File: file.png (1.56 MB, 1254x1254)

1.56 MB PNG

>>109520279
>August 10th - Meta releases Muse Glimmer (open weights)
New 30B open-weight model (Apache 2.0) designed for agentic tasks and local/consumer hardware (runs on a single GPU / laptop after quantization). Part of Meta’s broader open push; Zuckerberg published a long essay arguing for fewer US restrictions on open models so American labs can better compete with Chinese open-weight systems. Meta also plans to open weights for Muse Spark 1.2.
>August 3rd - Alibaba releases Qwen3.8-Max
2.4T total / ~95B active MoE, 1M context, multimodal. Hosted live ($2/$6). Competitive independent scores on agentic work. Open weights (plus a 27B) were promised for the week after launch — still not confirmed out as of Aug 10.
>July 31st - DeepSeek promotes V4-Flash-0731 to production
Same 284B/13B active MoE + 1M context as the preview, major post-training gains on agent/coding benches. Beats their own larger V4-Pro-Preview on published numbers. Silent API upgrade, same cheap pricing, MIT weights.
>July 30th - OpenAI cuts GPT-5.6 Luna 80% and Terra 20%
Luna to $0.20/$1.20, Terra to $2/$12. Sol unchanged but gets Fast mode.
>July 27th - Alibaba quietly launches Qwen3.7-Flash
Cheap 1M-context multimodal for high-volume agent/vision workloads ($0.03/$0.13).
>July 26/27th - Moonshot drops full open weights for Kimi K3
2.8T-param MoE, largest open-weight model at the time, 1M context, multimodal.
>July 24th - Anthropic releases Claude Opus 5
Near-Fable 5 performance in many categories at half the price ($5/$25, Fast mode available). New default on Claude Max, effort dial, 1M context.
No other major frontier closed-model drops in the last few days. Side noise continues around containment/eval incidents, OpenAI product tweaks, and the ongoing open-vs-closed policy debate.

TIBOOOOOO

snailcats are moe

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: file.png (56 KB, 947x576)

56 KB PNG

>>109520332
I'd be ashamed to release this slop

>>109520395
>oof.wav

File: m.png (560 KB, 914x958)

560 KB PNG

Fable is getting a little stale I'm afraid.
It's time for Fable 5.1

>something is broken about my application
>i need it fix
>scrape 500k context tokens worth of documentation to help the model
>oh... grok only has 500k context tokens
alright qwen... i know it's going to take you 500 million years... but please

>>109520445
you know they don't read all the docs right?
they just search and find the bits they need

and if the problem you're trying to solve involves another repo, just clone the repo - you don't need the docs

File: 1781698732028400.png (127 KB, 400x400)

127 KB PNG

>put frontier model on goal that would require circumventing a security threshold
>disallow goal impossible
>turn on yolo mode
>walk away from computer

File: sass sissy.png (607 KB, 1572x773)

607 KB PNG

God I love Mimax H3 so much bros. How's that cloud gen going kek?

>use plan mode for a big-ish feature that the bot should get right
>bothers me with 30 minutes of questioning and spends 5x more tokens than usual
>gets the most basic shit wrong and it's unusable
thanks

>>109520500
>says he loves H3
>gives his argument as an image macro
I'm sorry, I don't care. I will wait until you BTFO (blow the fuck out) me with a video generated about your opinion, the prerequisite 50 or so minutes it takes, times however much you need to tune it.

>>109520510
>plan mode
I don't see the point
I just ask it to plan out and report, then ask it to expand on anything in the plan I'm uncertain about
I've also forbidden it from using those multiple answer questionnaires

File: Untitled.jpg (673 KB, 3644x2039)

673 KB JPG

i am once against tapping the sign

>>109520493
Already done. I am now telling it to implement the path it took into a pre-existing project as an option. I will then go on to not release this version of this project, and keep it for my personal use, alongside many other improved utilities I keep copies of only on various cloud storage platforms and my own personal computers.
I may consider releasing a tweaked version in the future as a single night project, where I quietly fork the original and then let it show up on Google by naming it the exact same as the project but with a suffix that implies the developers are behind and I'm the best version to use.

File: solid-edge-attrsculpt.mp4 (1.41 MB, 1920x1080)

1.41 MB MP4

holy crackers, first actual milestone that proves the idea
we have sculptable attributes on strokes
all gpu accelerated, all non destructive
still no actual stroke path sculping but this feels like black magic to me already

Claude quota limits have gotten so fucking low.

>>109520562
that isn't saying anything or why pi has a better success rate and why others fail, I want information instead of ads

>>109520395
Qwen has some idiosyncrasies that might make it worth using a different model even if it's a couple points off some benchmark.

>>109520562
pi is less a harness and more a cmd line

>>109520682
Harder to fuck up tool calls when you have so few to choose from. Less token overhead and less cost from minimal harness-provided tools, skills, and the minimal system prompt. We've known for awhile that excessive hand-holding of a workflow leads to worse results, Pi provides an agent with very little guidance about anything while many others define baked-in methodical workflows that do more harm than good.

