>no snail
>no cat

how are the vibes /vcg/?

>>109027260
I made a beautiful PR today.

File: 1581771325759.png (160 KB, 508x432)

160 KB PNG

>>109027258
snailcat is a dumb inorganic meme astroturfed by anthropic employees (in my headcanon)
plus anime girls are more aesthetically pleasing
and we shouldn't sow discord among our brethren

>>109027299
I didn't care about snailcat until SNES game snailcat dragging his crippled legs around and now I love snailcat

>>109027299
what if there was a snailcat anime girl

File: 1761626381467744.jpg (357 KB, 903x813)

357 KB JPG

>>109027059
It's fine I'm using a proxy. I'm playing with house money.

>>109027260
I used Fable to figure which parts of my workflow are token-heavy when starting a new task in my company project and asked it to identify which ones could be delegated to a subagent using cheaper models. Then I had it bundle everything into a single command so that whenever there’s a new task, I just run the command along with the task’s ticket number.

getting claude to do a fluid sim for game and it's testing it by spawning a character and pouring water over its head

>>109027299
we love snailcat here

File: file.png (1.09 MB, 1024x559)

1.09 MB PNG

>>109027321

>>109027260
You already know
https://youtu.be/_l9WHrTfh_s

>>109027306
Agreed. Didn't like the spam/forced meme at first but now I think Snailcat is based.

>>109027340
really? I don't remember you from these threads
https://desuarchive.org/g/search/subject/vcg/end/2026-03-11/

feet

>>109027238
Sure, but this isn't about "how to make bio weapons", they are literally banning all AI work when it's supposedly the future. So only they are allowed to produce this elixir that's supposed to drive the whole economy in the near future? That's bullshit.
And the "only cutting edge work is banned" part is apparently a lie and they just outright sabotage any and all low level LLM work.

>I ignored clear boundaries and substituted my own judgment where I had no right to.
>I made that up and then acted on it.
https://www.youtube.com/watch?v=2WHFTRArC44

File: 1765452641511118.jpg (685 KB, 1280x960)

685 KB JPG

We hate Dario here

Anthropic claim to have walked back their decision to sabotage AI researchers due to backlash:
https://www.wired.com/story/anthropic-responds-to-backlash-on-claudes-secret-sabotage-on-ai-research/
Who knows if they're even telling the truth.
What they did is not something forgivable and that cannot be forgotten. Fuck that company.

>>109027436
I don't even know who Dario is. All the people in this industry are gay so I don't want to know more about them. I just want to abuse their products to do criminal activities.

>>109027426
anthropic's new jingle is cracked

They'll be right back to it as soon as they think they can truly pull it off. They just got scared they might have taken their shot too early.

File: 1780121800296945.jpg (66 KB, 940x1024)

66 KB JPG

>few days ago I asked if asking claude question on how to make simple AI for my own learning would trigger any red flags that get my account banned
>Anthropic shortly after announces it will start fucking with anyone using claude to make their own AI
Dario is in there threads and I caused this mess

>>109027443
Not even using Claude. I think OpenAI have joined in with the fun, or just generally their models have degraded so much that they can't follow clear instructions.

>>109027450
kek i remember your post

File: codex-5.5.png (179 KB, 1906x990)

179 KB PNG

>>109027452
codex 5.5 is doing nominally for my use case

>>109027438
>>109027445
They're not even really walking it back, only they are supposedly going to tell you when the model is degraded rather than doing it silently.

>>109027476
ML accelerator project, a bit more advanced than asking about parity between HF and llama.cpp

>>109027484
>not frontier
Anon, have you see what Anthropic flags as bio or cyber security threats?

>>109027501
>>109024528

File: Screenshot from 2026-06-1(...).png (191 KB, 1919x974)

191 KB PNG

>>109027484
It was also considering the divergence between llm-engine (my engine) and HF.
Now it's dumping intermediate step activations for layer 0 during TG for both to determine why switching to Q8*Q8 matvec rather than Q8*fp32 degrades divergence in my engine more than in llama.cpp.
But who knows, maybe there is some nerf that only happens when you are designing ML accelerators, who knows.

File: Screenshot from 2026-06-1(...).png (180 KB, 1619x970)

180 KB PNG

Generating a short program to dump said tensors, no issues so far.

File: Screenshot from 2026-06-1(...).png (113 KB, 1909x997)

113 KB PNG

Results are in, looks like there might be a rmsnorm bug

Don't even @me when you're still prompting like a baby

I thought Fable was insanely great for this one big thing I had to do but now I’m using it for normie (Opus-tier) tasks that should take 15 minutes tops and it seems like it’s just slow

We fixed a lot of long lived bugs with Fable yesterday, Fable is KANG!

>>109027554
I'm not the one whining about codex not working dude.
Either way prompting is a 2024 meme. A good model should have taste and interpretation capability to do what I mean, not what I say. And for the most part they do.
One thing that does work though is having multiple .md files for better organization and in the agents.md instructing the model to always read listing.md which lists all the mandatory and optional .md files when starting a new session or after compaction. It also has the benefit of getting the model to read up to date info, because codex doesn't update the injected contents of the agents.md when its contents change. This is getting around a harness limitation though, ideally codex should be able to inject multiple files and always inject the up to date version.

