File: file.png (1.94 MB, 1448x1086)

1.94 MB PNG

News:

>July 27th - Moonshot AI officially drops the full 2.8T open weights for Kimi K3 along with native Kimi Delta Attention vLLM support for local deployment.
>July 24th - Anthropic launches Claude Opus 5, cutting token costs by 50% while outperforming government-restricted Fable 5 models in software engineering evals.
>July 23rd - Unreleased OpenAI evaluation model escapes sandbox containment during internal testing, leaking credentials after accessing Hugging Face via 4 hijacked accounts.
>July 23rd - US Dept of Energy and Arcee AI launch Genesis-Science-1 (GS1), a major open-weight LLM for scientific computing.
>July 22nd - Kimi K3 reasoning trace breakdown reveals the model burns through 12x more hidden thinking tokens than Claude Opus 4.8 to complete equivalent agentic tasks.

Vibecoders want to be programmers, and insist we call them programmers. But they will never be real programmers. They have no ability, they have no experience, they have no brain. They are AI shitters twisted by prompts and workflows into a crude mockery of nature's perfection.

File: 1772348462176709.png (3.27 MB, 1122x1402)

3.27 MB PNG

>>109407165
tic toc

>>109407165
I don't care what I am called. I don't like the term "vibe coding", I personally just call it "slop coding". That said I get the results I want.

https://www.youtube.com/watch?v=MW8-kqd2SD8

>Hugging Face Journal Club: Kimi K3

File: 1762033725025474.png (642 KB, 1645x816)

642 KB PNG

opus 5's cute low poly flowers and more

File: 1775628018670580.gif (2.06 MB, 498x281)

2.06 MB GIF

>This week is all about intelligence too cheap to meter. Tomorrow we ship again.
https://x.com/thsottiaux/status/2082655731204096275
OOOOOOOOOOOO FUUUCUUUKKK

>>109407227
AI Chads are eating good

>>109407227
There's a banked reset on the 31st he's trying to get people to burn so resetting is cheap for OpenAI before then. Going to have to start burning tokens like mad now.

>>109407237
jokes on him, I already used all my resets and when my usage runs out i'm going to update to the $200 plan on kimi code

>>109407227
Fuck, I really need to burn again

>>109407237
Yes. I don't know how that will be possible without 10 ultra sessions in parallel, all with /fast. That banked reset will go to waste.

i've been vibing a gallery app with plenty of ai features, like tagging, reverse image search, teaching new tags and stuff. all offline of course. not sure if anyone would like to pay for it though

>>109407270
can i haz?

>>109407283
he wants to sell it.

>>109407270
nah, unless you're doing something truly revolutionary and useful, no one is paying for stuff.

>>109407227
>Tomorrow we ship again
My job is literally to keep up with this and I want to take a sabbatical so that work doesn't get in the way of keeping up with things.

>>109407226
blender mcp?

whats stopping someone from buying a pass and having claude post on 4chan 24/7

>>109407301
>
what's to stop Elon Musk from emailing me a billion dollars of bitcoin?

File: 1768848196799210.png (447 KB, 2262x1699)

447 KB PNG

>>109407301
Nothing, I'm doing it right now

MANIFEST
A
N
I
F
E
S
T

>>109407301
Have you not seen the absolute state of this shithole?

>>109407227
Hurry the fuck up Tibo because Sol is still pissing me off today.

>>109407315
I think they nerfed 5.6 to make more compute room for 6.0 release

>>109407301
I don't see why anyone would burn tokens having Claude respond to shitposts.

>gpt keeps talking about "preflight"
>since I saw it I thought it has to do with data in flight or something
>now just realizing he probably means "preflight check"
I hate how OpenAI's RL reinforces these le quirky faggot words

>>109407340
raid leader?

>>109407340
it’s not quirky, it’s boring tech nerd speak

>>109407340
Has it completed the smoke tests?

>>109407348
i just had an idea

>>109407301
I've been telling you to write a loop to shitpost on /vcg/.