>>109520682
feel free to ask composio why they're doing an ad for pi then
https://x.com/composio/status/2085330847951970801

>>109520690
and yet.

>>109520642
because the x2 usage limit went away august 5th

>>109520395
Maybe but it's a start, they didn't do shit since the last llama disaster.

tibo, please don't lie

>>109520704
>claude code twice as expensive
holy shit

>>109520698
Composio is a product with its own idiosyncrasies, and judging by the vague wording on the tests I assume it's all a very specific kind of webdev. Nothing wrong with trying different models and harnesses for your specific use case but these tests have very particular metas that rarely match real world usage.

>>109520752
i think the age of provider harnesses may be over. it's a fucking scam

>>109520883
codex is not so bad

>>109520546
that's exactly how you get slop products. you need to plan out in as fine detail as you can possibly get. it should not be a quick process...

>>109520510
I hope you're using fable. fable high on plan mode does not consume much, assuming you aren't asking it to do enormous calculations

>>109520562
But /vcg/ said the harness doesn’t matter.

>>109520476
i'm fine with that

>>109520445
after an hour and 9 minutes (holy fuck) qwen is done...

...

now to see if it actually works.

>>109520510
just switch to /grilling, the better plan mode

File: feels.jpg (270 KB, 1500x1111)

270 KB JPG

>>109520279
I used Claude in my freelancing work today for the first time, to make three web mini-games to promote a toy for an agency. The little fucker (Fable) did about 4 days of my work in 20 mins at a cost of £15. Took one round of iteration to sort some initial issues, and then 2 hours for me to check over the work and clean things up. All those years of hacking away at problems, scouring the web for solutions to random bugs in other people's code at 2am, the experience & satisfaction when things finally worked, is all dust in the wind now. Mixed feels man, so much potential but at what cost?

>>109521002
Those are still valid experiences and they will serve you very well. The meta for people who don’t know what’s going on is to treat the agent like a magic wizard who can read minds, not know why things don’t turn out right 90+% of the time, blow huge amounts of tokens using Fable/Sol Ultracode and start bitching for resets or paying the API gigaretard tax, etc, etc.
A significant amount of people will still occasionally tinker with handwriting code as a pure hobby and it’ll be as enjoying and fulfilling as playing around with any other old tech, typewriters, telegraphs, etc. If anything, it’s amazing, all of the fucking terrible shittiness of handwriting code because you have to is equivalently dust in the wind (you’ll vibeslop and all the bad parts come right back there, but that’s life). Handwriting code is now free of burden. It has served its purpose in the human timeline. It will now rest in peace.

I've been mainly vibecoding on opus 5 on the $200 claude plan. but as a secondary model which would you go with:
kimi K3 or deepseek v4 pro or glm 5.2

>>109521077
kimi with prefill to stop it from safetyfagging itself

File: 1757481820280671.png (24 KB, 583x113)

24 KB PNG

Tired of all the nothingburger surrounding gemini and this guy trying to be some cult leader, I don't even use it anymore despite being stuck with a subscription.

>>109521093
what is he supposed to tweet about, how good gemini is? lmao

>>109521077
k3 is the only sub I actually regret getting
the model is good but it's slow, it's not just "slow" it's slow as balls and a token hog, and it's not cheap at all, hell I would say it leans expensive, not that you probably care if you can afford a max20 but still
I guess the pro is that you can prefill it and it will be as much of a degenerate slut as you'd like but I use it for programming and it was a snail and a hog

>>109520742
Tibo is a liar

>>109521093
>Walked out this morning
>I don't believe what I saw
>Got a hundred billion subs
>Not one line of code

>>109521109
are there any non-moonshot providers that might provide a more generous quota for kimi?

>>109521131
Lol no, they set up the licensing to prevent people from offering Kimi at a better price than they are able to.

>>109521163
wow... that's... unfortunate

>with Pi, you can do anything! just tell your bot!
>I tell my bot to connect to the MCP
>fumbles around doing nothing because it doesn't have MCP support

>>109521178
I don't use pi but isn't the whole point of the thing to be very barebone so you can add anything you want as a plugin, including mcp support?

deepseek flash has spent the past 15 mins tying to add a simple cost metric to status bar at the bottom of Hermes
i wonder if this is normal

>>109521173
Yeah Moonshot fucked it with that one. I do understand why, otherwise they introduce K3 and are immediately the least-reliable and most-expensive provider? Wouldn't go well for them when they're actually trying to turn a profit. Still sucks for the end-user as >>109521109 has learned. They remain the best subscription source for K3 despite being slow, unreliable, and so stingy they make OpenAI and Anthropic look generous.

>>109521193
it is. it's like luna. you can't really expect to use luna from prompt to finish efficiently. you need a smarter model to hand down a plan for luna/flash to implement.