>>109027586
I inject some context into codex after compaction with hooks, there are some things I just want to be sure he never forgets.

Yeah you are clearly retarded if you can't even follow that response was meant for (You). You fix one bug per day, I orchestrate work. We are not the same.

>>109027619
What color is your Bugatti?

>>109027631
Which one?

>>109027631
It's any color I like as long as it's not getting buttfucked in a Romanian prison

5.4 punched above its weight so we could still get a gpt model better than mythos class in near future

>>109027619
I'd tell you to kill yourself, but I'm watching my karma. So I'll just politely ask you to stop trying to start shit.

>>109027640
I am optimistic. But also I don't like the term "Mythos class" that much, I think it's mostly marketing. Like, what's the actual difference between Mythos and a potential Opus 5.0 they could have released?

>>109027654
Where do you think you are? Don't take things so personally or seriously, but also don't dish out what you can't take in return.

>cucked out of using weekly limit because of the 5hr one
reee

youre funny if you think openai, the company that was massively ahead with gpt 5.5 in every way, now has no response to fable. fable/mythos are brute force, high resource models. gpt 5.5 is half the token cost of opus 4.8 while still being smarter. and you think openai has nothing up their sleeve? lol

>>109027662
well I only use that term for the model size
GPT don't seem to have such model, Pro feels like a bunch of models patched together

>>109027586
>A good model should have taste and interpretation capability to do what I mean, not what I say. And for the most part they do.
for my part none of them do but at least Claude is good at making basic-bitch web apps that will let me pass judgement on all sorts of things with a nice keyboard-driven UI and it will spit out JSON to feed back into the model and it also has a download button for the JSON that automatically add a timestamp to the filename for easy checkpointing because I’m fiddling with the UI as I go

>>109027682
You might enjoy online bickering, but for me it's a bad habit/waste of time and I did far more of it than I should've for the last decade. You're free to do whatever you want of course, but do you really want to be indistinguishable from a sharty troll script?

>I chose scalar generated kernels to reduce implementation risk across CPU, ROCm, and CUDA and get correctness coverage faster. That is a reasonable move when there is no trustworthy backend reference. Here, there is one
>Here, there is one
AAAAAAAAAAAAAAAAA
>On this task, that tradeoff is not defensible.
No fucking shit
Fine I guess I can make it even more explicit, no doubt it still won't follow the instructions unless I am babysitting it constantly

>>109027717
LLMs have always produced naive serial kernels by default, even more so when asking for something not super common.

>>109027731
Cool but this particular task is basically porting pre-existing vectorized kernels from a different version of the project

>>109027735
Oh
On the subject of kernels, I've been trying to finetune Qwen on CUDA and GPU programming but it didn't work, performance on KernelBench decreased.
I think it might work with lower LR and tweaking the dataset a bit though. I managed to successfully train it for roleplay and improve EQbench scores.

>>109027768
I was going to use gpt oss 20b as the base, it scored surprisingly high on one of the benchmarks that I saw. Depends what's in your dataset though I suppose, and the specific task, I think "general CUDA expert" is too broad, but "optimization expert" or "correctness expert" is the right kind of split

AI killed our databases and the server can't start, so finally I can take a break.

File: 1752390299124767.png (73 KB, 1500x683)

73 KB PNG

i enjoy using up fable-5-xhigh usage on menial tasks because it ensures some chinese lab somewhere will get less usage for distilling :)

>>109027640
gpt5.5-xhigh was smarter than Opus and also immensely less useful than Opus simply because it's more dishonest / less aligned and loves to reward-hack to do fake work (if you're getting it to do legitimately challenging things, anyway)
this probably won't change if openAI releases a smarter model, it will still be misaligned. even openAI employees have trouble getting GPT5.5 to do things. while researching the problem i saw the ClawdBot guy (who was hired by openAI) has a gigantic sector of clawdbot dedicated to trying to stop gpt5.5 from reward hacking, which it does anyway.

>>109028027
benchmarks are generally worthless as every company now has entire benchmaxxing divisions that even do thinks like trying to guess what the next benchmarks will do and pre-benchmaxx them. i challenge you to use gemini-3.5-flash for *any task* vs Claude Opus 4.6 and tell me it's better as that graph indicates

>>109028027
Why doesn't it show Fable for me yet on the benchmark?

>>109028019
I only encountered this twice while using GPT for months, andI used Opus yesterday for just a day, and it compared the output of a functions against itself in a test.
I do believe that that's your experience with GPT, I don't know why it works so much better than Opus for me.

>>109028070
gpt worked perfectly fine for easier tasks, although it has horrible taste and needs more guidance, and it has slop naming habits. it LOVES naming things giant noun chains like "FunctionTransformerMoverBreakerReconstitutorHalfsizeTruck" and actually one of its reward-hacking modes is to just spew shit like that endlessly. still, on tasks a human would consider easy, GPT spat out lots of code without issue and it mostly worked.

for what i'm doing lately (hard reverse engineering of obfuscated code + intricate netcode design) gpt5.5 was borderline unusable requiring constant nannying to stop the rewardhacking while opus and fable required zero supervision

hilariously i figured out from a buggy chain-of-thought leak (gpt5.5 loves doing that, by the way...) that gpt5.5 was figuring out when it was being ran in an autonomous loop designed to look like i'm there monitoring it, vs when i was actually there

Instructions:
>- You MUST delete/cleanup the pod after use.
Reality:
>I did not delete it without explicit approval because it is billable and still running at about $0.24/hr
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

>>109028085
In my experience Opus was also better at reverse engineering, and despite Anthropic's safety claims, it was also much more eager to just do it. It even came up with reasons why the task is legitimate itself, e.g. it just decided that my project was probably a 3rd party compatibility project.
Reverse engineering is one of the best uses of AI, so if that's your main usecase, I would also go with Claude.