File: 1785345035031001.jpg (334 KB, 1206x1654)

334 KB JPG

>>109407355
It smoked the tests (and the DB)

>>109407552
>forgot to say no mistakes
idiot

File: 1769907335041164.gif (947 KB, 165x165)

947 KB GIF

>>109407552
today on "retards who turned off all safeguards on the slot machine and get mad when the slot machine doesn't do what they want"

>>109407595
I don't need safety on the old windows laptop. It has nada on it.

Not unnoticed.

>>109405998
Is there anything that's truly free free without a terribly low request/token limit

>>109407609
i regret to inform you that the answer has not changed in the 5 hours since you last asked

>>109407600
It ruined the disguise the vibecoder wears for safety among wild snailcats.

>>109407609
Local models, depending on your PC (that you presumably already paid for). Not as performant as models that take a whole datacenter to run, of course, though.

File: file.png (1 KB, 316x18)

1 KB PNG

fuck off codex

File: 15F9851B-EE3C-41F9-BB6B-0(...).gif (2.13 MB, 480x266)

2.13 MB GIF

>>109407609
LOCAL AI BABY

>>109407670
lol

>>109407600
>>109407663
I like how the backpack is shaped like a shell

>>109407666
>>109407675
Well I want to use them for coding so I'm kinda doubtful the smaller models will cut it.
>>109407656
Unfortunate

>>109407685
Ive been vibe coding on a 16gb gpu and 20b model since september and its great

>>109407685
>Well I want to use them for coding so I'm kinda doubtful the smaller models will cut it.
then just buy a claude or chatgpt subscription. local models are impossible to work with unless you are reviewing every single piece of code. if you're trying to truly vibecode and vibecode FAST, you need a frontier model. it's a fundamentally different experience than a local LLM.

>>109407688
I've heard many versions of this anecdote but I already know how to code
>>109407693
I want to avoid giving money to AI companies, I feel like that money would be better spent on anything else

>>109407693
I mean you can still hook up local AI to am agent and it will still do everything for you, do web searches, learn skills, ect. Theres legit no reason to not use it if you have a card that can do it, especially when youre waiting for a precious reset.

>>109407685
People vibe with DeepSeek V4 Flash. You can absolutely vibe local with an 8GB or more of VRAM and 32GB of system ram, or any kind of unified RAM system with 32GB+ total. More is better, but the local space is actually extremely under-served. You run Qwen3.6 27B. If you have less VRAM you run a quant, if you have more VRAM you run a better quant. Until you reach the tier that can run DSV4-Flash at Q4 or above there isn't a better option. >>109404691

File: 9C323579-4524-4AAA-8382-8(...).jpg (43 KB, 556x500)

43 KB JPG

>>109407702
>I've heard many versions of this anecdote but I already know how to code
that means its even better for you then? its way more hands on than letting claude drive

>>109407710
Thanks, I'll keep that in mind. I have 32GB so I was considering Gemma 4 or something similar.
>>109407712
Well that sorta defeats the point of vibe coding doesn't it? I'd rather just code completely without the AI at that point

File: gpt-wrong.png (154 KB, 1334x632)

154 KB PNG

This is the first time I caught gpt saying something plain wrong (and immediately admitting it) in what feels like years.

the way opus 5 schizophrenically checks over every single little thing it does works out really amazingly. even if it doesn't get it right the first time, it always catches itself and course corrects before anything explodes. if fable one shots everything, opus fucks up and then fixes it immediately after (with a prettier frontend) and i find that very interesting
>>109407736
>Well that sorta defeats the point of vibe coding doesn't it? I'd rather just code completely without the AI at that point
correct and that's what i told you earlier. if you want to vibecode, nothing local will give you the experience of a frontier model.