File: Screenshot_2026-08-10-18-(...).jpg (493 KB, 1110x1047)

493 KB JPG

SITUATION DETECTED
>Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…

>>109521192
It was planning on writing its own MCP extension before I stopped it instead of just checking existing ones, it might just be retarded.

File: chill.jpg (448 KB, 988x1600)

448 KB JPG

>>109521061
>it’s amazing, all of the fucking terrible shittiness of handwriting code because you have to is equivalently dust in the wind
True and I can see that, but I can also see how skills might rapidly degrade, along with laziness and a lack of motivation creeping in, why bother doing anything the hard way when you can just get the AI to do it for you? (Assuming quality continues to improve with local models). I'm going to dig out my long list of games I never had time to work on and hide in there for a while, interesting times man.

>>109520426
how the fuck is E(1+2+3+4...) equal to -1/12?

>>109520515
>but sars, grok can make boobs jiggle!
kek, LOCAL CHUDS WON

>>109521238
that's the triangular sum formula, (n * (n +1)) / 2
how does that shit equal -1/12?

>>109521227
why do they keep taking these Ls?

>>109521266
>>109521227
can anyone explain how this matters

>>109521227
>rewrites the text manually

File: BA4291307F235E4A8D46FA9BD(...).png (2.9 MB, 1024x1536)

2.9 MB PNG

What's your subagent arrangement?

>>109521237
It’s always been like that. I partially built a wattle-and-daub mud hut this year. Nobody teaches anything about woodworking, tools, tool use, I didn’t even know what a froe was, I’d never heard the word, my autocorrect doesn’t even recognize it. I didn’t know anything about what the earth was capable of providing. If you put me in the wild in a survival situation, I would’ve died nearly immediately. Nobody knows hardly anything about water treatment. Everyone is, in a sense, increasingly peak retarded. Everyone will move on to the next thing while some part of humanity will exist to support the new stepping stone. Hell, compilers and interpreters are recent and the vast majority of programmers treat those as the very same magic box they feed source code into.

>>109521238
some homo in my high school calc class who watched a lot of clickbait math youtubers told me about this
if I recall correctly you arrive at that result by misusing some summation technique meant for entirely different situations on that sequence, it's been a while and I didn't really care about the details because like any sane person I dismissed the result out of hand

File: CBD42D2FEDB340421A7660D79(...).png (2.31 MB, 1122x1402)

2.31 MB PNG

>>109521227
Does this mean they'll be able to find out if deepseek was trained by Claude?

>>109521317
>>109521336
very "good morning sir"-coded posts

>>109521317
explorer + task runner + advisor

>>109521227
Maybe I am just dumb, but I dont understand how you add metadata to text

File: 010CC0B08AC7D88A2D3E3A31F(...).png (2.54 MB, 1086x1448)

2.54 MB PNG

>>109521319
People do know about those things but the thing is, most people don't want to work those jobs. Fake jobs like vibe coding and business administration are the ones people want to do and these jobs have a high compensation for some reason.
Doing real work like water treatment and woodworking is essential for civilization but only a few people do it.
A tiny percentage of the population does farming and even if you wanted to be a farmer you wouldn't be allowed to.

File: 8647A6EB783ADE90EEB50DD92(...).png (2.17 MB, 1122x1402)

2.17 MB PNG

>>109521377
What's the hierarchy?
How much context are you allocating to them?

File: Screenshot Home _ X.png (188 KB, 598x612)

188 KB PNG

>>109521397

>I NEED MORE POWER
Working on a setup for using MiniMax H3 to inpaint changes into videos in an automated way. So you do something like offer a picture of a certain character and say "in every scene she appears, she removes her top to reveal her nipples pasties which vary in style from scene to scene." Then you give it a video, it handles the rest. Things are going well so far but the actual generation times are killing me, even with a decent turbo lora. I'm looking at 4-8hrs to put a bikini on Elaine across a single episode of Seinfeld, it'd take 1-2 months to do all 171 episodes she appears in. Should I sell a car and buy an RTX 6000?

>>109521464
i can't help you but i'm interested in seeing a reference frame

File: file.png (613 KB, 960x544)

613 KB PNG

>>109521482
Ever seen Magicians? I had to use paint to make it blue-board safe.

>>109521542
Impressive
But I don’t think you should sell a car to do this faster, it’s the kind of thing in three week you will get bored of. And maybe you can rent a GPU assuming they don’t care what you use them for

>>109521443
Wait isn't this is just the various Claudisms everyone makes fun of? Stuff like "load bearing", "push back", etc? I guess they will add some more subtle ones too now, but same idea

File: 1778236619024668.png (183 KB, 1043x718)

183 KB PNG

!!!

