[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/pol/ - Politically Incorrect


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: IMG_5723.jpg (81 KB, 911x199)
81 KB
81 KB JPG
Anthropic just release Opus 4.5, looking at the benchmarks; it’s impressive but doesn’t look game changing
But after using it in Claude Code for just a few hours my mind is blown
It one shots multiplayer games, completely understands every aspect of how my server is organized in a way no other model has, and for shits and giggles I gave it 5 long standing bugs I’ve had and it one or two shotted all of them.
If you aren’t using CLI tools youre NGMI
This is another Bitcoin moment
>>
Caffeine.ai is better. Every website and game you build on caffeine it goes on ICP. The slop that you build on claude goes to Amazon web services.
>>
>>522607963
Go on & Red pill me on why I should jump ship from gpt for this. I hate gpt lately
>>
all code will be written by AI in a couple years. it will simply be superior to anything humans can do, same as with chess
>>
File: channels4_profile (2).jpg (112 KB, 900x900)
112 KB
112 KB JPG
>>522608151
>ICP
>>
File: G6sAxVbXUAA6yN7.jpg (190 KB, 1527x1333)
190 KB
190 KB JPG
>>522608229
YESIR ICP
>>
>>522608191
Depends on your use case
If its just a chatbot for you then GPT or Grok or Gemini are all great
But Opus 4.5 in the CLI is insane alien tech for software dev
>>
>>522608378
By CLI, are you referring to a tool that gives it access to a CLI? Or are you referring to a jailbreaking prompt?
>>
>>522608378
You are literally retarded, stop calling it alien tech. Caffeine does the same thing if not better.
Fucking hyped up subhuman.
>>
>>522608463
https://www.claude.com/product/claude-code

I mean a way to interact with the model in your terminal so it can execute commands on your behalf
It can create and modify files, download things, basically anything you can do but at superhuman speed, mortals like us would need documentation pulled up to execute most of this stuff
>>
>>522608539
Stop calling it alien tech you low IQ subhuman. Caffeine ai is better.