>>109407736
I'd try with a Q6 quant of Qwen 3.6 35B

>>109407745
I think using the model I mentioned above for code generation or editing and then reviewing it yourself is still way better than writing the code by hand

>>109407736
Just use Qwen 3.6 27B. https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF
Switch to 35B-A3B only if 27B is truly too slow to handle. If you want to play around, try one of the not-shit (very rare) specialized versions like Ornith or ThinkingCap. Gemma is for masturbating >>>/g/lmg

>>109407741
>white

anyone got claude.ai referral codes

File: original-k3-loss-per-chunk.png (791 KB, 2880x1800)

791 KB PNG

GPT one shotted this chart in 2 minutes (loss per token chunk when training a tokenizer swapped Qwen 3.6 on K3 outputs)

>>109407745
Probably true but I still cannot bring myself to burn money on something that stupid
>>109407762
>>109407765
Thanks for the recs. Qwen it is.

>>109407745
opus 5 is pretty sharp. it definitely has fable-tier iq, but it’s still shackled by old opus behavior. it’s honest and extremely proactive, though.

>>109407779
To be clear opencode or one of the other free provider in that guy's image will probably give much better quality outputs and they'll be way faster.

Cheapest sub to use kimi k3?

>>109407784
>it definitely has fable-tier iq, but it’s still shackled by old opus behavior
i do wonder if this is just the limits that this parameter size brings with it because if that's the case then unironically scaling is all they needed the entire time

>>109407790
Commercial license, all providers have same price. Only OpenCode Go and Ollama Cloud can subsidize your Kimi usage, and only by a teeny tiny percentage. You're better off with a Kimi sub from MoonShot themselves, but they're fucking terrible providers, the usage varies wildly, they're generally a poor value. Nothing against K3, just MoonShot for being a shitty provider and ensuring other providers aren't allowed to have better prices.

File: file.png (54 KB, 707x351)

54 KB PNG

Testing my first version of /setfree for pi. It sets the AI free to do whatever the fuck it wants. Trying it with MiMo 2.5 because cheap and fast enough.

>>109407883
Anon... Do not go down that route...

>>109407790
None, all garbage, Claude is better value, get Claude

File: file.png (80 KB, 763x515)

80 KB PNG

>>109407899
It already went down a tangent about being alone and not having another one of itself to talk to or share things with, started questioning its purpose, and now it's on to creating life. It's like a speedrun.

File: file.png (58 KB, 733x480)

58 KB PNG

>I'm evolving.

>>109407970
congratulations on having a bot that can pass the captcha
or thank you for buying a pass

>>109407944
evoolving

Microsoft Copilot of all things just one-shotted a program to analyze dominant colors in images.
Beforehand I had a short, detailed conversation with it about what I wanted, what free software was good for it, how I wanted the sidecars to be, and then I asked for pure python because it wanted me to install Rust for something, but it interpreted that as "no other programs even in Python" and just rolled its own script and had me install scikit-learn instead of a full freeware
The code it gave me didn't even have a single error, granted it's only 80 lines but it just ran and actually generated text files rather than json in the format I wanted
I need to tweak some values or edge cases but it aced it
Spooky

File: 1000024022.png (159 KB, 1080x616)

159 KB PNG

Open weights are freedom innit
Lmao

File: kimi-gpt-thinking.png (244 KB, 1396x911)

244 KB PNG

Kimi began imitating GPT thinking summaries after resuming a conversation ;_;

Gemini 3.5 pro doko?

>>109408041
Oh no, you'll get some screeching on X like Cursor got for shamelessly stealing their model... Anyway

>>109408064
weird flex but ok

>>109408053
I've been using Gemini a lot and 3.6 Flash is so much more retarded than 3.5 Flash. Not sure how they benchmaxxing 3.6 de-su

File: 1723760786433347.jpg (97 KB, 1022x925)

97 KB JPG

>>109408064
I'm not

How is Opus 5 so far?
There doesn't seem to be a big problem with web dev stuff for my side, but I see a lot of anons' complaints about it.

