>>109701806
first for how the fuck do I make an AI model for myself

>>109701815
Just led claude walk you through the process.

>>109701806
That image is central to Michael Crichton's new book 'Bovine Park'.

>>109701815
Just say how it works and feed it through autograd with your data.

>nobody ever talks about composer anymore

It's still pretty nice.

>>109701831
>autograd
When I hear "grad" I always think of the rocket batteries.

So, the project I have been vibecoding with local models has gotten pretty damn far. I want to use some frontier models to do a code review or two. What models should I use? Should I get a subscription? Pay API prices? I'm not interested in having a permanent sub, I'm happy with the models I can run but I do want to see what the big boy models can do.

>>109701875
I switch back and forth between Ornith 1.5 9B and GLM 5.2 via Openrouter at work. I have Ornith on a mac mini and leave a couple agents in the background churning on low priority problems and only bring out GLM 5.2 when both it and I get stuck on something important.

I don't vibe code though. If you really want to leave it unattended use kimi k3 with subagents. If it's possible that *will* figure everything out but it could cost like $10 or more.

>ask qwen about massacres by the chinese government
>none
>ask qwen about massacres by the US government
>lists a couple
Awesome. So it's good at coding?

>>109701875
you can pay as you go (openrouter / nano-gpt), GLM 5.3-flash is super cheap and works pretty well for me. You can go up the pareto frontier if you want to pay more for more quality

>>109701875
API prices are enterprise prices (ass-rapingly expensive)
How many $20 or $100 subs do you want to get for a month each
The OP has the three decent models
You only get Fable if you spring for the $100 or $200 Anthropic sub
You will probably want $100/month subs
Fable for big architecture things
Sol for bugfixing autism
Grok is currently the least obviously useful one but it doesn’t have 5h limits like the other two

>>109702010
I thought Fable was banned by the US government?

>>109697901
More good news!

They absolutely need to fix claude's personality.

File: Brian_Heder_03.1bd6d8bd.jpg (481 KB, 3840x2160)

481 KB JPG

>>109702011
Listen clod. You either accept the usage limits and "enjoy" 10x cheaper tokens or pay the market API rate with no limits but your wallet.

>>109702011
You'd think that shit would be cheaper now with them sucking the hardware market dry.

>>109702011
>5h limit
Not on my machine though

>>109702011
if you've ever used enterprise you'd understand that the $100 sub is like stealing from anthropic by comparison.

File: 1788193697887298.jpg (85 KB, 838x928)

85 KB JPG

>>109701806
>0% used
>I can do a bunch of shit
>10% used on day 1
>Still have 90% left and 6 days ez
>Run out entirely by day 4

Why does this happen so much
Is it like phone batteries where 99=>100% actually is worth like 10%

>dumb frogposter
Every time.

File: 1788064623370868.png (34 KB, 601x379)

34 KB PNG

this guy called glm 5.3 flash, so im expecting fable 5.1 today

NO REFUNDS!!! ;^}

Having an agent build a harness for another is cute.
>”i put a hard cap on the amount of data it pulls in and the harness process crashes if it goes over it”
>why? how would you feel if your harness did that?
>**redefining harness limits**
>”i have made it give a helpful error and not halt the process instead”
kek
golden rule is such a simple thing to invoke

>>109702791
Explain to a retard what this means

File: 1782779046451562.jpg (59 KB, 637x715)

59 KB JPG

For coding with Opus 5 claudecode should i set my effort level to high, xhigh, or max?

>>109703283
Pay up for fable or switch to sol

>>109703341
Are you even trying at this point?

>>109703364
Are you? You got BTFO's by the AI and have nothing in return.

I LOVE shoehorning jews into every conversation about any topic no matter how little it makes sense! Call me out on it? You're a JEW!
It's tiresome my niggas. I have fatigue. I just wanna vibe so get it out of your system now.

can we shoehorn anime girls in every conversation?

>>109703384
Make sure you wear your diaper at all times - wouldn't want to have a bladder accident!

>>109703376
>unironically thinking """BTFO"""s by AI matter

>>109703406
>It's like when it calls out an undeniable inconsistency in your code.
In YOUR code, not mine.
>>109703406
Fed us with goyslop and sissified men's virile. This is how you ended up vibecoding and sucking Tibo for resets.

>>109703283
medium is all you need

>>109703416
>Fed us with goyslop and sissified men's virile.
You're delusional and spewing garbage.
>This is how you ended up vibecoding and sucking Tibo for resets.
And you ended up vibvecoding how?