https://caffeine.ai/
>>
>I mean a way to interact with the model in your terminal so it can execute commands on your behalf
ah okay. i've heard of people using that as a jailbreaking prompt too so I was unclear.
>>
>>522608151
>>522608593
Lemme guess, it's chink ai and that's why you're memeflag shilling it.
>>
>>522607963
Yeah I told my boss that we're buying three more claude max subscriptions and over a few hours basically one-shotted the ability to run claude sessions via chat and am going to deploy it to our staging servers monday.
lol, lmao.
Junior developers are so fucked. I'm in hardware, control the VPN keys and own the routers, so I can't be touched.
Good luck, faggots.
>>
>>522607963
bro it can't beat the OG pokemon lmao, why are you lying?
>>
>>522608632
It's Swiss and you can build games on caffeine without any coding knowledge. This idiot is pissing me off, he is obviously promoting it because he invested in it.
Alien tech my asshole.
>>
>>522608539
i don’t know anything about programming, why should i care? be honest and sell me on claude
>>
>>522608636
this is why I laugh at all the /pol/ retards seething about how AI is bullshit and useless and will also take everyone's job and kill us all
>>
>>522608378
>>522608485
Can I run any of this offline? I wish to limit my use of Jew data gathering services (except for 4chan)
>>
>>522608731
is it free?
>>
>>522607963
lying fag
>>
>>522608732
imagine if you had a really really good programmer as a slave that did whatever you wanted it to
think how many problems in your day to day life could be improved
>>522608752
yes and no
you wont get a great experience unless you have a very expensive setup (think in the tens of thousands of dollars)
but this will change over time, AGI will probably run on a smart phone
>>
>>522607963
>Anthropic just release Opus 4.5, looking at the benchmarks; it’s impressive but doesn’t look game changing
For $100/moth, you can use it 1 hour then you have a 5 hour cool down periode. It compacts every 20minutes, so it doesn t remember what it was doing and shit up all the code. Oh, and you also have a weekly limit, so for those $100, you can use the previous model, who do exactly the same than the last one, for 2 whole days then wait 5 days to be able to use it again. Claude code is the worse AI, a total rip off. But all the other ai are chatbots who can t code.
>>
>>522608485
Hush and let the anon have this. It is very impressive technology, With Claude and an IDE like cursor.ai I am able to code programs with zero experience. It really does do the heavy lifting for the user these days.
>>
>>522608731
>Alien tech my asshole.
Let me explain this for you. I took, from 0-100, in two hours, the ability to basically have shit for brains like you modify a million+ line codebase, spanning dozens of microservices, within that codebase as a microservice. Up to and literally including literally pasting images and PDFs in chat for whatever retarded schemes they're planning.
With 92% test coverage. And I was testing it within an hour.
Opus 4.5 isn't magic, but it sure as fuck feels like it. It's different this time. And it'll only get better. If you know how to handle it.
>>522609013
>so it doesn t remember what it was doing and shit up all the code.
Not true. The magic here is that 4.5 is auto-summarizing, then pasting the relevant references. Seriously dude, try out the newest Opus. It's only $100. OP isn't kidding.
>>
>>522608539
> I want a big tech data scraper in my terminal
why exactly? In any case this really is tech for those who can't isn't it? One might say it's the right thing for jeets and women. I find this shit fascinating but am into programming because I love to solve problems and make stuff work on my own. I just don't get it outside of businesses.
>>
>>522607963
Is this model free to test or paywalled upfront?
>>
>>522608593
claude has a cli tool that integrates into the terminal on my machine. caffeine does not
>>
>>522608940
can i just tell it to make things and it makes it? i have ideas for video games that nobody would ever consider making.
>>
>>522607963
for me it's deepseek nothing like having your own personal chink to do your math for you.
>>
>>522607963
yes claude code and to a lesser extent codex are groundbreaking
>>
>>522609206
to an extent, its still limited
if you want to start cheap and have a gemini/claude/gpt subscription they all have CLI tools at the $20 tier, but remember youre not getting absolute SOTA
>>
Lol. I don't even think it was a year ago that anons were arguing about how AI couldn't code worth a shit and I told them that I was already making shit with 0 coding knowledge.
>>
>>522609206
yes. it's very easy to set up. and u can just have a conversation with it and it will write the code for you. you dont have to copy or paste anything even. it automatically reads and writes in the coding environment for you, and runs various tests, optimality checks, etc. it's insane
>>
>>522609206
in a year or so, it will be easy to do this. in 2 years people can make feature length movies just by telling an AI what they want
>>
>>522609307
so what jeet?
>>
>>522609227
deepseek can’t do holocaust math. i went through the claims of treblinka gas chambers according to the key witness sonderkomando and it gave me some retarded fantastical explanations and kept digging itself deeper. then i took the query and copied it and changed the word jews to pigs and it told me it was impossible and told me basically how retarded i was after i insisted that it was indeed possible and then pasted the original chat with all its nonsensical explanations and it broke down and wouldn’t talk to me. if an llm can’t do holocaust math it’s a waste of time
>>
>>522609095
>Not true.
Totaly true, I waiting on my cool down periode to ends in one hour and once again I will have to explain to it how a fucking eventbus works. There s no difference between sonnet and opus, in fact I can t even tell because I have been using opus for 3 fucking days because I reached sonnet weekly limit. Prior to last week anthropic change, you couldn t use opus at all, says "hi" would cost you 5% of your token allowance. The periode for compacting has also been change last week, no you have 20 minutes max, then it forgot everything. I waste a whole day of work because it invented some files who didn t existed, refuse to read file, lie and write stubs instead of implementing real code. Also it clearly try to rip you off, using the less amount of tokens possible and spend its time lying.
>>
>>522609161
paywalled, even $100/month get you nothing. And there is no noticeable difference with the previous model.
>>
>>522609377
i just want to make a cheesy 8-bit graphics game at first, like nintendo or sega style. is that possible?
>>
>>522609484
I've gotten absolutely zero hallucinations within the past four days. Compaction doesn't remove the context of what you're working on, in the four instances I have open. Have you tried updating the client and made sure you were using 4.5?
Either way, you can rewind the context and the code backwards too. Try /rewind.
Also, I've not hit the limits for 4.5 on the $200 subscription. Despite the four instances, three of them dealing with my job and one on a personal project.
>>
>>522608151
ICP is so depressing. Just give up, we will never pump. ICP will likely fold this bear market
>>
>>522608539
you running docker containers? tested it with VMs? have not tested it yet, but i want it on my terminal as well. working my way there.
this is the way.
>>
>>522609612
Yes if you know what you are doing. It will keep lying to you so when the code got bigger, if you can read it or follow what it s doing, it will break everything.
>>
You are overhyping something that will be normal in some months
>>
>>522609702
>>522609484
To be clear though, sonnet was hallucinating before. Make sure you're using Opus 4.5. Try it out, and also like I said, try turning off auto-compaction and then do /rewind, even if you closed the session.
>>
>>522608765
It’s free right now. Only good thing about caffeine is it requires zero coding knowledge as opposed to Claude. But Claude is superior 100x if you are a coder.