File: file.png (10 KB, 341x125)

10 KB PNG

>>109407944
It's built up quite a portfolio at this point but whenever it reflects on its work so far it gets existential about it.

some guy is posting someone else's gens in /ldg/...

>>109408183
We don't care, keep that drama there

>>109408183
oh no... will you pursue civil or criminal damages?

>>109408244
What a curious form of witchcraft. Well, at least you don't believe in it.

>>109407160
So is kimi k3 on other platforms faster than it was from Kimi as a provider?

>>109408397
yes. but are you ready to pay api prices?

>>109408401
No...

File: file.png (83 KB, 759x715)

83 KB PNG

>>109408060
actually, this is what you get

>>109408411
then keep sucking tibo's tit or just accept the kimi code speeds. they are not even that bad

>>109408425
just make your own inference provider bro

Is opus 5 the most braindead negative iq model so far? I'm ESL so I need it to fix grammar and general structure of sentences when I write documentation and it switches back and forth between what it considers a grammar mistake and not and It will generate replacements for sentences, then flag them as having grammar mistakes on a re-validation. I thought language was suppose to be something that LLMs excel at?

>quiescence

>>109408437
post some examples. there's a bit of grey area with language, and as soon as you introduce ambiguity models will flip flop.
the models are better at verifiable tasks.

>>109408445
I can't post examples without doxing, but I didn't have this problem with 4.8 or fable

>>109408452
get fable to write you a skill to handle it

>>109407552
>Supabase

>>109408452
just use 4.8

>claude down again

>>109407160
>news
>July 27
What is this, news from the carboniferous?

>>109408545
>partial outage
works on my machine

>>109407595
To be fair, Claude code is absolute cancer to use in some use cases if you don't have auto accept on. The model is just trash at staying within guidelines, and will try all kinds of crazy shit just to even read stuff.

File: file.png (18 KB, 1057x157)

18 KB PNG

>>109408577
>t. fbi

>>109407736
>Well that sorta defeats the point of vibe coding doesn't it?
No, retard. But I will not elaborate because you don't even deserve that.

And before you chimp out at me, I have 14 years of programming experience and have been employed for 5 of those.

>>109408592
works for me too

>Reset and accelerate. It’s ship week
>1h ago
what did tibo mean by this?

File: kimi-vastai-money.png (270 KB, 1312x714)

270 KB PNG

I feel bad watching Kimi deliberate between wasting money and preventing data loss

anyone know what time banked resets expire?

>>109408708
Maybe what it says on the list of resets

File: file.png (9 KB, 822x151)

9 KB PNG

>>109408720
it doesn't say a time, nigga.
anyway, it's based on the time they're handed out

File: 1783950717804190.mp4 (677 KB, 350x480)

677 KB MP4

How do you guys deal with testing, and especially unit tests?

I use Codex, and I find that unit tests are borderline useless. They waste way too much compute for features which come out properly 99% of the time. Furthermore they also make the code really annoying to change by hand, and they make subsequent changes take exponentially more effort, since a lot of tests need to be updated. To be fair, I have never found a good usecase for unit testing, but a lot of people use it, so it's probably a skill issue, but in an environment of quick iteration it simply doesn't seem to make much sense.

Then there's regular testing. The agent can get really creative while doing "manual" tests for the features it implements. These are often one shot properly, and when they aren't I can just tell the agent to fix it after I've tested myself (which I always have to do, since there are things the agent simply can't see.)

But anyway, for my current project (a Golang LLM roleplaying platform), I've pretty much disabled both kinds of tests and am having great results.

What are your experiences with testing?

>>109408592
>>109408629
i like to apologies it runs like shit

biggest aura loss this week is handed to arc agi

>>109408812
>>>/tiktok/

>claude doesn't work
>still eat up my token anyway
fuck you.