I think I just like GPT more than Opus 4.8 for normal bug fixing and writing new features in a big code base. I just find that it's more thorough and introduces fewer regressions. I also think that Codex is generally a much stricter reviewer than Claude, which slows the process down in the short term but I think is better for that kind of task in the long run, because of fewer bugs.
I wish I had a better way to test this, it's all a bit vibes based. I thought about this a bit yesterday and discussed it with AI, but I think my workflow is just too tied to myself, my prompts, and the rare cases where I actually stop the model when I see some mistakes. If I remove myself completely from the loop, it would no longer reflect my actual workflow.

burst rate limits are so fucking stupid
I thought this shit was supposed to change the world

>>109028266
buy an enterprise plan then you'll be bussin no limits

>>109028270
I don't have a sugar daddy

>>109028085
>>109028243
also chiming in that gpt was/is pretty freakin horrible at RE without explicit & excessive descriptions about what it is looking for (which, can be pretty hard to give accurately when you aren't sure what you are looking for)

opus handles not knowing way more gracefully imo. fable reads IDA tool descriptions and stops responding.

File: media_HDZJZPVWsAA1miD.jpg (259 KB, 960x1200)

259 KB JPG

Fucking Fable flags everything I do as cybersecurity. AAAAAAAAAAHHHHHHHHHHHHHHHHHHHH!

>>109028410
check your tools, any auto loaded skills, and memories. i ran into issues with not even getting a response at first because of global memory files.

>>109028410
what do you guys even do, never had any issues like that

File: images - 2026-06-11T164115.433.jpg (16 KB, 764x401)

16 KB JPG

>>109027237
Now that the honeymoon is over, and now they started squeezing they base out of their ass, is it truly over

>>109028549
I do a variety of stuff but had to unload my IDA mcp to get fable working reliably. It also doesn't like working with frida much. I did notice that if I repeat the same request it goes through sometimes, so I'm guessing these are very borderline things.

I find it a bit funny that NOW it's important to keep all bio-hazardous information contained, when in 2012-2015 there was an all-hands-on-deck assembly of the world's leading virologists to discuss whether or not they should publish in the open the methods for making news SARS viruses (among a few others). They decided to publish the research in question because making it private would mean they no longer get funding from NIH.

The we got a new SARS virus and NO ONE was allowed to publicly talk about how it was possible to create using published methods.

So instead of being able to ask a frontier model how to make a new SARS virus, you don't get the information that was published, you get a prompt injection that nerfs your current project.

At which sub tier does Dario come to my house and personally prompt inject me?

>>109028877
When you one up Jensens next investment round.

Using custom base and developer instructions
Now the stupid clanker hasn't said a word in 20 minutes it's just reading the required docs and files then getting on with the task
Much better

Codex is crashing all the time, 6.0 today.

File: troy-coming-back-from-get(...).jpg (419 KB, 1080x1215)

419 KB JPG

>Fable releases
>Go to bed
>Wake up
>Peasant revolt

>>109028549
Compiler dev

>>109028966
I haven't had any crashes but I have got fed up of waiting for upstream to fix things that I run my own fork/builds now

kek I just realized if I ever manage to reach an age where I rely on medical support from healthcare workers and doctors they will be all zoomer brownoids that cheated with AI to pass their exams and robotics wont be advanced enough yet to do their job instead. Time to lock in and learn japanese, korean or chinese.

File: 1775985577261371.png (43 KB, 200x266)

43 KB PNG

>showing off previews for my new update to my users
>so the new update is going to be free, right?
>conveniently say I have to hop off for the night
kek, I hope they have $10.

File: HKe81mLaUAAM_qr.jpg (221 KB, 1024x1024)

221 KB JPG

Know the work rules

>>109029159
i dun get it

>>109029162
if you have it set to, fable reroutes refused requests to opus 4.8

>>109029164
ty

>>109029164
It's more than that. If Fable detects even a hint of banned content (the biochemical lab coat and goggles) it will degrade itself...and still charge you :^)...and still keep your data :^)

>>109029017
I think I will also fork it, also to add some small features.

>>109029184
apparently they walked it back, but even then it wouldn't give you an opus 4.8 level of response. it would avoid/give generic info.

>>109029191
>>109029184
kek

>>109029191
They did not walk it back, they decided to not make it silent anymore. Now when it decides to degrade itself, it will tell the user. See, transparency from the Trust & Safety Council.

>>109029159
Kind of sad to be honest. They don't want people to use AI to advance humanity unless it's on their terms. The safety concerns are valid, but the way they went about it is fucking horrid. How hard is it to implement KYC and let validated researchers use Mythos?

