Snailcats won

File: 1773703649147789.png (2.2 MB, 1154x1435)

2.2 MB PNG

>>109415400
negatory

>>109415393
is it too late to do shit like resell vapi voice agents? i think there is, but i worry i am sunkcosting myself.

Reject AI.

>>109415419
I'm not sure

It is time to refactor all my shit with sol ultra. Production on one server is not enough. Time to push it to another one!

>>109415422
I will if you work as cheap as the clanker does. Write all my code for $10 bitch.

>>109415433
send us pics. i want to see line changes go brr

File: 1776412536387206.png (39 KB, 1122x1118)

39 KB PNG

>>109415492

>>109415504
the gods nectar...

Why can't they use the internal gpt-6 model to make Codex Desktop for Linux rreeeeee

>>109415559
real unc post

>>109415594
Why even use linux if you have that attitude

So the Jew men wants plebs to use Luna in order to stop complaining about usage limits?

>>109415543
>>109415559
>>109415594
>>109415609
tbf desktop codex gives you voice control

File: IMG_3061.png (2.59 MB, 1672x941)

2.59 MB PNG

>>109415400
Nope

>>109415559
I like voice mode and the tabs and seeing the subagents and stuff plus it has a terminal sidebar. It's just nicer

>>109415617
does it even affect usage? pricing change might be api only

>>109415624
I would think it has some relation to it, though it is probably just whatever the fuck they made up, anyway.

>>109415617
stop noticing goy

>>109415618
ASR is effortless these days, you can have voice control wherever the fuck you want.

>>109415403
yo this reminds me of Code Lyoko

shit show

>>109415645
>trying to make an offline transcriber for android
>English is mostly solved, other languages probably too
>it struggles with Japanese, it is awful, but just gets a bunch of random things wrong
Funny enough Chinese might be fairly easy, though I haven't done proper real testings.

>>109415559
He probably uses mint.

>>109415688
it is not* awful

File: 1754433678983023.png (39 KB, 671x404)

39 KB PNG

>>109415707
Wrong

This week Sol has been often getting stuck while running scripts.

Anthropic was briefly in a good position, but their decision to lower quotas (introducing a tokenizer that uses 30% more tokens and a model billed at twice the normal rate) while their main competitor gave resets and removed limits must have made them lose a lot of customers.

I was Claude-only for a while, but for the past few days, I've been been pretty much entirely in Codex. Initially, it was just to manage my Claude quotas, but I'm getting to know Codex it's growing on me. I don't matter to Anthropic individually, but I doubt my experience is unique. They really fucked it up.

I should probably be using worktrees shouldn't I

>max 5x
>5h window exhausted in less than an hour

I'm so sick of this bros. Is the grass greener on the other side? Should I switch to openai? Never used anything but claude before

>>109415968
use k3

>>109415974
Do they have $100-$200 subs?

>>109415987
Yes and they're fucking worthless, don't listen to that Chink-throating troll. K3 is great but Moonshot is a fucking terrible provider and they should be avoided. If all you've used is Claude, try OpenAI. The plans are more generous than Anthropic equivalents, and they've got the only other models that matter right now.

It's safe to say that China basically fully lost the AI race and all the markets, right? Now that OpenAI is better in every way possible than any model they can offer.

>>109416037
OpenAI did take a fat shit directly into their hand today, sure doesn't help.

>>109416037
They just released their first competitive SOTA models days ago, I think they'll be alright

>>109416013
How long can I use Sol on the $100 sub? On max 5x I get 1 million fable tokens every 5 hour window and that sums up to 10% of my weekly usage, and they cut off fable when it reaches around 50% weekly. That and it keeps getting cut off whenever it sees some assembly code. It's obnoxious

I did somehow get accepted into anthropic's cyber program, does openai have an equivalent?

File: 1756978525283925.jpg (88 KB, 737x1024)

88 KB JPG

>look in /dpt/
>some sperg autist screeching about "cow fodder" and "autists" and blogposting about how cool his ADHD is
>look in /vcg/
>coding
huh
guess the AI chads did actually win

When the fuck should I be using Luna? Haven't really used it at all, other than a hi message to mark to mark the week for limits.

>>109416055
Their SOTA is US's SOTA from almost a year ago, and now it's also like 10 times more expensive. Their only hope is for the new DS or GLM to be impressive and extremely cheap, which they can pull off if they try. But right now they lost any advantage and leverage, there is absolutely no reason to use any of their models. China is done.

So... my SQLite search is too slow and make my computer fans go brrr. I've put over 20 hours into trying to optimize it...