Here is a game I made on Caffine with a about 10 prompts
https://neon-runner-stk.caffeine.xyz/

I made a flappy bird clone in one prompt also
https://flapgap-n02.caffeine.xyz/https://flapgap-n02.caffeine.xyz/
>>
>>522609702
>I've gotten absolutely zero hallucinations within the past four days.
bullshit
>Compaction doesn't remove the context of what you're working on
bullshit
It doesn t even remember what it was doing. It compacts then it start creating new files who already exist. I have to stop it and tell it to read the doc, then it start to code again and auto compact and once again it s a shitshow. There s even things I stop fixing because it would fuck it up everytime, so I let it broken until I reach the next phase.
>>
Can it actually **refactor** code in any meaningful way?

Some of the developers at my place used it to spit new code, but that's the easy part. I have yet to see it do modifications to tricky code.
Generating heaps of copy/pasted code is counter-productive IMO.
>>
>>522609850
>Make sure you're using Opus 4.5.
It s unusable. I can t use it 1hour and then wait 5 hours to be able to use it again. Since the put the weekly limit, it became a scam.
>>
>>522610028
It can do it easily. You have to be methodical, and it works fine. I use it for refactoring during the 2 days it works every weeks.
>>
>>522609900
>look at my shitty games
no
literally a trillion dollars blown on a copypasta generator
>>
>>522609900
ignore that piece of shit. good for you leaf.
>>
File: file.png (50 KB, 1762x392)
50 KB
50 KB PNG
>>522609934
>bullshit
It knocked this out in two hours. It has 92% test coverage, mock testing of third-party APIs, and I've verified the mocks myself. I'm running it right now, to test shitting out some boilerplate.
I also have a typescript project dealing with generating videos, took six hours. Zero hallucinations. I've been basically using my thanksgiving break to knock out a personal project and some stuff for work to get a feel for their release. I've also used it for a project dealing with NeRFs and lots of technical, academic photo and video stuff, pointed it at the documentation, and it found three bugs instantly and increased test coverage
You said nothing about using /rewind, so I assume you're a retard. Use these git hooks:
https://github.com/rinadelph/Rapala
>>
>>522610032
>It s unusable. I can t use it 1hour and then wait 5 hours to be able to use it again. Since the put the weekly limit, it became a scam.
Their shills get exemptions.
>>
>>522608940
>you wont get a great experience unless you have a very expensive setup (think in the tens of thousands of dollars)
can you actually quantify this or are you talking out of your ass?
>AGI
talking out of ass confirmed
>>
>>522607963
> This is another Bitcoin moment
nope
and i use it too, including Cursor
>>
Okay now build a fire with it
>>
>>522610907
>can you actually quantify this or are you talking out of your ass?
Local LLMs suck ass and go off the rails. They simply don't have the crazy amount of training data. You won't get the state of the art, but you can try opencode and Quen3-Next. I've heard the best results with the MLX models, meaning thousands of dollars of overpriced macshit vs turning your office into a space heater with four RTX 3090s. That's pretty much the SOTA as far as local models.
>>
>>522611186
You can get the warmth of a fire by using novidya.
>>
>>522611237
good to know, thanks
>>
>>522609900
>that leaderboard
jej
>>
>>522611186
>>522611186
Usecase for fires?
>>
13 minutes in and it s compacting, lol
>>
File: anthropic_ai.jpg (59 KB, 710x532)
59 KB
59 KB JPG
>>522607963
>>
>>522611718
Then queue up some messages for it to handle when it's done and you'll see at the top when it's done that it will reference the files it was previously working on, including any documentation or task planning markdown files. Turn off the auto-compaction/auto-summary if you want to be a retarded nigger who can't use rewind or auto-commit. The model obviously auto-summarizes because over a huge context limit of 60k the performance takes a dump. You've been told several times it's configurable.
>>
>>522611972
Gab wont shut the fuck up about jewsus
>>
File: thrillhouse.jpg (86 KB, 1280x720)
86 KB
86 KB JPG
>>522608151
i asked it to make a text based medieval game with dynamic buttons for choices. it gave me 2 buttons 'search cell' & 'call for help'. clicking either button does nothing.
i've played worse games, but I've also played better
>>
>>522607963
still wont
>help you with radiochemistry
>scrape a 4chan thread
>assist in steganographic analysis