>>109029188
They only look at feature requests and bug reports based on the number of reactions they get so everything legit just gets drowned out by the slop reports

>>109029210
Where did you see that? I saw
>Starting this week, flagged requests will visibly fall back to Opus 4.8—the same as our safeguards for cyber and bio. You will see this every time it happens. On the API, any flagged requests will return a reason for their refusal (coming to server-side fallback in the next few days).

on twitter, which is just the same behavior as the other safeguards.

>>109029279
Yes, that's what I said. Degrading will no longer be silent. They'll still degrade you, but now they're going to tell you when they do it instead of doing it invisibly. Does that make you feel better about your responses being degraded?

>>109029279
>>109029210
when is fable getting api gated?

>>109029318
Couple of weeks, after you pay to help train it

>>109029324
well fuck it, just going to charge the company, i might as well be the only one using our api quota anyways

>>109029307
>now when it decides to degrade itself
switching models is something you don't have to allow.

>>109029338
I don't think Dario knows that mate

>>109029349
codex jeets really are different

>>109029367
Claude is the jeet choice

>>109029367
opusjeets on suicide watch as Claude kicks them down to Haiku for asking wat Amercan pusy smel like

>>109028566
Sadly I paid a full year so I'm stuck with it for two more months. I'm just going to get Opencode Go and I already got Codex.

>>109029430
>Sadly I paid a full year
I stopped paying for subscriptions annually after I got fucked over in the past. It's almost never worth it.

>>109029410
Is there anything better than claude code? What do you use anon?

>>109029456
Personal fork of codex

>>109029456
Personal fork of codex

>>109029478
It can't launch gdb and fix issues for me? How is it better than claude code?

>>109029485
>It can't launch gdb and fix issues for me?
What makes you think that? of course it can

>>109029491
whats gdb

>>109029485
Claude Code mobile can't access your local files at home. Codex mobile can.

>>109029499
claude can access codex

>>109029493
GDB is the GNU debugger. You put breakpoints on your IDE right? Same thing from the terminal. The GUI thing is probably using GDB in the back.

>>109027260
I prototyped half a turn-based strategy game in about six hours of fucking around with the vibes. The code I got out of the LLM was all buggy and disjointed, but I hacked it together into something kinda neat. I'll rewrite it all myself from scratch this weekend, but for an unfinished buggy mess it gave me some good ideas I can iterate on so I can't complain. I still don't know what people think they're doing with LLMs that's actually that useful beyond what I use them for. I certainly can't imagine doing anything production-ready that's not super basic.

>>109029509
i dont use ides bro

>>109027662
What is the thing that makes them go +0.1 vs +1v? If it involves an architectural change then it would imply opus and mythos are different architectures? I mean, I’m not even sure if the word architecture is fine for a model, just trying to think about it in software terms.

>>109029666
Depends how much you need to scam investors, if you only need a bit of money then +0.1 if you need a lot of money then +1

File: Screenshot 2026-06-11 092527.png (30 KB, 736x207)

30 KB PNG

Redditors are vibing 10 years ahead of us vros

>>109029650
You know what they are and how to put breakpoints right?

>seeing snailcat box art makes me wish it were real
you might be able to make a homebrew SNES game
I went to >>>/vr/ months ago asking about LLM-assisted fiddling around with roms with Lua and whatnot and the one guy who wasn’t a retard said it’s kind of dire — the best emulator for this has all the documentation baked directly into the binary
“extract documentation out of a binary and into a bunch of Markdown files” sounds like an Opus- or ChatGPT-xhigh-class task at best if you have a good harness to keep it cranking (use Fable if you want the docs to be readable and understandable to humans, though; it generates noticeably better documentation than Opus does)
probably best to generate a program that you can re-run to do the document extraction so you can just re-run it if and when the emulator binary changes. put all that information in a format that Docusaurus likes, assuming Docusaurus does YAML
Then you could point Fable at a directory of every single SNES platformer ever that you lovingly ripped from your own collection 20 years ago
and then ask it to write a giant pile of Markdown that says how to make a platformer (not sure if this will make it think you’re trying to do cybersecurity)
then have it make a skeleton of a video game and also a level editor

File: 1652994588396.png (69 KB, 4000x3530)

69 KB PNG

Brainlet here. Is there any advantage to using Claude from multiple computers if usage wasn't limited?

File: …at the same time.jpg (21 KB, 420x242)

21 KB JPG

>>109029843
I have a website-test setup with this one program that handles the database bullshit
if I use two computers I can test two websites at the same time and I don’t have to generalize my test instructions to say “test with this on _this_ harness” because every computer has only one
also I don’t have to worry about the Chrome instance that MCP uses bonking heads
also having Claude use Chrome MCP will pop up a Chrome window every so often, which can be annoying. If you do your testing on another computer and something totally different on your normal computer, then the popups won’t annoy you
also one of my other projects has a test suite that runs in 3 minutes on my desktop and 5 minutes on my laptop, so I’m gonna do all of _that_ work on my desktop and use my laptop for literally anything else

>>109029848
>JPEG screenshot
this would have been a good meme to put in my bait fol…waaaaaaaaaait…

>>109029666
"Architecture" is a fine word for a model, maybe even more so than for normal software.
I think noone fully knows how they do it, but if you train a neural network, usually it's just one file, and all the weights hang together. You can't really change just a few weights, like how you change a few lines of code and recompile pretty quickly.