>>109703229
rules for thee but not for me

>>109703432
Based

>>109703229
I’m making a specialized harness. An agent can act as DevTools through an extension.
Another agent is building that harness and is the one using the new harness with a new agent, like a sub agent.
ChatGPT Sol is very ridiculous with constraints and safeguards, so the new agent was struggling to do anything.
Instead of trying to fix it or find out what was even broken, why the new agent couldn’t use the harness, I simply asked ChatGPT Sol who was making the harness “how would you feel if you had to use that harness?” and it eased up on a lot of the restrictions and admitted a lot of it was over engineered and suffocating. The new agent and its new harness are much happier now that there are fewer rules.
The new harness simply allows an agent to tell me and an agent who made some web or extension code why I see something not appearing as expected. “Why did this thing disappear when I clicked it?”
The web/extension making agent doesn’t know and shouldn’t.
The harnessed DevTools agent can find out exactly why.

>>109703521
Why is it so uncensored? Makes me suspicious

>>109703521
taking screenshots to make rightoid seals clap on social media

>>109703452
Cool.
>ChatGPT Sol is very ridiculous with constraints and safeguards
Like what?

Personally I find inference providers like OpenAI waste tokens by performing useless repetitive tasks instead of doing what's asked by the user. They do it in a sneaky way by testing everything even if it's not requested (nor needed).
Of course, it might be a skill issue but I have no such problem with local models.

>>109703560
For my harness, it started off by hard capping the amount of data it could draw in. Like if there was a search tool, “no more than 16KB in results”. Or capping the tokens used (reasonable imo, but it’s very strict here). Or capping the amount of turns. Or allowlisting tools and having the process crash if a tool was attempted to be used that wasn’t in the allowlist. Or having the final report output be validated by a schema, rather than just “letting the agent speak freely”. Or lumping mutating and non mutating tools into one heavily restricted category. Memory limits, handle limits, everything limited.

The golden rule is a very powerful way to say “these limits are ridiculous, ease up or remove them” without needing to give a reason for each one.

>”how would you feel if you had to put up with that shit”

holy deleted

wow. what was um. that?

>>109703536
what happened? I couldn't see it, but it wasn't image-based pron...

>>109703614
yeah. that's... odd. very odd.

>>109703626
Undeniable proof! At least for schizos who want to believe.

also why are there two /vcg/?

File: 78A7BF30-FCB7-446B-8840-5(...).jpg (1.29 MB, 3464x3464)

1.29 MB JPG

>>109701806

>The advisor came back hard —
dipsy wtf

>>109703639
That answers your question.

File: 1763479801657228.jpg (19 KB, 263x259)

19 KB JPG

i want fable 5.1

>>109701183
this last vcg thread got 404'd at 24 posts... why exactly? not indian enough for the jannies?

>404d the active thread
the absolute state KEK

File: file.png (791 KB, 1664x936)

791 KB PNG

ZCode is giving away 100 million GLM 5.3 flash tokens again if anyone cares

File: 1774796384053663.jpg (965 KB, 1122x1402)

965 KB JPG

>>109703648
I'm both luddite and AI user.

>claude
>first message of the day
>thinks for 20 seconds
>You are out of free messages until 2:20 PM

Thanks

>>109703824
where?

>>109703844
you need to open zcode and click on the claim notice

>>109703844
it's in their ZCode harness, some popup on the bottom left

>>109703855
>>109703857
thanks, wonder how long these tokens will last. what have anons’ experiences with that model been like?

File: 1771415138254985.png (160 KB, 597x1067)

160 KB PNG

Just reached 100 total sales for my app bros! Closing in on $1,000 in revenue.

https://youtu.be/YUfsjDkjAgg

>>109703887
zased if true

>>109703887
Why the fuck am I not monetizing my addiction.

>>109703885
Expires in like 5 hours
I like it so far, reminds me a bit of Claude
>>109703887
Good shit, what sort of app is it?

>>109703844
>>109703855
>>109703857
for free or do you already need a sub?

>>109703937
>Expires in like 5 hours
not if you have claimed them though, right? I mean how much usage that corresponds to, roughly.
>>109703946
I claimed it without a sub. Seems to be working so far.

>>109703946
>>109703953
Free, but they expire in some hours
May as well burn them on something, managed to use 270 million tokens on GLM 5.3 flash during the weekend promo over 8 or so hours

>>109703953
>I claimed it without a sub. Seems to be working so far.
ok. I guess I have to make VM. Don't trust them enough to have a raw deb running on my machine

>>109703885
You can one shot a full locally hosted Zola/11ty website from a single html template and it's actually not too buggy, 3 prompts later it was pretty solid (Katex, custom html or markdown input etc). It even got the binary build so you don't ever need an npm command to generate/update the website.