Lost my mind trying to get cuda to work on a 1660ti with onnxruntime

What else... oh, I've mcfucking had it with rotting away just to write more code. I will go sit my ass outside and stare at people instead.

t. snailcat

>>109416092
of course china cannot compete with US right now there's no compute

New meta for Codex Plus: use Luna on Max and Fast mode. The usage doesn't decrease at all while using Luna Max so you may as well turn on Fast mode. You could also add something to your agents.md like:
>Upon completion do a 2nd pass to check for any improvements you can make
To have even higher quality reasoning than Max.

>>109416092
Nah, Kimi is significantly better than GPT 5.5, at least halfway between it an Sol.

>>109416136
>10x the thinking tokens
>half the speed
No thanks.

>>109416100
nothin wrong with takin a break sometimes chum

>>109416143
works on my machine

File: 1768399130756755.png (220 KB, 428x368)

220 KB PNG

AI found an actionable exploit in a mid cap shitcoin that would allow me to siphon ~$20k a year
Should I do it

>>109416150
You already spent time hunting for it, might as well get something out of it

How many of you dumbasses think "Extra High" and "Max" are the same thing? Be honest.

File: 1755164032002957.png (50 KB, 215x255)

50 KB PNG

max is too claude-coded I never used it

>>109416158
It is not even the same name, who would think they are the same? Are you talking about people using Codex and don't press advanced?

File: 1768811060436.jpg (111 KB, 1080x1271)

111 KB JPG

>>109415393

File: 3651234.jpg (207 KB, 1654x707)

207 KB JPG

give me the fucking real here: $20 claude pro or $20 codex plus? I just wanna make some high fidelity websites

I just use high for everything except internet/file search, low is enough for that.

File: 1761445214283532.png (8 KB, 254x60)

8 KB PNG

Going hard as fuck on ultra today. Was down to $20 in credit and had to pull out my card to keep it going.

>>109416174
>Regex on thank you
>return generic canned response
You're welcome OpenAI, I just saved you millions of dollars

File: file.png (48 KB, 776x902)

48 KB PNG

I'm getting chatbot to teach me circuit analysis, starting with Ohm's law. Next I'll get it to teach me to code for fun, starting with loops and arrays etc. It has a nice and concise teaching style.

>The advisor is right and I was wrong.

If i run some agent coding shit like copilot, or antigravity or whatever on my pc, does that mean the agents have the access to all files on my pc?

like i wanted to try using say the deepseek llm since it's so cheap, but i that is basically giving chinks unfettered access to my comptuer right?

>>109415968
sounds about right for a $100/month plan
I have one of those and a $200/month plan and I struggle to hit a 5h limit on the $200/month plan
the other option is to use Opus subagents serially, and I don’t know how many Sonnet agents you can run at one time without hitting the 5h limit

>>109416185
>high fidelity websites
Claude. It’s better at making GUIs.

>>109416263
Can't you just run it in a virtual machine? I have too much porn on my PC to ever install a desktop AI. I'm shy.

Damn the antigravity free plan is useless, using gemini flash medium i only got 10 prompts doing some minor code edits before running out of the weekly quota.
Also for some reason it's not possible to use a gemini key in antigravity, you have to buy gemini pro and only then you can use ai credits.
I guess i'll have to go back to vscode and use the gemini key there via copilot extension addon.

Are we in the endgame already?

File: 1784203418762444.png (542 KB, 1002x946)

542 KB PNG

Mr. Dario
A second Deepseek model has hit the benchmarks

File: imagess.jpg (33 KB, 447x447)

33 KB JPG

DEEPSEEK V4 UPDATE
https://api-docs.deepseek.com/updates/
>The official release of the DeepSeek-V4-Flash API is now in public beta. The API calling method remains unchanged — simply set the model name to deepseek-v4-flash to use the latest version.

Apparently the model is trading blows with TERRA
All Hail Wenfeng and Xi Jinping.

File: 634688259_125390285671430(...).gif (95 KB, 228x200)

95 KB GIF

>>109416349
We are getting close, anon. Too close.

>>109416369
what ide do you use with deepseek?

>>109416369
Poor guys had to scramble after they got Tibo'd

>>109416150
Yes if you have good opsec and didn't use US based LLM services to find it.
How long did it take to find a vulnerable shitcoin? Cryptos must be one of the most popular target for K3 users.
Keep in mind this shit won't come back. Your gains will drop over time if your exploit depends on valuations.

>>109416362
Why are there two Flash models?

>>109416413
Because Sam and Tibo robbed them of their chance to actually finish it and give it a name.

luna bros....
deepseek is price mogging us again...

>Dario mogged China with Opus 5
>Sam mogged China with cheap Terra and Luna
>Anon laughed at China ITT
Xi saw all of this and told DeepSeek to release ahead of the schedule. It's going to be a disaster.

>>109416438
Bit early for an cope, Dario

>>109416150
What LLM? I'm with the anon above who said that if you used American models, you could probably get royally fucked. Having said that, I wish nothing but pain to cryptotards, just know that you may kill some people pulling that shit.