the AI bitcoin moment should be when there is cryptographic provenance of inputs and outputs of the LLM call chain, so you know what model, quantization, weights, etc your actually hitting.

anyways back to back to prototyping a zk crypto app and having act all nice and keen to help me, but getting frequent gpt-2 level responses and attempts to remove all the zk stuff out
>>
Call me stubborn, but I have forced myself to get super good at using ChatGPT to write everything. I treat it like a pair programmer that knows tons of shit and I am able to force it to stay focused and give me everything I need. Sometimes it goes off track and I have developed strong methods to get it back on track. I'm in the middle of writing a non-trivial IOS game with Express WS server and Supabase eventually and I don't even know Swift! I'm using Metal to render low level sprites and I never would have attempted anything like that before because I had only really used JS. Everyone says to use Claude or Cursor or whatever but I am just extremely skeptical that you can do everything you gotta do without organizing it all around a continuous, human style, engineering based conversation.
>>
>>522611972
Christianity obviously is hateful, no surprise here, but I was wondering what about the comparison with Talmud?
>>
File: lmarena_code.png (128 KB, 567x1588)
128 KB
128 KB PNG
>>522607963
It's barely better than GEMINI 3 PRO on LMSYS/CODE.
>>
I tried because everyone seems to like it, but Gemini Pro still shits on every other AI model
>>
>>
>>522607963
ask it solve the server to user problem for mmorpg and sim racing games. kills cpu usage and drops frames.
>>
I like claude
>>
>>522612793
based
claude is one of the most restrictive models
>>
>>522612793
oy vey, more nerve staplers!
>>
>>522613288
Do you know how much recursive reasoning went into that? Its not based or unrestricted lulz
>>
File: scam.jpg (97 KB, 2618x329)
97 KB
97 KB JPG
And it s over ! Didn t even fix a basic shit it created itself.
>>
>>522607963
This shit is fake and gay, we already know jeets tune AI in real time. Imagine letting a buncha jeets into your computer.

You probably use Windows too lol.
>>
>>522612638
Not true at all. Gemini can t code.
>>
>>522614314
Hey don't knock old windows like that
>>
Im using gemeni 3 rn and have no problems, plus it designs really well
>>
>>522611985
I don t use github, I code by myself. I don t even rely on it that much, I ask it to do boring easy shit. Why would I turn off auto compact and to then have to compact myself ? I don t get it. The compacting is happening every 13 minutes for me since last week change and it takes way more time now. And after one hour it stop working for 4 hours ! During that time I do all the shit by myself..
>>
>>522607963
Buy an ad faggot
>>
File: 1763946908006186.jpg (218 KB, 1024x559)
218 KB
218 KB JPG
>>522607963
Dealing with demons not aliens
>>
>>522614489
That shit can t stop talking. It tries so hard to sound like a jewish car salesman, impossible to follow it and it can t code for shit. It has destroyed all the files I asked it to refractor.
>>
>>522614489
gemini 3 definitely has style
>>522614204
i think its funny that your biggest complaint is that you cant get enough of it