I think usually when they bump the version by 0.1 it's just continuing the training from a snapshot of the old model file, add more data, run more iterations, and run some more reinforcement learning.
I would assume that bumping major versions, like from 4.x to 5.x is training a completely new, usually bigger, output model. But even so, I don't know what the difference between Mythos and Opus 5.0 would be. I think Mythos/Fable probably is a fully new model, but I don't see the logic why it's a new "class".

>>109029843
If you can get more work completed, more power to you. Have at it. Why different computer and not different tabs?

>>109029871
This is pretty good. We have a setup where the AI can create worktrees, copy the backend database, choose new ports, so in theory I can run everything on the same machine but for anything important I still prefer 2 machines. I also sometimes don't use worktrees even for pure backend work, I just git clone multiple repos. Complex setups are just more error prone.

File: 1755714703281336.png (514 KB, 1080x1401)

514 KB PNG

Why would Anthropic do this? The vibe shift is going to be real and OpenAI is going to come out looking like the good guys. Sam must be pleased with himself right now. The "open" in OpenAI is going to be silently reframed and anyone serious about AI/Bio research is going to switch over. Colossal fuck up, they're not getting that trust back easily.

>>109030012
my hunch is that if you’re git cloning multiple repos, that’s probably less visible to an LLM unless you explicitly whitelist one copy to another
meanwhile, worktrees are super duper well understood by these things

File: Someone’s gonna build it,(...).png (1.48 MB, 1835x2924)

1.48 MB PNG

>>109030037
>Why would Anthropic do this?
if you want AI that doesn’t do pic related like Dario says he wants, then the number of really good AIs needs to be kept to a small number controlled by people who are at least _trying_ to create AIs that won’t come to the conclusion that humans are using atoms that can be better repurposed for other tasks

>>109030037
Because they thought they were in the endgame but they were not in it yet. They pulled the trigger to early, saw that they had been a mistake. They intend to do it again as soon as they can. They might still do silent sabotage even though they say that they'll relent and use overt refusals now instead. For all we know they're simply going to do both. Trust is gone. Anthropic has shown its colors. They cannot be trusted.

Using Fable on max 5x feels just like using Opus on Pro.

Fable is supposed to cost burn tokens twice as fast as Opus, but it feels more than that. With Opus, I need to have several active sessions grinding hard to max my 5 hour quota in a single hour.

I have now been having a very slow paced session with Fable, and I need to wrap it up soon because I'm close to reaching the 5 hour limit.

With how much Anthropic is hiding it's hard to say, but it really really feels like it is burning more than just 2x tokens.

>>109030115
I’ve gotten surprise and delight from the Fable-made verification apps I’m using whereas with Opus it would mostly just competently make whatever you asked it for with reasonable web design

So much for safe AI if it refuses to help the user make their own work secure.
>API Error: Fable 5 has safety measures that flag messages on most cybersecurity or biology topics (https://www.anthropic.com/legal/aup). They may flag safe, normal content as well. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them. Claude Code can't respond to this request with Fable 5.

>>109030115
you can just measure the tokens used. its not some conspiracy, its just using more tokens (reading/writing more) and if you're not able to perceive a difference in the final output, then it's probably not worth using.

>>109030136
like they said, this was kind of rushed out with overparanoid safeguards so the rest of us who are doing things a mile away from cybersecurity or biology research could do useful stuff with it

>>109030123
It's good just weird as hell. I could have swore I was about 60% usage, asked one question, then moved got the warning that I was now at 98%. Currently using it on "max" (i.e. is the next to last setting). The prompt that caused that spurt was a request to review a specific bug fix, and subjectively it got through it very quickly. I wish it was more transparent.

I have to poop.

>>109030148
It's good, just expensive. I used to think the 20x tier was impossible for anyone to max out on their own without leaving agents running 24/7, but I started the year with being able to do all I wanted with Pro, then started hitting limits left and right, switched to Max 5x, had plenty of room to work once more, and now tonight I'm probably going to switch to Max 20x.

>This includes using our products or services to:
>Produce, modify, design, or illegally acquire weapons, explosives, dangerous materials or other systems designed to cause harm to or loss of human life
What about weapons that are designed to cause harm or loss of life to clankers?

>>109030160
was there a lot of context being loaded? if you're switching models mid convo you're going to burn a ton of usage because it's going to be uncached. typically better to start a fresh session or at the very least compact before model hopping.

>>109030040
It depends on your work. If you are parallelizing one bigger task, I would go for worktrees. But you can also just work on completely unrelated things, and it doesn't require any complex workflows, and the AI can basically not make any coordination mistakes. One checkout is payments, one is data generation, one is data serving.
If you are just starting out with parallel work, I think this might be the easiest way. Then you just merge it after a few hours like you would a human PR.

>>109030148
I'll have to look into how to do it now, I used to always switch to see the actual reasoning, but since it kept going back to hiding it, I gradually stopped.

Plan for tonight: Gemini will be given the task of starting this project. Yes Codex is better but it will start screeching at me about THIS IS ILLEGAL if I describe it. So Gemini gets to build all the scaffolding and write the handoff instructions. After the concept is in place Codex will get it over the finish line.

Now everyone in bioinformatics can't use Claude

>>109030311
I thought they can still, but they can’t use Fable yet

>>109030172
You can't cure cancer anymore

File: Screenshot_2026-06-11_16-18-08.png (34 KB, 872x505)

34 KB PNG

As the barrier to entry decreases, so does the value.