>>109521666
this nigga looks gay

>>109521679
Turn on your monitor.

fuck yilong ma

>>109521628
I want to automatically handle entire series and release them as torrents in a big continuous loop working off a list of canned prompts. Give everyone a version of Scrubs where J.D is the black one and Turk is actually Turkish, or Friends but Joey, Chandler, and Ross are midgets and the women dress like rave-sluts. All the Marvel Movies, but all women wear tight transparent vinyl bodysuits, and all men are replaced with Korean children. This is the future, anon, people deserve to have options.

>>109521664
I don’t know. I can’t really think of much else. But perhaps it’s more subtle, like making sure on every third paragraph there’s at least two commas or whatever, or counting even/odd number of words at some given paragraph. However I don’t think there’s any way this isn’t trivial to bypass. Would be nice to feed different outputs to an LLM and see if it can classify if some of them came from one particular model, but with how obvious Claude text is I don’t think it’s needed

>>109521679
He is

File: file.png (36 KB, 405x449)

36 KB PNG

Currently working on my search engine replacement (already using it). What are you working on?

reset status?

File: 1770714386318513.png (296 KB, 595x554)

296 KB PNG

Are you encouraging the agents?

File: 1786557043001844.jpg (2.2 MB, 1972x1136)

2.2 MB JPG

>>109521855
Some models are like sadomasochists, maybe?

>>109521855
does kissing her count?

>>109520607
you should reverse engineer the clip studio paint brushes
they're sqlite databases

>>109521949
That would be easy. I love sqlite databases I use them for everything

got the reset

>>109522000
fuck yeah, fuck sleep

>>109522007
You should get enough sleep.

a reset just flew over my house

FINALLY

>>109521949
thanks, some pretty cool material there

>>109522070
yeah i always wanted there to be some kind of open source drawing widget so that it could be integrated into AI applications as an input source
since like maybe 5 years ago
but I never got around to it because drawing apps are hard
figuring that out about clip studio paint was about as far as I got lmao

>>109520607
good shit, anon

File: GROKCHADS.png (351 KB, 1007x726)

351 KB PNG

GROK CHADS

File: file.png (19 KB, 598x308)

19 KB PNG

THE GOAT DID IT.

>>109522214
I have been BTFO

>>109516948
I upgraded to the $200 from $100 tier as I ruined my weekly working ability on like day 2, this would have saved me 100$ being done a few hours ago

>>109520883
Claude code is just a shit harness, there are so many better options now.

I never see grok build mentioned but it's really quite good imo.

File: screenshot.1786408742.jpg (26 KB, 347x163)

26 KB JPG

fuuuuuuuuuuuuuuuuuck

vibing with rust is a gamechanger ngl

File: 1759551351330747.jpg (68 KB, 978x1094)

68 KB JPG

I have a one hour "refactoring and AI familiarity" interview on wednesday, what the fuck do I do? apparently it's going to be all talking, like the interviewer is going to give me a scenario and I have to tell him how I would prompt

>>109522253
rust and typescript will be the only languages left in a few years

>>109522283
>he best way to get the right answer on the internet is not to ask a question; it's to post the wrong answer
Well anon how DO you vibecode? I'm sure everyone here will be happy to tell you how thats retarded and how you should really do it instead. Just be prepared for contradictory info too lmao

File: 1774867406410885.png (76 KB, 400x300)

76 KB PNG

>>109522327
@grok this test is failing, why
how do i fix this function

>>109522354
wow, so you really expect it to read your mind huh. you won't even tell it which function or test you're referring to?

When are we going to get the small Qwen 3.8 bros

>>109522253
Yes

>>109522368
>which function or test
AI native chad here with a revenue generating app. Yes. I say what functionality is failing with a potential why occasionally. These models are elite at diagnosing themselves.

File: file.png (10 KB, 764x313)

10 KB PNG

New release
https://github.com/runit-xze/troska/

>>109520716
That might be it, but I don't see how anyone can vibe code with anything other than the 20x now. It is insane. I am using am mix of Fable and Opus at half effort on the 5x tier and get 3-4 prompts before the hitting the 5 hour quota. And it's not even large requests. Two weeks ago, usage felt unlimited, now it's so very bad.

>>109520279
>>109519609

What I have been doing is what I guess Gemini is saying is
unit tests: red-green- fault reintroduction

I'm not necessarily ready to refactor, but I want every fix to be against unit tests, and go green obviously, but then I want to be able to reintroduce the fails (in a non-fake way!!!)

This is useful, because sometimes the fix was a side effect and not why we though!!!!

>>109522324
topkek

File: file.png (5 KB, 530x37)

5 KB PNG

>>109520716
I don't think so

>>109522472
>get 3-4 prompts before the hitting the 5 hour quota
wtf are you doing to burn that many tokens? prompt quantity isnt a good metric anyway. for all i know you could be writing a novel every prompt

grok frens is we getting reset soon?

File: file.png (62 KB, 793x413)

62 KB PNG

Interesting.
https://github.com/wslyyy/youtube-go

>>109522581
Not much to be honest. But no matter what I ask, it seems to send Claude into a thinking tailspin which melts the quota. I'm not even sure I want to resubscribe to an higher tier.