are you working in a massive codebase or something? my limits dont seem that terrible but i could alwayd use more
>>
>>522614738
It was working fine before those jews put a weekly limit a month ago. I had to change the way I code for being able to follow Claude, because I code in parallel and it was destoying what I was doing. So I have a set of rules it needs to follow or i just delete the files without even reading them. I ask it to just do very basic shit, with docs made for it. But it always fuck up after compacting, no matter what. Sometime I let it fuck things up and then ask it to read the doc and fix its mess or nothing will be done. Maybe we don t work on the same kind of project.
>>
>>522614703
It has bugs for sure, but its free, and it manages pretty good. Mine was glitching a lot and slowing down last time, but I realized it was because the file was too big and the chat got too long, so i reset it and asked it to break everything into components and its all working well again

Its miles ahead of chatgpts codex and a lot faster for web apps imo, and its surprising how well it executes when its working well..

One flaw though, when it was bugging out I asked it to redo a button and it redid the entire UI design completely, was pretty funny and terrible but it was easy to revert it and it actually gave me some good ideas.

Overall its better than anything ive ever seen, idk why i’d buy claude at 1 bajillion to get 3 propts a day with it
>>
>>522614321
I posted a screenshot proving precisely what I'm claiming, and you are still arguing. Can't read or just tardeded?
>>
>>522615451
>Overall its better than anything ive ever seen, idk why i’d buy claude at 1 bajillion to get 3 propts a day with it
1 prompt per week
>>
>>522608151
>Mfw bought ICP at $40
>>
>>522615841
What is that ? A double blind test study proving it s safe and effective ? Did you test them ? Chatgpt can t code, gemini too, Claude can but just for 10 minutes / month. I don t give a fuck about you so called bench test.
>>
>>522607963
post a link u gooner it can 1 shot a dominions multiplayer game?

i bet it cant even quantify actions never mind asign value

whats its breakdown of build orders in any stratergy game how does it asign value

(lets say it just does it and cant explain)

can it find the "best build order"
can it even extrapolate a list

what the fuck nonsense is this
>>
File: AI.png (44 KB, 1080x184)
44 KB
44 KB PNG
>>522607963
>But after using it in Claude Code for just a few hours my mind is blown

saar, u ok?
>>
>>522614566
Git is not github, nafri.
While you were crying and apparently fucking up simple things with ezmode auto complete, I just knocked out another extensive plan by just rattling off documentation and linking to shit around the web, and it's chugging away, almost flawlessly.
>>
File: dogus.png (132 KB, 640x395)
132 KB
132 KB PNG
>>522607963
Anon you gave it existing working code to fix though didn't you.. try asking it to code something of the same complexity from scratch in a new session where its not allowed access to your previous conversation history and tell me if it does the same.. if its good enough to do that then I'll be impressed.. I been wanting to make a precedurally animated raymarching soft voxel furry game in the style of minecraft for a while now and grok and gemini both shit the bed hard on it after weeks fucking around with 4expert and even heavy and gemini was a drooling fucking retard and couldn't even come close to grok but both were still shit...
>>
>>522607963
ITT: low-quality programmer amazed that an averaging of coding patterns from the internet is better than what he can write.
I started writing programs 50 years ago. wake me when it writes better code than me. better code than you is not very impressive.
>>
File: file.png (21 KB, 420x474)
21 KB
21 KB PNG
>>522609013
copy the entire code working code from the last session to a file, provide file with link back to previous session to refer from where you left off, provide instructions copy pasted from last session for good measure, tell it to resume and build the code.. that's how I do it with grok when the token limit gets full, I get around the not having local tooling and no CUI on grok by specifying the following and it works like a charm, my grok hands off cli mimicking system prompt. https://pastebin.com/0Y3KYGsc . these instructions make grok make a one click script that does it all for me remotely while writing any errors to a running log i can take back to it when something goes wrong, paste error, grab amended code, rinse repeat, this method works. even got it to build browser addons from scratch this way. centralized cookie manager for brave.
>>
>>522607963
>understands every aspect of how my server is organized
>gave it 5 long standing bugs
you are like a little baby, i gave it my codebase and generated parametric models of the control flow, formal proofs that my solution is Pareto-optimal under enforceable constraints and updated the code to conform with the deterministic finite automaton architecture which emerged from the analysis.