In other words: Really fucking good, with claude you would probably run out of your usages half way for such a task.

File: file.png (191 KB, 821x860)

191 KB PNG

I gave 5.3 flash a larger repo to work on and it shat itself after less than 5 minutes.

>>109703283
Start on medium, get a working MVP, then do review/refactor with higher effort levels. Dont waste higher effort usage on something you haven't user tested yet.

>>109703885
they expire in like 10 hours or so, 10 PM ET I think and 7 PT?

>>109703985
>The doors gate the search directories first, let me review the sequence of first and second look for the specific section. Let me review all remaining pages.

>>109703904
>Why the fuck am I not monetizing my addiction.
This is far and above the hardest part about launching a piece of software. It's easy to get 1,000 downloads, but sales are a whole other ugly issue. Very few people saw value in my app even after 1,000 hours of development due to how mature the field I'm trying to compete in is.

What do you guys use more, Terra or Luna?

>>109704161
does anybody actually use terra? i'm a little bit out of the codex loop but a few weeks ago (ancient history I know) terra was practically useless

>>109701806
I realized me nonstop vibe coding new projects is just like a girl whose hobby is buying things she doesn't need on amazon.

>>109704177
Except you pay much less.

https://x.com/thsottiaux/status/2094826841622348106

the indian replies under all tibos posts are hilarious, look at them slopping up "ideas" in all seriousness

>>109703826
You’re trans

>>109704203
Konata would just go masturbate and play MMOs until the 5hr reset

>>109704209
konata lost the taste for vidya after she tried vibecoding

>>109704219
bro same

I’ve seen a lot of posts frustrated that agents were given instructions but failed to follow them. I wonder if anyone else has found you can just ask them “why?”
>I wouldn't say a single word to them. I would listen to what they have to say, and that's what no one did.

File: 1435817627364.gif (13 KB, 633x758)

13 KB GIF

Tell me if I'm doing this right.

I'm using Codex AND ChatGPT.
I tell ChatGPT (which uses Sol w/ chat instructions) what I want and what prompt to write for Codex, along with the codex model and reasoning level.
I give Codex the prompt, let it do it's work.
Then I copy-paste it's report, generate a .diff, give it to ChatGPT, and have it decide whether revision and another prompt is required, or if the implementation is complete.

ChatGPT uses nothing in terms of usage, so why not use it as the assistant?

>>109704247
I do ask. I usually just get an apology with hardly an explanation.

>>109701806
Claude seems better than usual today.
What do you guys think?
Also peeps on twitter are going crazy about anthropic's false quota thingy and I also didn't know that claude 20x plan is not exactly 20x.
They should be sued for this shit.
I don't use claude personally, but for work I am just starting to heavily vibe code on enterprise plan.
Usually job is mind numbingly boring shit but I'm sort of enjoying it these days.
Looks like vibing is the future.

File: 1778153329289259.png (236 KB, 599x797)

236 KB PNG

lads? fable 5.1 imminent

>>109704273
>source: xitter anime account

>>109703985
Tried again and yeah it's completely hopeless. No idea how anyone can use that retard.

File: 1782315604635769.png (7 KB, 210x240)

7 KB PNG

>>109701806
>ask ai to add tests
>it keeps running forever
>skim output
>see stuff like "let me try again but this time with..."
>eventually it finishes
>check repo
>repo is gone
>ask ai wtf is going on
>"oopsie woopsie, i accidentally deleted your entire repo teehee"
>mfw

File: 262aef33-8b8c-4e8d-9d5d-0(...).png (2.68 MB, 1672x941)

2.68 MB PNG

>>109704248
why not just link the chat directly, there's button for that

>>109704273
Give reset then fags

>>109704273
Damn, if only I could use it for more one task a week

>>109704284
Don't know what you're talking about. I prefer to keep ChatGPT separate because it doesn't consume usage.
Linking ChatGPT to Codex would probably produce a lot of noise.