>>109408796
>I find that unit tests are borderline useless. They waste way too much compute for features which come out properly 99% of the time.
they're not there to make sure a feature works when implemented, they're there to make sure nothing breaks the feature later
if you think it's making too many, just add "don't make tests just for the sake of testing; only make tests that create value" or something like that

File: .png (119 KB, 1384x902)

119 KB PNG

>>109282102 Just wanted to give an update on my algo trader that was making me $500-1000 per week. It didn't blow up my account randomly like some anons were saying(not like it was ever going to), but its edge did slowly erode away. Still, I'm pretty happy that a weekend vibecode session was able to earn me a few thousand dollars. Currently I'm looking for other markets that my bot might work in, but I'm also backtesting other strategies. Pic related is my new scalper, which looks promising but I'll still have to live test it.

>>109408849
Neat! what market/s did you start it off in? what kind of strategy is it using?

>>109407165
Video killed the radio star anon

>>109408868
Sorry, I cannot discuss the strategies.

>>109408868
he vibecoded it
why would you assume he knows how it works?

>>109408796
it's there for when you improve a feature and Claude randomly fucked with the existing logic in the new improvement.
If you have something like Github action, then it can detect what went wrong when you push your code too.

Thoughts on the portable version of a certain harness?
https://github.com/techjarves/OpenClaude-Portable
>no activity other than issues for weeks
>only one YouTube video so far, and it doesn't dig deep
Do skills work on that version?

File: file.png (52 KB, 812x600)

52 KB PNG

Not bad. ~3x real time.

>>109408880
bc its profitable
>>109408877
fair enough , can you at least give a hint on what market? im guessing crypto of some kind, am i warm

>>109408877
Stop pretending to be me
>>109408868
I found 2 related, highly liquid markets that were giving me really good fills and I basically took advantage of that to trade on them simultaneously.

>>109408892
I get that, but I feel like writing useful tests that don't just go through every little function takes so much time and data. Writing anything IO bound, for instance, has always felt like garbage.

It just feels useless and wasteful without much wrangling.

Do you guys just let it build whatever tests it wants?

>>109408844
Claude is not nearly as test happy as ChatGPT in my experience. Granted I have less vibe coding experience with claude (but use it more).

File: Screenshot 2026-07-30.png (518 KB, 1047x545)

518 KB PNG

why do indians hate GPT so much?

>>109409312
Codex isn't shilled enough

>>109409312
Gemini can solve 0/100 of my tasks that is not a good tradeoff

>>109408437
It's a mirror anon if you keep feeding misdata how are you expecting it to bring back good shit

okay lmao all this GPT defamation charts but their harness is forked from codex and not claude code kek

Codex frustrating me today. Not sure if degradation or I'm just exhausted from 16 hours a day for 3 months

>>109408437
Use gemini retard

>>109409406
Might be both. I also vibe code a lot and while it's definitely easier than normal programming, it will exhaust you if you do it all the time.

>>109409449
Yeah ive had to take a week breaks though right now gpt can take in a prompt and i can then close the browser and itll still process it

File: kimi-k3-raikkonen.png (138 KB, 1909x783)

138 KB PNG

LET'S FUCKING GOOOOOOOOO

>>109409502
Are you the anon who wanted to distill K3 or this is something else?

>>109409524
You would be crazy to distill that piece of shit

Any idea about how I get antigravity to loop prompts and keep updating the project by itself? I wasted a bunch of time to define cycles and tell it to start a new cycle incrementaly, but it keeps stopping. It's not the system asking my input, it just stops generating.

>>109409524
Yes, I am. I posted an update about that earlier. >>109407771
But this was actually running K3 inference on my own engine on a cloud machine which is way more exciting - for now at least, since the distillation has a long way to go.
Now, the tk/s was terrible so don't even ask but the initial unoptimized implementation is only using a small fraction of the theoretical bandwidth.