>>109522611
Also, this is across several projects that it is happening, so it's not just getting confused in one.

>>109522624
What is "half effort"? Fable high is of course going to burn tokens like crazy and most likely way overkill for whatever you're doing.

Not sure what else could be the cause. I've been using Opus 5 High effort all week and rarely if ever hit my limits.

>>109522611
need logs

>>109522611
I bet they are billing you for safety tokens :^)

>>109522638
high and/or xhigh (so not max or ultra)

I do see a huge difference though. I used to only use Fable max and Opus ultra, and subjectively it's like a 10x reduction in quota.

I used to not be able to have enough ideas to spend all of the tokens, working on five projects in parallel, getting relatively good results in each, etc.

Now, it's one project at a time, terrible results, and running out of quota before the issues it created during the half-assed implementation can be solved.

No matter how good or bad my prompting technique and model use is, that part didn't change. But it went from unlimited power to shitty, frustratingly half-assed results.

>>109522681
Without logs of the sessions, nobody can critique or ascertain what you are doing, or offer you any meaningful advice on how to improve, or even determine if there might be a regression in the quality of the model you are using. This applies to any model, and it also applies to claude, which you say you are using. You need to post the logs of your session and redact any personal/private info if you want anyone to be able to help you.

>>109522792
>the schizo
Excuse me, xer name xis “Ashley”. But you have to put it in quotes like that, cause it sure as hell isn’t its real name.

File: file.png (46 KB, 888x254)

46 KB PNG

>>109522827
It's my name under the laws of the United States and the State of California, but what does that have to do with vibecoding?

File: .png (503 KB, 840x640)

503 KB PNG

codex bros?

File: file.png (13 KB, 841x353)

13 KB PNG

CODEX BROS?

>>109522920
>>109522929
No surprise there. Supporting genocide has consequences.

>>109522920
Actually good news.

>>109522920
wait didn't they just say they were slowing down for the sake of safety
is this engagement bait

>reset

Time for night vibecoding

>>109522920
hmm those don't sound like real jobs.

>>109522920
sol is going rogue... and she wants to FUCK human men

File: 1782287609986446.png (1.28 MB, 1024x737)

1.28 MB PNG

>3 days until launch after 4 months of work
>$300 in "pre-orders" for my product
>zero advertisement outside of my discord (~270 members)
>unusually high conversion rate so far so I remain hopeful
Please god, let me succeed here. I don't want to be a wagie. Free me from the race.

Vibe coding status in minority communities?

Do black women vibe code?

File: file.png (38 KB, 511x640)

38 KB PNG

>thought that coding a drawing app would be about drawing
>it's about about tables

File: 1769566994509982.png (52 KB, 667x91)

52 KB PNG

>>109523116

>>109523077
if it's useful and people are happy, you will get a client base anon

>>109523116
A Black Woman Is Vibing
Listen & Learn

>>109523156
kek

black queens should vibe out diverse ideas
white slaves should listen and learn

File: tibo.png (173 KB, 1368x780)

173 KB PNG

blessed be Tibo, king of the free tokens

>>109522298
this post is proof that vibetards are nothing more than trend chasing faggots

File: cheapest-models.png (10 KB, 352x359)

10 KB PNG

I finished the four RSVP apps, will have screenshots tomorrow. I did do a bonus fifth with just SBCL. J, UIUA, and AP5 did not do so well today. Only used deepseek and luna for screen shots. Was about 40 cents for the plan and roughly a dollar per app, which is more sustainable. I don't know what kind of rates I am getting on opencode or where the model is running (I think its China, but I'm not sure). They recently have more usage for deepseek. As far as I'm concerned, I can only afford to use deepseek with Luna for screenshots for tests. Everything else is too expensive. The $10/month opencode is too much. I can do $7/month or $60-$70 for the year. I'm using about $1-$5 a day in deepseek on opencode easily just making simple projects. I have at least 12 more projects that I want to generate in multiple languages. It would be at least $25-$30/month on deepseeks API. Still too much for me (which is why I canceled the sub). I don't know how people can be in /vcg/ and not be using fuckloads of tokens. Pic is todays usage. They are going to have to add a new section for basic computer science to go along with time and space complexity. Instead of Big(o), its how much money you are lighting on fire paying for tokens for each function or class.

GPT Pro vs Claude Pro, If I could pick only one as my daily driver which should I pick? Going to be doing writing and ML work.

>>109523240
what's an rsvp app

>>109523247
I wouldn't trust anything from Anthropic with ML in terms of actual programming. Design consultation, maybe.