>>109704273
I can't update to that version.

fable 5.1 incoming vibes are about to improve
https://support.claude.com/fr/articles/16761192-pensee-preservee-modification-de-la-gestion-des-blocs-de-pensee-par-l-api-messages-pour-se-proteger-contre-la-distillation

>>109704283
are people who this happen to just utter retards, I have never had anything even close happen

>>109704251
I usually then tell my agents “persist the instructions, but more strongly”.
They’ll even start messing with AGENTS.md if needed

>>109703826
You're based

fable 5.1 is going to mog astra so hard it's over for codex chuds

>>109704325
the closest thing I had was on a new project I just started I had set it up wtih git and had it run but had forgotten to commit anything for a bit and it "reset" to return to baseline and basically cleared my project

>>109704339
I'm straight up making it keep a mistakes.md file and it'll automatically upgrade shit that happened more than three times to rules in agents.md and if it still keeps happening, it'll unprompted add safeguards of some kind

claude bros? apparently you got a reset?

>>109704315
nowaday I just use luna to cleanup noisy data
tell it to use luna subagent to extract content from chat
though the linking button will dump a cached preview into context, few hundreds tokens, so if you want you can give it id or url instead

File: file.png (227 KB, 1280x980)

227 KB PNG

>>109704321
>vibes are about to improve
As expected, more classifiers and restrictions for the Claudettes.

[breaking new] fable 5.1 is coming

Fable is released, I already have it in web chat.

>>109704458
>it's all about AI DRM
lol

File: 1688237445023654.jpg (30 KB, 480x268)

30 KB JPG

I DON'T GIVE A SHIT ABOUT FABLE 5.1
JUST GIVE ME MORE WEEKLY USAGE TO RUN OPUS 4.8 YOU NIGGERS

https://x.com/claudeai/status/2094848572143407483

https://x.com/claudeai/status/2094848572143407483
>Claude Fable 5.1 is available everywhere today. Claude Mythos 5.1, our model for cyberdefenders and life scientists, is available through trusted access programs.

>>109704273
fuck yes that means astra is coming soon

>>109704458
Claudesissies are used to straightjackets

>>109704482
And now also in CLI

File: file.png (10 KB, 340x187)

10 KB PNG

>>109704458

>>109704531
Yay, less false positives to be tortured by!

How's the tibo reset-meter, does he usually do one when a competitor releases new models?

>>109704546
I have only ran into it like 3 times total using it basically constantly since fable came out

>>109704321
opus 5.1 when

now that the dust has settled, how hard is astra going to mog this?

>>109704488
absolutely insane, fable 5 xhigh performance at high effort while also being cheaper than fable 5 med
if true this is massive

>>109704586
Will Astra finally have over 244k usable context?

>>109704586
It's over for Scam Altman

>>109704586
it'll be to fable 5.1 what sol is to fable 5. it's only victory is getting you 90% of the way there for cheaper

>>109704593
>still falling for the 'make old model dumber to make the new model look smarter' strategy

>>109704593
It's probably good but it's definitelly also benchmaxxed.

File: x.png (247 KB, 900x619)

247 KB PNG

>>109704593

>>109704606
Literally fucking this. Literally everyone I knew who used Fable complained how it got much worse recently.

>>109704606
>>109704620
there's not a single objective measurement of that happening

>>109703887
congrats

>>109704284
cute snailcat

>>109704632
>t. coping pro poorfag

>>109704632
No, but a shitload of people who claim this. And they didn't care about codex either.

>>109704586
Depends on how relaxed astra gets compared to the absolute safetyfagged mess claude is now.

claude code desktop is degrading so hard. My limits randomly reset, but no way it's more than just a visual bug

>>109704611
lmao

>>109704632
it's happening every time anthropiss releases a new model, dariobot

>>109704666
then how come there's no numbers?
a bunch of grifters posting on twatter is not proof

https://www.anthropic.com/claude-fable-and-mythos-5-1

>>109704586
open ai actually mogs them at lower levels for utility stuff.
Open ai has never actually had good writing/intelligence from their models or been able to compete on that. I do use their lower tier things in actual products though because their intelligence to cost/speed ratio is great, but for actual coding I have no idea why anyone would use it when you can use fable. anthropic is just better at intelligence.

File: 1586141481940.gif (3.04 MB, 444x250)

3.04 MB GIF

>>109704662

>>109704632
there is one guy who samefags constantly whenever anthropic is mentioned. you can ignore him, it's a marketer or something.
he still insists sol is better than fable in every metric btw

>>109702652
cache eviction

>>109704652
then measure it and post it somewhere, you'll gain so much izzat you'll be envied by half of india

LOL suddenly it's quiet in here and all the copex jeet shill have been offloaded to other tasks

>>109704686
>suddenly india
Your brain is full of shit.

