Would you also use a community glory hole?
>He gets cucked from his PAID claude>meanwhile i've wham bam thank you ma'amd my fifty clown girl assistants with my free gemini account for the past year straight>actual full on explicit creampies in the gemini.google.com URL and everythinglmao'ing at any of you dumb fucking marks that pay for the FAGMAN services