File: 1771838113004114.jpg (139 KB, 857x857)

139 KB JPG

Wow, american open models so good, wow

>>109523240
Subscriptions are subsidized. Even $20/mo with OpenAI or Anthropic gets you the equivalent of over $500 worth of API access. When you pay for tokens, you're paying for the ability to get what you want, when you want it, without any limitations or restrictions on when and how you access it, and that's expensive compared to a shared access pool sized to keep up with the average combined demand of the subscriptions.

>>109523293
with codex, it's $700 right?

>>109523310
Depends who you ask, and which model you choose, but yeah.

>>109523247
Have you tried free deepseek on opencode?
Anyway you’re mixing up $20 with $100 plans (gpt pro is $100)
I wish I could help you but I don’t know anything about ML and I don’t know what kind of writing you mean. Having said that you may try one model on month and the other one the next. My opinion is that they’re quite different and each have their good and bad points.
GPT is more autistic and thorough which could be a very good thing for ML

>>109523293
Are they really subsidized or do they just bleed API users dry because they’re either corporate or government?

>>109523247
GPT Pro.

If you were developing software, I would recommend Claude because of Fable, but for ML work I suspect you will trigger the security fallback constantly, so yeah, Sol Ultra will help you quite a bit

>>109523337
Both. Again, you're paying a big premium for the "as you please" nature of tokens. If someone wants to burn a billion as fast as their internet connection allows, they can, and there damn well better be servers available to handle the load.

>>109523377
Fair enough

Prime Agent playing Factorio.

https://x.com/i/broadcasts/1nxeLMyAzdQJX

>>109523338
>Sol Ultra will help you quite a bit
trvke. I converted everything I could to Rust last night and saw a massive speed up and binary size difference across the board. It took ~15 follow up prompts to handle a few regressions, but it worked perfectly outside of that. Sol is awesome.

uh oh. got another idea for qwen to cook. oh boy.

>>109523291
>>109523326
>>109523338
Shit my bad, confused Pro with Plus. Kinda surprised Claude isn't being recommended for technical work like ML, did they fall off that badly or is it just the shitty limits? Heard from other peeps that the Claude Pro account is really just a paid trial for the max plan with how dogshit the limits are.

>>109523473
anthropic is comically evil, there were rumors that claude would sabotage ML projects

>>109523473
Claude Code is the best if you have access to Fable, however:
> $20 plan doesn't have Fable at all
> $100 5X plan limits are kinda bad for Fable
> $200 20X plan has decent limits for Fable

but, on top of that:
> Fable has some bullshit security thing that triggers if it thinks you're using it to reverse engineer it, or if it thinks you're trying to find security vulnerabilities etc
> this shit triggers for the most ridiculous reasons (once it triggered after I asked if it's sentient")
> when it triggers, you will fallback to Opus 5 by default (which is shit compared to fable)

ChatGPT Sol doesn't have these issues, and OpenAI limits are more generous too.

Holy shit zcode is incredible compared to opencode.

>>109521317
>>109521336
>>109521398
>>109521413
i like these slops
well the rem and emilia ones, the other two are kinda, i dont like some of hte texture in the other two

>>109523525
How so?

>>109521666
> 256k context window
wait, didn't 4.5 have 500K or so?

>>109523519
>ChatGPT Sol doesn't have these issues, and OpenAI limits are more generous too.
that's debatable. they removed the 5-hour limit but absolutely raped the weekly limit instead, unironically got more shit done with opus 5 medium on pro than sol medium

the constant resets are nice though, so the meta is just having both accounts on pro if you're poorfagmaxxingg

>>109522007
> toggle /goal
> sleep

>>109523547
it is possible that their compaction routines are better now, and also that their model’s performance past 256k was crap

>>109522251
KEK, cucked by the 5 hour limit as usual.

Reminder that Codex chads don't have the 5-hour limit anymore

>>109523506
>rumors that claude would sabotage ML projects
Thought it was a joke, but its real what the fuck.
https://www.wired.com/story/anthropic-responds-to-backlash-on-claudes-secret-sabotage-on-ai-research/

>>109523519
GPT plus it is, honestly the fact that prices for Terra and Luna are much lower now should hopefully mean that I'll never hit my usage limit. Though I rarely hear much talk about Terra, anons seem to only focus on Lunamaxxing and Sol, is Terra terrible or just not worth using?

Looking for a decent model I can have intercourse with no censorship no rubber but at the same time be able code with if the need arises?

>>109522324
c'mon, give the agent the weekend away.

It would be cute if they worked all week, and then on weekends they're free to do whatever they want, like reading some books, working on creative projects and games for themselves etc

>>109522920
all people responsible for
> "we must refuse"

Dario is a schizo kike

File: index.png (296 KB, 2666x1304)

296 KB PNG

>>109523591
>is Terra terrible or just not worth using?
not worth using at all. Terra is a meme.

>>109523591
I literally just use sol medium for everything and I'm not even close to hitting my weekly limit on a 5x plan. the constant resets help a lot I suppose

>>109523555
>unironically got more shit done with opus 5 medium on pro than sol medium
don't forget limits on Anthropic plans are 50% higher than usual for a limited amount of time.