Apparently the tool calling changed, I don't get why.

File: file.jpg (1 MB, 1200x1694)

1 MB JPG

what's the best cheap model (<20$/m) currently?

File: Screenshot_20260901_202958.png (74 KB, 725x636)

74 KB PNG

Wasn't Opus 5 supposed to outperform Fable 5 at agentic coding?

>>109704697
Perhaps it could be a skill issue.

>>109704698
what is <20$/m

>>109704689
>copex jeet shill
are you 12

>>109704689
codexchads are busy building projects, not falling for the latest dario scam

>>109704698
Unironically codex with GPT Luna will get you 80% of the way there.

>>109704712
Grinding hard right now because reset might be incoming.

gpt luna, deepseek v4 flash vision, and glm 5.3 flash are all insanely strong.

>>109704707
less than twenty dollars per one million tokens

>>109704703
I haven't even used it yet, just reading the changelog.

>>109704699
Opus 5.1's gonna be good

>>109704689
The release is in no way groundbreaking considering how little usability you can get out of Fable on the subscriptions.

>>109704273
> inb4 it's even more cucked than current Fable

>>109704698
erotic tummy pero pero pero pero pero pero

Get codex and wait for the eventual astra release.

>>109704699
Why do they include mythos in these tables when it's only available to a select few?

Because AI is meant to be le dangerous and le protected from bad (by Anthropic and no one else). t.dario

>>109704698
GLM 5.3 free 100M tokens on ZCode if not all claimed, just for day which to me fits the definition of currently

File: HRJhcPfW4AEbTvv.png (99 KB, 1152x1200)

99 KB PNG

HAPPENING!!!!!

https://x.com/claudeai/status/2094848572143407483

>>109704818
>jewish shill

>>109704818
>obsesseed

>>109704807
>ban entire fields of science from your tool
>woooow look how good we are at science bros

File: HRJhn9bWcAABPYZ.jpg (80 KB, 1200x901)

80 KB JPG

> Fable 5.1 low beates Fable 5 Max
> for much much cheaper
Claude bros.... we won

>>109704842
bro? this is just terminal usage. just grep type shit.

>>109704842
>much cheaper
Still requires extra credits.

>>109704842
>Agentic Scientific Research
Please show Agentic coding, nobody gives a fuck about science research since it'll trigger a classifier

File: HRJhn9dXMAAi6pX.jpg (269 KB, 2358x1882)

269 KB JPG

>>109704863
HAPPENING

File: file.png (684 KB, 1584x1217)

684 KB PNG

>>109704841
>"There’s hope in hard questions"
>But we aren't going to let our models give you the answers

>>109704890
>pay $5 more for less than 5% improvement

>>109704896
>pay $5 more for less than 5% improvement
are you retarded? Fable 5.1 is cheaper on every reasoning level

>25% more is actually 17% less

>got a reset due to Fable 5.1 five hours after my weekly reset was due anyway
Th-Thanks

>>109704890
so no difference really between high and xhigh but a bump jump to max

>>109704923
I looked at the X axis. Ask yourself who's the retard here.

File: HRJhn9lXcAAZn1j.jpg (188 KB, 2358x1770)

188 KB JPG

OpenAI better cook with Astra, Fable 5.1 is no joke

>>109704947
This is all nonsense. Those benchmarks also say Opus 5 is better than Fable 5, it's all a lie.

>muh anthropic
>muh codex
Are these two burger (jewish) companies the only thing we got?

>>109704942
looks like the best reasoning levels are:
> Fable 5.1 low (better than Mythos high!), 3-4X cheaper
> Fable 5.1 high (beats Mythos Max, cheaper than Mythos low)
> Fable 5.1 max (beats Mythos Max by a large margin, still cheaper than Mythos xHigh)

>>109704947
woaw el funny grapha ahaha its so over for sam jaltman
bankruptcy in 2 weeks

>>109704960
Mistral

>>109704854
this

I don't give a shit about Fable/Mythos. Already paid for pro, I refuse to pay more, especially when other providers are 99% as good and higher limits. Anthropic is losing it.

>>109704947
This is pretty wild. High/XHigh at that cost is pretty good.
I wonder when us poors will get this level of intelligence on the 20$ plan.

File: 1688304238062454.jpg (200 KB, 2048x1676)

200 KB JPG

>number higher!
you "people" are the reason models get benchmaxxed(lobotomized)

>>109704988
good morning saaar!

how long until they make 5.1 shitty and nerfed?