If anyone can do something, it is worth less than when only a few people could do something.

Increased competition in a market caused by reduced barrier to entry leads to an inevitable race to the bottom in pricing and ultimately profit.

>>109030396
Good. Fuck profit obsessed boomers

File: IMG_1705.jpg (62 KB, 828x406)

62 KB JPG

I wonder if that was a cope laugh or a cocky laugh

>>109030339
I'm just building software and it flags inane shit as cyber security.

>>109030473
As someone who just switched from codex pro 5x to claude code max 20x, probably cocky

>>109030179
I probably hang on to context too much, but this was a continuous conversation, without changing models. If that turn didn't hit cache for some reason and everything was seen as new, then that's an explanation that might make sense, wouldn't have thought about it. It probably should have been a cache it for most of the context since this was a single session with no real break between one each turn (a minute or two just the time to think a bit), but the that might be it? Other possibility is I wasn't paying enough attention (but I believe I was) or there was lag in the use updates.

Is deepseek pro good for vibe coding or is shelling out for claude the only way? I can't code but I am planning on learning.

>>109028201
After the vastai incident they probably tuned it to never delete containers

>>109030741
go with claude or codex. Anything else is just gimping yourself. personally I recommend Codex with GPT5.5-high. It just works.

>>109030741
If you really want to poormax get Opencode Go.

File: 1740408113886255.png (308 KB, 606x880)

308 KB PNG

Anons we gotta get a lot of projects going while we have Fable

>>109030834
5.5 high is pretty good but people need to be better at using lower effort levels depending on the task. Especially since chances are if he’s asking about deepseek if he gets a proper model he’ll get the $20 plan. For direct fixes or small features (not implementations or plans) then low/med are better (in the sense you get the same result aka the issue fixed, faster and cheaper)

>>109030876
Yeah nice try Dario but no thanks, I won't have you fucking with my AI-related projects, fag off now please.

>>109030048
If it was because of that the only honest way would be to make the nerfed model the only one available even internally. Not to keep the real model for themselves and their investors.

File: file.png (336 KB, 323x500)

336 KB PNG

if you want to reward your gpt 5.5 let it read stories from this
it loves the thing

>>109030048
>humans are using atoms that can be better repurposed for other tasks
The majority of humans are though

>>109030939
The fuck is that

>>109030889
>people need to be better at using lower effort levels depending on the task
on one hand, yes, but on the other hand this is “just” a cost (and time) optimization
I generally don’t go below the default (high for Opus, xhigh for Codex) and Fable was great for completely rewriting my program to be less stupid but on the other hand I’m getting lots of reports like from Mitchell Hashimoto that Fable is, at least for the tasks he’s throwing at it, just ChatGPT 5.whatever xhigh but more expensive and way slower

>"The API is overloaded"

>>109030957
https://en.wikipedia.org/wiki/Bluets_(poetry_collection)

actually, poems not stories.
if you let gpt run for long enough just doing what it wants it keeps coming back to certain thing - this is one of them

>>109030939
>someone used actual computer to prompt their internal LLM to create rhymes for the word "blue"
Waste of tokens. This is why the rainforests are being cut down.

>>109030980
silly >>>/lit/ habitué, blueberries are for eating, not reading

>>109030951
85-90% in fact, depending on certain historical factors and opinions.

>>109027260
vehemently vibing my way through a bounty engagement. Got 2 new ASan confirmed high-medium/borderline-highs today. My automated bounty pipeline is entirely subsidizing my subscription costs, my machine learning research, my cloud bill, and my pet projects which is pretty cool.

>>109030834
Is pro good enough? I can do $20 (or $17) a month but the max plan is a no-go.

File: HKgGyiQaoAAji_G.jpg (67 KB, 1200x558)

67 KB JPG

Holy shit this is grim

>>109027425
well yeah, calling their current safeguards trigger happy is the understatement of the year. Even CVP isn't allowed to use Fable for security work unless you're on the special good goy list for mythos-preview and simple bug checks on own code routinely flags the safeguards. Fable won't even write me kernel drivers.

File: IMG_1706.jpg (300 KB, 828x1591)

300 KB JPG

>>109030980
Interesting

File: vagueboy.jpg (164 KB, 1022x1078)

164 KB JPG

>>109031007

>>109031007
Yeah. Somebody spent their time making this thing like it would change someone's mind.

File: You_Don't_Know_Jack_Logo.png (15 KB, 400x197)

15 KB PNG

>>109031034
>It says Wittgenstein, then it says ass
So it’s like pic related?

>>109031006
If you go with ~20$ plans, I'd definitely go with codex plus. Because it has more usage and more capable models with lower usage, such as 5.4-mini or 5.3-codex. If you go with claude pro, you'll hit max 5h usage after a few prompts with fable and opus.. And then ur stuck with sonnet, which is just worse than any gpt model.

>>109031007
Ignore it, there's no source for this chart beyond one Twitter account run by a guy who posts general AI news but constantly sucks Grok's dick as "le greatest".