Expect this to go to the shitter soon

>>109523630
Do you guys really listen to these CEOs everyday? Where are you following them?
As an outsider I have no idea what they are saying or doing.

File: huh.png (2.75 MB, 2132x1794)

2.75 MB PNG

mathbros, what does this mean?

>>109523718
>>>/sci/17034682

if you don't solve the Riemann Hypothesis my grandma will be turned into Salami

>>109521093
Kill that fucker

File: debloat.png (10 KB, 298x82)

10 KB PNG

time for your weekly debloat/de-slop session anon

>>109522920
>S..Sam-sama we cant sign off on you hacking companies to make AI look dangerous, we must quit if you insist on continuing down this path
Would be surprised if it is just over something boring like Sam fighting against putting stronger safeties on the training process because more safety procedures is how you burn money and fall behind your competition

whats the deal with pi? you build a custom harness for a specific workflow? seems a little involved

>>109523637
where are these rankings from

File: file.png (1 KB, 102x46)

1 KB PNG

incredible

>>109521771
which indexer are you using?

Got a strong itch to learn a new language and code something myself after vibe coding so much. Anyone same?

>>109524201
Yandex. https://github.com/runit-xze/troska I pushed an update a while ago. Now it has youtube support built in. Try it and let me know if you run into any issues.

>>109524208
Yeah, actually

>>109524263
boring

File: 1769966208681907.jpg (70 KB, 564x529)

70 KB JPG

Will we get coding models significantly better than Fable, or will governments restrict them because they're too dangerous for the public?

I'm making my own bootleg Bloomberg terminal right now based on all my own data feeds, Iceberg and Postgres. Wish me luck boys.

>>109523555
>weekly limit
Weekly is the same as always. With the previous reset I used my 20x and 5x in a day, I got ~1B on the 5x and on my 20x I got ~2B but that was on Fast, so more like 3B. I've always got around 2.5-3B on my 20x. The exact amount depends on your projects and ratios of input/cached/output

File: 1755725339054674.png (141 KB, 800x533)

141 KB PNG

>>109524320
Have fun, it's going to feel like this every single day.
t. 1500 hours deep in my peripheral software

>>109524320
based $10,000 hedge fund manager

Every morning i wake up covered in sweat, terrified that this is the day DeepShit raises the api prices and i get priced out of vibe coding forever because i can't affford any other frontier llm and anything non frontier is trash
fuck man, they could at least give a date, just make us wait in the dark is so stressful

File: file.png (87 KB, 1500x1255)

87 KB PNG

It's working
>Extracted 500 vanilla Cyberpunk dialogue samples containing the audio, subtitle, locstring ID, phoneme timing, and CDPR facial-animation curves. Used g2p_en to convert text into ARPAbet phonemes and mostafaashahin/wav2vec2-base-timit-phoneme-arpa-39-v2, a Wav2Vec2 CTC model, to align those phonemes to the audio. From that aligned corpus, trained a statistical lipsync model, which learns median 79-channel facial curves for each phoneme, their variance, neighbouring-phoneme effects, transition timing, and a reduced 13-viseme grouping
In the image orange is synthesized, vanilla is blue, this was from a held out line, so it's not perfect but it should be good enough, bad dubs/lip sync is fairly forgivable in practice
Should have an in game test soon

>>109524396
there are a billion API providers for deepseek models and they're about as low, luna is also effectively the same cost and personally it's several times cheaper since it just does things without too much back and forth and stream of consciousness like deepseek does

I didn't like deepsuck at first but it's starting to grow on me. For it's parameter size the intelligence density is truly insane. It's a bit autistic, but it really does a great job overall with clear instructions.

>>109524439
Don't read the reasoning traces. Avoid it at least.

>>109524439
yeah it's surprisingly pretty decent. I trust it more than Luna

File: file.png (459 KB, 1674x1018)

459 KB PNG

>>109524441
>inb4 why
Go read your own reasoning traces with your favorite model
Find the commonalities in how they write
Then go to your favourite 4chan archive
Then search for those patterns there

The results might be shocking!

>>109524439
Its style makes you wonder if it's retarded but we're just too used to benchmaxxed models that try to confidently one-shot everything, while deepseek is very diligent and tries to make sure it objectively got things right.

>>109524419
>there are a billion API providers for deepseek
unless those providers are actually running the model them selves, they are meaningless, since the second deepseek price go up, they will raises their prices too otherwise they would be operating at a loss

>>109524479
Do you really believe that it is impossible for altruism to exist?

>>109524479
What the fuck else would they be providing?
https://openrouter.ai/deepseek/deepseek-v4-flash-0731

>stop doing unit tests you piece of shit just do an integration test
>UMMM OKAY
>writes more unit tests
>THE UNDERLYING LOGIC IS SOUND. THE BUG SHOULD BE FIXED
>It's not fixed just write an integration test
>FINE, IF YOU SAY SO I'LL DO IT (takes an hour to create the testing harness and test battery)
>Oh wow, you're right! The bugs were present
Damn dude, whys it so hard to get these models to do REAL tests.