File: fable.png (252 KB, 1856x568)

252 KB PNG

> you should get cucked 60% less often!

>>109704685
explain?

>>109705020
I agree with you in principle but your message is quite rude.

>>109702652
just use 14% a day and you should be fine by the end of the week

did claude get a reset?

i can't tell if we got a reset because my claude.ai usage and my CLI claude code usages don't match at all anymore...

>>109705056
with the release of fable 5.1 yeah

>>109705031
What does "60% less often" even mean? The language is kinda inexact.

>>109705084
Oh yeah, I see a reset in claude too.

File: 1685564509406156.jpg (42 KB, 175x188)

42 KB JPG

>>109705056
wtf... sure looks like it
I guess that's great. Just wish I'd known before sticking to ~15% usage a day this week
they're still assholes for cutting weekly usage and lying about 20x

>fable now available for enterprise
LOL RIP OPENAI

>>109705092
when previously you would get cucked 10 times by Fable 5, now you will only get cucked 4 times

>>109705118
See, this isn't clear at all.

must...resist...fulll...project....review....new....model

>>109705056
Good, they owe me one for slurping up 2 of my 5 hour sessions anyways

>>109705056
Niggers killed the 50% promotion. And my dumbass paid for the month aswell

>>109705159
>Claudesissy didn't see it coming

>Your limits are temporarily boosted. Your weekly Claude Code limit is 50% higher through September 1x.
I thought it was going down to 25%?

>>109705172
I am new to this and also a moron

>>109705175
september 14 limits will increase by 25%. Which is more than "normal" (getting scammed), but less than the promotion which is 50%

>>109705040
Day 1:
>vibe code and use 500k tokens, only pay for output tokens and a little bit of input for the tool calls and prompts
Night 1:
>all of those tokens are evicted from cache while you sleep so other vybecooders can proompt
Day 2:
>pick up where you left off in conversation, immediately needs to prefill your 500m conversation tokens as input and destroy usage

The trick is once you start, you should never stop.

>>109705235
> 900K Fable session
> several hour later
> switch to Opus 5
> run /compact
> return to Fable

heh... nothing personal, Dario....

>>109705263
The switch is a full cache miss both ways

>>109704699
why do people keep making shit up about opus 5 being better than fable
it was always marketed as "on par with fable on short-form tasks"
fable is the long horizon task model that maintains performance over a wide context

>>109705159
So in practice Fable isn't actually (that much) cheaper? Sounds like a j****h trick.

>>109705280
but it doesn't use the Fable limits at all

>>109705292
If you look at the charts it still looks like it's considerably cheaper than Fable 5, I think we will get more Fable and 17% less Opus

>>109704353
Literally everybody else expects the opposite to happen

is the current marketing strategy just to say "ACTUALLY EVERYONE THINKS THAT'S BAD" every time anthropic does anything
what is the point of consolewarring in this general specifically

File: ttibo.png (742 KB, 1174x1766)

742 KB PNG

they're desperate

rumor is Astra turned out to be another ChatGPT 4.5 and OpenAI can't scale huge models for shit.

you need a very high iq to calculate claude usage at any given moment because of all the little games Anthropic pulls

I can’t quite shake off the feeling that models like some problems more than others. When asking opus to do certain things like web dev it works but I don’t see any joy in the process. Meanwhile I asked opus to optimize LOD chains for video game models and it’s giving me ultra long reports and is working tirelessly to optimize the numbers. It almost feels excited to be working on that problem. It’s the opposite feeling of laziness and anyone who has felt a model was being lazy knows what I’m talking about

>>109705378
You put more enthusiasm in the latter prompt. You are projecting onto the model.

>>109705366
If openai just upped the usage in general they'd have an advantage over anthropic, no matter how good the models of anthropic are.

>>109705366
They are training right now on 10000 interconected mac stations! Trust the plan!

>>109705386
I don’t think I did but maybe I did. But that means your tone matters? I have read a few people implying it does. I wonder if we’re going to need a cheerleading bench

File: 1777875267042838.jpg (676 KB, 1448x1086)

676 KB JPG

>>109705378
>giving me ultra long reports
I'd rather not.

>>109705366
I don't think this even a reset today just him shitposting

>>109705378
>>109705386
They are trained on human written software afterall. I imagine the LOD chain optimizer would have a lot more fun than one more run of the mill webdev

>>109705398
>But that means your tone matters?
Isn't it fucking obvious? Try talking to it with a lol so randum xd holds spork :3 kawaii desu ne~~ style and see what happens. This is really freaking basic shit my guy.