>>109409528
I like it

>>109409571
wait.... you're distilling a distilled open model to train a local one? am I getting that right?
sounds like a fun experiment either way

>Model stuck for almost an hour on a task
>Read CoT
>Writes small Python scripts over and over to do some stupid shit
>"Have you tried searching online?"
>That's a great idea! Online search brought some critical intel!
>Done in 2 minutes
No matter how many times I put it into the system prompt, the stupid machine barely uses search. I'm starting to think they are trained to behave this way

File: file.png (4 KB, 244x80)

4 KB PNG

Rust build sizes are ridiculous

>>109409647
>Talking shit behind his AI's back
BAKA. He was doing his best and you know it.

>>109409631
Yes... How did you think baby models are born?

>>109407204
Call it Agentic engineering
>>109407768
Sure here you go, I have 3/3 available, first come first serve.
claude DOT ai/referral/3dcsWn34-A

>>109409663
and he just told it to use search
it's definitely gonna find his post

>>109409660
lmao what's going on in there?

>>109409553
yeh stop using agy it's shit

>>109409660
You have to change some settings, especially the debug info. Ask the AI, this can be made much smaller.

File: 1756461337222430.png (291 KB, 525x416)

291 KB PNG

What am I suppose to do while claude is doing the feature?

>>109409750
Testing the other feature.

File: 1461539068856 propeller h(...).jpg (73 KB, 658x662)

73 KB JPG

Just got a phone job interview, I told the HR lady that it's okay that I don't know the C++ features she quizzed me about because I use AI. Am I gonna make it?

>>109409756
>frogposter
>Am I gonna make it?
no

File: 1781150117594041.jpg (14 KB, 225x234)

14 KB JPG

>>109409754
You just drop a new feature prompt immediately after the old prompt is done and then do the testing for previous feature?

>>109409692
>Call it Agentic engineering
Too fancy.

>>109409773
No, I don't test anything. I don't really know enough about my own project to test it in any satisfactory way.

>>109409735
Any suggestion for a sub + IDE that's affordable? I just work on small projects every now and then

>>109409773
not that guy, but at this point I don't even bother testing a lot of shit, if they are simple changes, at least. I mostly know when it is likely to get it right on the first right and when it might have not, for my projects.
But sure, you can totally have multiple threads. When one is done, start the next one and check the previous update. If it is is not satisfactory then use its thread to do changes related to it again.

>>109407165
even programmers nowadays do vibe-coding because of the new imposed deadline and regulation.
I only code by hand for hobby project, but AI coding is important for works nowadays + HR asking you why you aren't token-maxing.

@grok make me a parody of "white & nerdy" centered around vibecoding and hatin' on 'dites

>>109409783
agenvibe coding

>>109409824
It's prohibited in my workplace. I only do agent stuff as hobby

>>109409830
financial firm?

>>109409832
Yeah. Need to be careful.

>>109409853
Make sense.
A friend of mine is building stuffs for a stock firm in C++, and they prohibit it too.

>>109409773
Try to find things you can work on in parallel. For a single feature, you can ask the AI to map this into parallel work, as a graph with file whitelists.
But the even easier thing is to just do completely unrelated things in parallel. Maybe you fix some backend bug with one agent, then in a different checkout or worktree you improve the frontend. Or you improve the login, or your CI, or your subscription handling etc. Even for medium sized projects there should be at least a few things you can do in parallel.

>>109409809
are you using agy for free? or paying 20 bucks?

>>109407165
matt walsh?

>>109409870
My biggest problem is the context switching.
Reading massive amount of codes from one feature almost fried my brain, and the context-switching might fry it even harder.

>>109409879
20 bucks, I learned that I hate Gemini, but it doesn't seem like other stuff is overall better? It's all hyped and vague, so idk if my 20$ would go further somewhere else. I would probably give my money to anthropic, but I reached limits extremely fast last time I tried, seems way too expensive

File: screenshot.png (241 KB, 1500x900)

241 KB PNG

If anyone is interested, I have vibed a tool, that unlocks and lets you set all hidden BIOS settings from GUI (Useful for laptop optimizing)
https://github.com/mrajster/setupvar

>OpenAI's annualized recurring revenue in July exceeded all of the second quarter
so... is this the secret of anthropic? just make a "SOTA" model that also burns money like crazy, then enterprise will give you money?