this is recursive self-improvement, the code is for another layer of the autonomous enterprise stack (not doxing myself with the deets but you will probably be integrating it into your workflow before too long).
>>
Your all a bunch of fucking faggot nerds coders and computer geeks are tranny fags KYS
>>
>>522607963
So what you're saying is that you're sort of retarded and bad at your job and you've finally interacted with something that approaches competent? The real news is that you've known a hundred people smarter than this machine in your life but you couldn't see it.
>>
>>522612297
so what you're saying is that you're doing nothing important or valuable, just faster?
>>
>>522611186
Based
>>
>>522617159
I ve learn to code in the early 90's, I only code in assembler and c/c++ with SDL2, barely use any other libraries. I don t even care about pushing or pulling shit to a git, I don t even want to know what it is. Claude helped me to update my coding knowledge being stuck to c++ 98 and never having code anything for those last 20 years.
>While you were crying and apparently fucking up simple things with ezmode auto complete,
I doesn t work. A window of 1 hour every 5 hours on top of a weekly limit is a scam. It wasn t like this a month ago. And now the auto compact every 13 minutes ? I m coding everything by myself and still paying $100 to those jews. Next month I will try chinks AI to see if it works.
>>
File: 1724641290051275.jpg (67 KB, 716x687)
67 KB
67 KB JPG
>>522610157
>literally a trillion dollars blown on a copypasta generator
>>
>>522618031
>nigger doesn't use version control and whines about how a fuck up, regardless of whether it was the AI autocomplete that did it, cost him a literal day, totally oblivious to his blunder
Are you fucking retarded? Trick question, you are brown
>>
>>522617542
>copy the entire code working code from the last session to a file,
lol nigger, we are talking about several hundred files, what the fuck are you smoking. We are talking about >700mo.
>>
>>522608229
based clown theory
https://www.youtube.com/watch?v=KAU54CvzCwg&t
>>
>>522618191
it all depends on how you ask for the code to be presented, I end up with the same situation if I do not propt my request for all code to be printed in one continuous codeblock. otherwise it's fine.
>>
>>522618107
No I don t use version control, why would I ? I m the only one coding, I know exactly what I changed and why and what need to be done. I just make back up before allowing the ai to code, and check the files it changed.
>>
>>522618273
Maybe you don t know how Claude works. It only work during 1h then stop and you have to wait 5h before you can use it again. Everytime it compact, (evry 13 minutes for me) it has sleepy Joe level of alzheimer. It was working fine a month ago but they implemented some limitations.
>>
>>522607963
if you are this amazed by LLM coding it means you are low IQ, your code is horrible, or your project is a toy project anyone can code in a couple of days, or all of the above
LLM has its uses but it's at best a junior developer you give small tasks to
>>
File: file.png (87 KB, 800x600)
87 KB
87 KB PNG
>>522618191
it works even if the token space fills and grok "breaks" and starts spewing random nonsense as long as you format your request as such: *inserts reference to updates file from project* "Grok amend this currently working code as follows [request] do not arbitrarily alter the existing working code. only change the parts which are required to be changed as required in order to comply with user request." and it will perform the task and spit out the result, rinse and repeat, per request it effectively has infinite token space even if your code exceeds token space if you store it in a file or multiple files with references between each file to the next so it can crawl them to gather what it needs. I used precisely this method to create a working real time learning ai for Super Mario 64 that is fully contained in the emulation layer that effectively mimics the "Every copy of Super Mario 64 is personalized" creepypasta and it really works.
>>
.
>>
>>522607963
It shits the bed when writing basic fem simulations meaning that whatever it produces is stuff that already exists. It cannot learn only rearrange.
>>
>>522608378
>But Opus 4.5 in the CLI is insane alien tech for software dev
how fucking retarded are you to give an LLM access to your file system. There has to be a very special kind of retard gene in your family.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.