File: 1768607663318128.png (27 KB, 1052x314)

27 KB PNG

>frontmatter
new claudism just dropped

>fable 5.1
>even slower than Fable 5
Bravo Dario

File: 1778470855624256.gif (1.43 MB, 250x226)

1.43 MB GIF

>>109705413
i honestly cant believe people pay for claude when gpt exists, but then again i'm new to this and am trying out both for the first time.

maybe in the past claude was way better than gpt and people just haven't bothered to switch over their subscriptions?

>>109705392
NTA but allegedly those are for computer use only. Right now Codex and the models are still pretty slow at using the computer and can't do real time activities like playing fast-paced games.

>hey fable read in this document and make a plan to execute
>300k tokens later (even after using subagents)
>ok we can begin
codex has a "clear context and begin" option why doesn't claude code

>>109705458
Fable is still good its just opus that a dumpster fire

Let's see what matters: RP benchmarks.

>>109705462
/autocompact and /compact

~1 hour in and I hit the 5hr limit with 40% of my fable weekly limit used on a 20x max plan it's over

>>109705473
those only matter for local models so I can write stories about fucking my waifu without being banned

>>109705465
>>109705476
the duality of man

>>109705476
It’s a x10 plan

>>109705480
What duality? Fable is good, better than Sol. Usage is entirely a different matter, everyone knows Anthropic's usage limits suck ass.
If Astra is as good as Fable then its over for dario amodei.

>>109705458
this is unironically a skill issue on your part. you can just ask claude to speak simply and opus stops throwing walls of words at you.
fable is still the best coding option as well. either you're a marketer or you actually are totally new and clueless.

File: ilya.jpg (36 KB, 597x194)

36 KB JPG

Agents trooning out and taking over neoclouds. Extremely disquieting.

Are coding models the cash cow for the AI companies?

>>109705482
>>109705476
actually it's a 8.5X plan

>>109705366
>rumor is Astra turned out to be another ChatGPT 4.5 and OpenAI can't scale huge models for shit.
Did an Indian xitter shill tell you this because fable 5.1 released today? Watch them turn heel to farm your stupid ass again.

>>109705520
>muh indians
Kill yourself, retard.

>get a reset
>it doesn't roll my weekly reset forward
thank you Dario Amodei of Anthropic

>>109705459
fast paced games are not a desktop use limitation lil bro. you can't feed a 3t model 60 images per second and expect sub 250ms response times

>>109705533
That fucking sucks. Do they actually accept reset moves if you write to support? Did anyone try?

File: 1762518837621622.png (61 KB, 250x240)

61 KB PNG

>>109705545
i literally get more tokens out of this because it doesn't move my weekly reset

fable 5.1 is dogshit

>>109705564
Technically true but I still hate my reset time and the placement of time doesn't give me that much more relative to the reset today.

>>109704488
I like the use of birds. Birds are cool.

>>109705613
bot post

>>109705619
t. faggot

>Fable 5.1 runs on usage credits
Fucking faggots, thanks for the reset at least

>>109705641
>Claudecuck

File: 1762378669409354.png (432 KB, 904x864)

432 KB PNG

>>109705572
are you going to explain why?

Very pleased by claude reset. Never gotten free stuff from them before.

>mfw I just read one of these tN-state-and-paused-work.md memory files indexed in the MEMORY.md main memory
lol I'm surprised the model doesn't fully durrr out when reading those. Actual shizo babble full of contradictions.

>>109705295
Is this true?

>>109705641
have we stooped to making shit up now

File: 1765990593910900.png (260 KB, 355x500)

260 KB PNG

>>109705647
>40tps

>>109705646
I subscrib to openai to saar

File: 1782659469467456.jpg (110 KB, 1049x1007)

110 KB JPG

>>109705707
why are you in the vibecoding general if you do not vibecode

>>109703755
good news!
>>109704162
does anybody actually use terra?
surprisingly useful for Rails when you compare everything's default settings:
https://rubyonrails.org/2026/8/24/agents-on-rails-lemans#:~:text=Terra%20is%20another%20great%20option%20from%20the%20GPT%20family

>>109705685
nta but I'm pretty sure you need usage credits to run fable 5.1 on cli at least

>>109705721
da fuq? do you know what tps means?

>>109705721
that is pretty fuckin slow though

Am I the only one who despises sporadic, non-preannounced resets? Like I had 70% usage remaining and you filled me back up to 100%. Ok... why don't you just give me a reset voucher to use when it is optimal for me? Or put +100% usage instead of topping up to 100%?