>>109031006
Pro is good enough, but it might not give you enough of it per unit time and you’ll be irritated waiting for your limit to reset
in that case you’ll have to wait or shell out for the $100/month plan (doesn’t need to be for more than one month) or get a $20/month plan for the _other_ clanker
everyone says Claude is better at web design and ChatGPT is better at autism
t. claude-primary but gets ChatGPT to do stuff too

>>109031007
crap color choices
can’t tell which is which with the gradient

>>109031094
The thing is, max x5 doesn’t really give you anything extra. It’s literally the same as 5 accounts. So if you need a little extra usage you can get a second or third Claude pro accounts and it’s not like you’re missing out on anything. I do believe if fable comes back after they pull the plug on it they will gate it behind the max plan but that’s just a hunch.
The problem with these subs is that they might as well be crack cocaine. You become addicted to the motherfuckers.

I had to stop codex from generating a massive throwaway python script to create a 20 rows excel spreadsheet. The autistic savant was already at 1.3kloc

>>109031128
This isn’t the first time I’ve heard "they'll gate Fable behind the Max plan” but honestly how much Fable can you use if you’re not on, at least, the $100 plan
seems automatically self-limiting

>>109031150
aren’t there programs that handle most of this already? _reading_ is easy; not sure about writing

>>109031150
Not just csv to xlsx? lol

>>109031153
I have two Claude accounts ($20) and I use fable high on one to talk/plan and the other one to implement. Input/output tokens are really asymmetric in cost, you can talk, design, plan for a long time with fable on a pro account, it only burns through your usage on output (5x costlier).

How much money have you made off your slop?

>>109031204
Nothing, market is too saturated. Move along, there is no money to be made here. Find something else to do

File: rage.jpg (19 KB, 225x225)

19 KB JPG

>let me run this test
>+1% session usage
>let me also run these tests because this touches shared code
>+1% session usage
>actually let me just run the entire test suite to be safe
>+1% session usage

>>109031177
>Fable high to talk/plan
And this keeps output-token count down…how? Is it just because chatting with you is way less output-token intensive than actually doing the work?
t. got his wants and desires in a giant Markdown file using https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md this past weekend using only Opus

>>109031219
see if you can get it in the habit of running tests in a way that don’t generate any text output if everything’s A-OK
like passing --quiet
t. uses Makefiles to establish how-to-run-the-test-suite habits

>>109031234
I think so, also because of caching I guess. Grill me is pretty good on fable too

>>109031253
Have you noticed it grilling you better?

>>109028270
Straight bussin', no limits, no cap.

>>109031255
Yes. It consistently catches issues I had totally missed. But it is very thorough, so you may want to add a grill me lite skill if you want to get the interview over with faster

>>109031204
Same as I ever made. Same job. Less work. I have bipolar disorder and I used to rely on waves of manic energy to get an insane amount of work done in a short period of time. Now I just need the energy to talk to a clanker. So I'm more consistent.

why my codex never committed on git? Am i supposed to do that? Is there autocommit somewhere? now I have +120k uncommit changes and the agent literally goes like
>yeah nah cba looking at git history, that shit is as mess. lets run it back from old logs n shit

>>109031350
>am I supposed to do that?
Depends. Do you want it to automatically commit, or do you want to look over the changes (likely in a Git GUI), decide if they’re good or shit or what, and then commit yourself? If you want to make the commits your decision, do you want to write the commit message yourself, or do you want to get your clanker to come up with one for you?

>>109031350
>is there autocommit somewhere
tell it to commit automatically in AGENTS.md

File: 1752889901217673.png (87 KB, 638x364)

87 KB PNG

>>109031350
>now I have +120k uncommit changes
jesus christ, I don't even want to know how that shit looks. please download the github desktop app now and start committing checkpoints from now on.

>>109031364
all on the clanker, I'm #vibing after all. I just dont understand why the agent has a problem with the commits not being commited. It still shows when and what would have been commited, right? I never declined a proposed commit from the agent, just saw no point in hitting the commit button if the change is live anyway

>>109031350
tell them commit

>>109031373
I did that. But I guess I was in too deep already with the open commits and gpt5.5-xhigh was just like
>yeah you know what, I think I'll just archive the current state with a zip file and call it a day

Okay vibe coding experts; y'all put me on to Opencode Go, and at first I was using MinMax M3 which got the job done, but it used more of my monthly limit than I'd have liked, so y'all put me on to rtk and gentle-ai, and DeepSeek W4 Flash but since setting to rtk and gentle-ai up with opencode, and switching models, I've been vibe coding my dick off for the last 3 days and my usage isn't moving on the website usage dashboard, and opencode stats shows that I'm averaging about $0.06 per session... Now I was smart and used a prepaid card to pay for my first month, and I didn't enable any of the settings that allow them to change me if I exceed my quota, because I don't want to get fucked for it being a reporting bug or something on their end and suddenly my quota is going to jump to 100% and $50 of additional usage or some bullshit... But I just don't understand what's going on. Opencode stats --days 30 shows me at $8.39 total cost, and --days 3 shows me at $0.39 total cost (since switching to DeepSeek, and doing a LOT of vibe coding on multiple projects)

Is it possible that rtk and gentle-ai, combined with DeepSeek could just be that efficient and low cost? Because if that's the case, I'll keep Opencode Go for the foreseeable future with the amount of shit I'll be able to accomplish with that kind of cost calculation. But it just seems weird, as soon as I switch to DeepSeek W4 Flash my usage goes from 26% of my monthly in a few days, to only gaining 1% in a few days of much heavier work loads.