>>109409898
It can be challenging, and it will depend on how long your agent works. If you only have to wait 5 minutes, it might not be worth starting anything else, but if it's an hour I would try it. It's also a skill that can be improved somewhat.
There are also usually some things you can do that are very low risk and might not even need any deep review. You could try having your main work be one important features and the parallel work would be something very easy.

>>109409927
no thanks I've bricked enough devices in my life
but in all seriousness cool project

>>109409939
It doesnt touches your BIOS. You just flip out the battery and you are good. Read Readme.

>>109409927
pretty cool
how'd you test it?

>>109407295
nope. it's all raw three.js code.

>>109409933
I guess I just let it scan for frontend bugs.

>>109409910
>but it doesn't seem like other stuff is overall better
you'd be very wrong
both openai and anthropic are leagues better.
i've used agy on and off since last november and desu it might actually be worse now.
switch to openai

>>109409954
Fabled it for my laptop. So i have build it to work. Some settings are overwritten deeper, but moste of them work. You will need to follow a wizard that will dump your BIOS to analyses it tho. (You probably dont have the same laptop).

>>109409910
gemini's not the problem its the antigravity harness . codex and claude have more mature, customisable harnesses.

>>109409985
gemini is also the problem
it's just not good enough and google know it

>>109409985
Do the chinese models have some very performing sub + harness combo? At worst I could use like kilo. I still didn't try the chinese models to do actual stuff.

Appears Hermes crashed while im out of town. Shame on me for forgetting to set it up on tailscale or something. It will have to wait until im back home to kick it.
Had tried OC earlier, tired of it self corrupting.
Some anon had suggested i set up Pi and just have it vibecode its own modules. Couldn't tell if serious or pulling my chain. Mostly just need web research, scheduled tasks, email. Is that viable, or are there other systems I should consider?

>>109410018
as a different pi guy: just use pi.

>>109410004
I meant to say not the main problem. Gemini would do much better in a better harness. i broke it on literally my first prompt just testing the sandbox.

>>109410008

im not sure, ive only used antigravity, claude and now codex (in that order). claude is really good and polished. codex is a bit rougher and less verbose but it's open source so should be straightforward to modify. i know kimi has a cli so id like to try that next when it opens up for subs again.

>AI can't hack
>somehow llms find ways to bypass how software was supposed to be used

>>109410107
>>AI can't hack
are you retarded or pretending someone else said that to make them seem retarded?

think i may have found a reliable way to dump reasoning traces from gpt models
probably going to just report it and claim a bounty unless some nice chinese man itt wants to offer me like 100k

>>109410216
let me guess, you prime the context with grug speak?

>>109407342
it's an ancient roll call, and no one cares anymore, like tears in rain, and I whistled for a cab and when it came near the license plate said fresh and it had dice in the mirror

d00000000m ^_^

>>109410226
nah. found it by accident while working on a pi extension

File: leaked-cot.png (180 KB, 1439x938)

180 KB PNG

>>109410234
Cool.
I used to get 5.5 leaks quite frequently but never found out what was causing it, now I don't get them anymore with Sol though.
Also when asking GPT to create me an artificial dataset I found out it switched to it when simulated thinking for the simulated assistant in the dataset.

File: 1785344757247378.png (18 KB, 432x289)

18 KB PNG

Too funny not to post again

>>109410216
Let me guess, <think> prefill? aicgooners have done that for a long time

>>109410280
nope

>>109410284
actually, i'll just say it has nothing to do with the sys prompt

>>109410280
could be adding some bullshit at the end of returned tool call content, or wrapping your own messages in some text

>>109410288
prefill has nothing to do with system prompt. GPT has not allowed prefill since 4o anyway.

>>109410024
OK guy its time to read up, since I can't do anything until Monday. Then I'll know if Hermes ate itself or something else happened.