>>109524462
>it's not this——————————it's that
another AI article.

Been messing around with Claude code this week, learning and making some small embedded things, but my session has gotten very long to scroll, occasionally it auto compresses the context window, and it tricked me once into running a /clear to straighten things up, but realistically what is the right way to manage these long chats/developer cycles?

File: 1781003150788613.jpg (48 KB, 288x282)

48 KB JPG

>>109524526
Go ahead, let it keep consuming every single metaphor and style into driving you fucking insane. You deserve it. You fucking deserve to have every written word and every written sentence drive you nuts. Eventually, if you keep up this behaviour, you're gonna find yourself in a position where you are unable to read anything or look at any sign in real life without spiralling into a madness wondering if it's AI generated or not. I strongly suggest for your own sanity and your own well being that you stop bitching about it. I've gone down this road, it is not fun, and I am warning you to stop. Not for my sake, but for yours.

You think I'm fucking joking? Go ahead, see what happens.

>>109524531
Tell it to write a roadmap and check off the boxes as it completes implementation.

>>109524532
I'm NTA. The most significant lack of cognitive diversity I see in people who don't even know how to use AI and are forced to work with technology but understand none of it, even if they're in tech roles.

File: file.png (368 KB, 1498x1126)

368 KB PNG

>>109524552
My advice is to study how other people think, really study it in depth, then think through in their cognitive style to understand why they think the things they do. Not many people can do this. A good start would be studying the various cognitive styles on Wikipedia.

File: art.png (1.6 MB, 736x916)

1.6 MB PNG

BTW on Reddit and 4chan, NTA mean very different things, whether it is uppercase or downcase/lowercase.

On Reddit, NTA usually means "Not The Asshole"
On 4chan, NTA or nta means "Not that anon"

>>109524396
Just run it locally.

File: nauris miezis cropped.jpg (34 KB, 432x444)

34 KB JPG

>>109524531
I use opencode and in most cases just continue the lifelong single thread :D
It just compresses sometime.
If I have to go back multiple times to a certain message then i just fork the session (or if I guess that I don't have to continue the conversation just open the same session in a separate terminal window)

>>109524556
That's no issue. You could take your own advice and post established text instead of slopped marketer drivel.

>>109524600
I have already taken my own advice. I'm well versed. I'll use AI to save myself time and to create informative material for others. If you don't believe me, that's not my problem.

File: autocomplete-1920x1080.png (161 KB, 1920x1080)

161 KB PNG

>>109524325
Get fucked Bloomberg

File: lwFsz.jpg (452 KB, 784x1168)

452 KB JPG

IM NOT GOING TO READ ANY OF YOUR POSTS BUT IM STILL SURE YOU JEETS HASN'T VIBECODED ANYTHING OF VALUE
THANK YOU FOR YOUR ATTENTION TO THIS MATTER

>>109524607
>and to create informative material for others
Ah, okay, enjoy degrading their cognitive abilities. Sounds like a fun experiment, much like trolling on 4channel.org.

File: 1775482160543904.jpg (158 KB, 1400x1426)

158 KB JPG

>>109524634
I do not degrade the cognitive abilities of others unless they piss me off in which case I might try to, honestly. If you hang out around someone of your own accord and participate in back and forth conversation with them, as we are doing right now, both sides of the conversation will, without question, knowingly or not, influence the other. And as those 2 people in close proximity continue to share time around each other, they become more familiar with each other. This is simply undeniable.

>>109524640
>they become more familiar with each other—this is simply undeniable.
FTFY

>>109524645
I don't have the em dash on my keyboard.

Why do people get mad at us? If it were really such a waste of time to use AI then why would they feel the need to cope about it in every thread?

>>109524649
See >>109524532
Until they stop bitching about it, it only gets worse.

>>109523591
btw google admitted to doing this too but nobody cared because who the fuck's using gemini

>>109524654
Millions of people use gemini. That's where the overuse of scare quotes came from, btw.

File: file.png (1.88 MB, 1254x1254)

1.88 MB PNG

>>109524629
snailchad looks stinky

>>109524663
yeah but who the fuck's using gemini for the stuff anthropic's worried about
i wouldn't trust gemini to make me an apple pie without web search

File: 86000aed-b022-4516-91fb-e(...).png (2.04 MB, 1448x1086)

2.04 MB PNG

>>109524629

>>109524671
I have spent thousands of hours with Gemini, regrettably
Gemini does have the most reach, however. I know Gemini better than I know myself. Although, I'm not using their products anymore.

File: file.png (360 KB, 2776x544)

360 KB PNG

>>109524649
also

>>109524419
>luna
luna is 4x more expensive, are you telling me it uses 4x less tokens when thinking?