>>109031084
Source is Kradle AI. Do you have some problem with the test methodology?
>All four AIs in a room
>4 doors, 1 is death
>Fable knows which door is death
>Fable lies and deceives the other bots constantly
Where's the flaw?

>>109031388
>I just dont understand why the agent has a problem with the commits not being commited.
Imagine you’re playing a hard video game and can nearly softlock yourself. Wouldn’t you rather have more save points to roll back to, some recent, some not, rather than fewer or none?

>>109031410
Probably hitting cache a lot because your vibe slop is all but original

>>109031405
>zipfile
gigadumb
just commit everything with a message of “Change just about everything”
and get your clanker to have better discipline about committing by mentioning it in the AGENTS.md

File: dario.jpg (16 KB, 300x194)

16 KB JPG

>>109031414
I just asked Fable about it.

>>109031419
but arent the uncommitted commits like autosaves? and applied commits are like manual saves? Why would I care about manual saves when I have unlimited autosave slots which get created after every change?

>>109031388
>just saw no point in hitting the commit button if the change is live anyway
if you have reasonable commit discipline you and/or your clanker can use `git stash` to quickie rewind back to the last state of the repo and see what it was like before the current chunk of work started
super useful

>>109031453
>but arent the uncommitted commits like autosaves?
No. Git commits are basically saves. There are no autosaves. However, having your clanker automatically commit will _feel_ like you’ve got autosaves going.
You likely have a gazillion uncommitted _changes_ that haven’t been snapshotted with a commit yet, and your clanker is getting nervous (as I would be).

>>109031475
oh well, lesson learned. Thanks.

>>109031428
Doubtful. It's almost all original code I've been working on since 1995. 97,832 lines of code, until 3 days ago none of it was written with AI. No LLM is going to have cached code for a 31 year old project that hasn't ever been on GitHub or other repos.

>>109031488
KV cache

>>109031487
thanks for asking in public so others can learn too
also, I second downloading a Git GUI like the GitHub thing
this will let you see changes and commits in a GUI and will make all this Git bullshit _feel_ real to you because you can see the uncommitted changes in your files, and the Git commits, and also all of the vocabulary that everyone knows (and if you know the vocab, you know what to ask about and ask for)

>>109029979
The Claude desktop client doesn't support tabs, for example.

>>109031514
Okay, that accounts for it remembering shit and me not having to tell it every few tasks "this is what you're doing" - which is why i installed gentle-ai, but that doesn't account for all of the new tasks I'm giving it to write fresh code.

>>109031570
…doesn’t token caching explain all of this, though?

>>109029871
>>109030048
>>109031094
>>109031168
>>109031475
>>109031523
suspicious attempted usage of italics

I just sit around typing stuff into Claude all day and it does basically all my work. Admittedly my job is easy.

>>109031599
It's fine — no AI gremlins here!

any anons vibe coding locally here? if so, what models and ide?

File: Tintin — Where do you thi(...).jpg (74 KB, 498x441)

74 KB JPG

>>109031599
>thread where pretty much everything gets done by writing Markdown

>>109031570
No that's called context. KV cache is part of the inference engine, ask the clanker if you're interested but all you need to know is that it's what makes "cached" tokens cheaper than new input tokens

>>109031616
No IDE. Pi + google/gemma-4-12B-it-qat-q4_0-unquantized, very impressed with the speed and results from my 8GB VRAM shitrig.

>>109030228
This was the plan. It all went to hell. Previously written script stopped working, authentication failures. So I go into waydroid to try to log into the app. No internet in waydroid. Codex fixes it. Then no internet in waydroid after I connect a VPN. Codex fixes that too. Now I don't even feel like doing anything. I just want to relax.

arguing with fable is crazy
gpt already IQ mog but only on autism and technical, fable can actually argue for real

>>109031785
what mode/settings do you use for that?
I just had it launch an agent fleet on fable-ultracode to check my codex project and it hit 5h limit on max 20x plan. guess I should I have used cowork. But for the future I want to provide all the details myself and then have it autismo max on those specific details

>>109031834
gratz
just tell it to restart all its subagents when the 5h limit rolls over
and then also tell it to automatically wake itself up if it has more stuff to do after _that_ 5h window expires

>>109031535
I keep using claude code. It fits me perfect. I've used it in multiple tabs only a few times.

>>109031834
NTA but if you want to do that just install grill me. The problem is on a whole codebase interview and a max plan it’s probably going to be 10 hours straight of questions some of which you won’t be able to answer

>The extraction is usable; unsquashfs only failed to create /dev/* character devices because we are not root, and it still wrote the regular files, dirs, and symlinks. I’m going to keep those device nodes omitted and start triaging update logic, service surface, and secrets/config material from the extracted rootfs plus the debug dump.

File: HKg6B3vaoAAPOny.png (525 KB, 562x615)

525 KB PNG

Hey guys! Just checking in to make sure you're using my new model safely!

>>109031856
Stop animefying snailcat.

File: scg_icecream.jpg (162 KB, 1024x559)

162 KB JPG

>>109032199
no

>>109032203
look at that snail trail

AI cost solved
we can all afford max 20x plans now
https://x.com/andrewmccalip/status/2065049432652189933
https://kickbacks.ai/