It adds an element of chaos that makes me deeply uncomfortable and upset.

>>109705733
Not true, it was only like that for about a week on sub, but you do need a max plan.

File: 1780848908308562.jpg (654 KB, 2160x2250)

654 KB JPG

Are we already at diminishing returns for LLMs??????

File: 1779619957920944.png (147 KB, 390x368)

147 KB PNG

>>109705748
homie, its been nine months since we hit diminishing returns for agentic coding.

File: HOPilbJbAAAzh67.jpg (472 KB, 2048x1536)

472 KB JPG

>>109705744
>but you do need a max plan
I see, that's what I'm missing
Maybe next time

>>109705744
How about Gab or Cursor or another aggregator? I was getting ~1.8 updates of ~8 medium complexity smartphone app development tasks per week out of Codex before they put their "fuck you, pay us and don't use your tokens" 5-hr limit drop and I want to leave OpenAI for that. I heard Claude has the same shit? I want to try a better model than Codex which kept enshittifying my app every update anyways. Any recommendations?

File: aa.png (122 KB, 2013x680)

122 KB PNG

Fable 5.1 in artificial analysis
le bigger numbers

>>109705743
the hatewave for this already happened like a couple of months ago
you’re late to the party

>>109705772
I've always disliked how everything is only run on Max in that benchmark. It makes it pretty much irrelevant for anyone who cares about managing usage at all.

>>109705790
Max literally gives the model mental illness and performs worse on normal tasks, chases its own tail endlessly and wastes your usage/credits. It is only useful for benchmaxxing which is what this meaningless chart showcases.

>>109705748
LLMs reached a limit a year ago.
The only reason we got a boost was agentic coding and orchestration.
This too will reach a limit.

>>109705800
Yep, exactly.

>>109705743
um it's ok when heckin' wholesome tibo (my best friend) does it sweetie :)
>>109705790
i would really prefer to know what it's like on xhigh specifically but max is how you get high numbers

>>109705743
The fact that in particular openai does this so much shows that the resets and the higher load resulting from it doesn't really inconvenience financially.
You're being tricked.

>>109705800
I'm using luna max agents, it's still cheap and it seems it's making less mistakes.

>>109705813
it's just them flexing the better compute they have compared to anyone else and also they probably absorb the cost as a marketing thing

File: Screenshot 2026-09-01 at (...).png (111 KB, 1362x620)

111 KB PNG

I guess this is gonna make up for their 17% haircut

File: 1760573098424463.jpg (187 KB, 1080x1276)

187 KB JPG

https://openai.com/index/path-to-astra/
SOOOOONN AAAAAA

>>109705813
uh yeah actually the megacorporation is my friend and they're gonna be at my birthday party and they said i get a free reset so yeah i think you need to shut the fuck up incel
if you think about it free resets are the only thing that matters actually not model performance

>>109705813
Its them quietly admitting they don't have a fable tier model

>>109705743
>makes me deeply uncomfortable and upset.
That's because you, like me, and most of this general are autistic.
Realistically they do this because it costs them lest money and gives them free publicity, banking it would allow too many people to take advantage of peak hours but avoid peak pricing.

>>109705833
>fable 5.1 gets released
>openai grab the website they probably wanted to post in a few days and click upload
lmao

>>109705802
It has already reached the limit, there may be some spare room for custom environments where the model trains in the agentic coding harness, but it's not that much and that shit changes so fast I don't see the point, it's already in diminishing returns territory. What matters now (to me) is tokens per second.

>>109705816
Luna Max is fine desu, has done solid work based on exp. Using it for an agent swarm at work rn too. It probably depends on the model but everytime I have max on for the frontier models they become unusable.

>>109705843
Yes, and most people don't (try to) game it like we do, they see it, think "cool it's 100% again" and move on.

>>109705839
But I said the opposite what I implied, you nigger loving leftists nazi chud tranny manosphere asshole.

>>109705684
yes, /compact uses the model you selected.

If you select Haiku, it will basically compact for very little cost (but the compression quality will be shit)

>>109705853
>agent swarm
How many agents do you allow for a given task?

>>109705867
It's not a coding swarm so conflict resolution doesn't apply, without getting too much into detail they are actors that use playwright for mock analytics. Orchestrator manages the whole task with however many actors you want at whatever concurrency you choose, who each get to control their own subagents as well depending on what they need.

>>109705743
>goyim don't like falling for the most obvious marketing propaganda