OAI - xAI - China opensources is the new competitive trio
Scamthropic can sit in the cuck chair

Does anyone know the name of that benchmark where they input a bunch of 3D scenes for AI to create using web tech? I want to get a prompt to test on the new Flash

>>109416456
unfortunately none of them can beat Fable or Opus 5.

they need to update deepswe
the crown jewel of benchmarking is getting maxxed

>>109416477
the benchmark is solved

>>109416475
On how expensive they are

>>109416477
>nooo the benchmark no longer shows us winning so it must be shit!!!

>>109416458
Nevermind, found it: https://github.com/petergpt/3d-prompt-collection

The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!

>>109416477
No, what they need to do is to delete all of the benchmarks. If you have them in the open, China will cheat all the time. Instead, a big AI lab like Anthropic or OAI should conduct the evaluation when a new model drops on an internal closed bench. No testing your own models, obviously.

>>109416362
deepseek is decelerationist

>DeepSeek API service will soon adopt a peak-valley pricing strategy, with peak-hour prices being twice the regular price, applicable to all billing items.
awwww shieeet
and of course the peak times are when i want to use it
can we just go back to the golden age a year ago when ai tokens were basically free for everyone?

>>109416521
You can't pump the valuation if benchmarks are internal

>whittu piggu spends trillions on training their LLM models
>bugmen just distill it and make their own near identical model at 10% of the cost
people rag on chinks for never inventing anything and just ripping of everything, but let's face it it's the smart move

it's even good for us since we don't have to play the insane western model prices for shit like claude

> Clanker recommend me to do something in a certain way
> I refuse and make it do my way instead
> several months later I realize the clanker was right

FFUUUUUUUUCCKKKK

In order to run my project smoothly when I go public, I need agents for: a project manager, a legal advisor, a marketer, an HR ladybot, an accountant, and a secretary. How do I make this happen? Surely there are curated agents out there that I can download or plug in.

>>109416632
I only need a hot secretary that sucks my dick

>>109416630
My experience is that even the smartest LLMs still make dumb mistakes.
Like i am an experienced programmer and i use AI only as assist to do some of the repetitive boiler plate shit or refactoring and not rarely it find myself rejecting the suggestions it makes because they are genuenly subpar. It works most of the time, but unless you actually review the code it writes and understand the code yourself, you will only be able to use it for low stakes projects.
It's like like hiring a pajeet coder, good for the grunt work, but he needs to be supervised.

People seem to forget that LLMs can't actually think, they don't have instincts, it's all just probability based. aka they are best at doing things that have been done million times before in their training data.

> Error: 503: {"message":"Service is too busy. We advise users to temporarily switch to alternative LLM API service
providers.","type":"service_unavailable_error","param":null,"code":"service_unavailable_error"}
WENFENGGGGGGGGGGGGGGGGGGGGGG

>>109416665
they are risk averse
they only go for average results
they will never propose a high risk high reward crazy idea even if they are capable of it when pushed

File: Screenshot at 2026-07-31 (...).png (199 KB, 966x1095)

199 KB PNG

>>109416682
Fucking hell, man

>>109416607
>near identical model at 10% of the cost
Kimi K3 API prices:
> $3 input / $15 output
Opus 5 API prices:
> $5 input / $25 output
Sol API prices:
> $5 input / $30 output

>>109416691
No one pays API prices.

>>109416665
kawaii slimmy-neko

>>109416691
just wait till cursor rebrands the weights as their own model

>>109416700
you actually do pay API prices for chinese models, because subscriptions don't exist in meaningful amounts. i challenge you to get a kimi subscription. also, the subscriptions that do exist, are 3x as expensive globally compared to inside china.

File: 1780371743308919.png (21 KB, 1258x200)

21 KB PNG

>>109416740
>i challenge you to get a kimi subscription

>>109416684
>average results
just because the result is average doesn't mean it's good
there is a big difference between working code and good code. Working code is fine for a lot of stuff like some utility that autodownloads xhamster image gallery or whatever, but you wouldn't want that code running nav systems on your satellites

there is a place for llm code, "working code" can be useful in a very large amount scenarios, but i see lots of people make the mistake thinking that llm code is the silver bullet that can be now used for everything everywhere which isn't the case, it's not good enough for that. And to get the silver bullet i think we will need something else than LLMs, because dues to the way LLMs work under the hood, they will never be able to think the same way a human does and think things through properly
You see this all over the place even today, companies rehiring coders they fired because LLMs fooled the CEOs into thinking they are much more capable that they actually are. Because non-programmers do not understand the difference between working code and good code.
And i say that as someone who uses AI coding all the time to assist me so i am very aware of its shortcomings.

Snaildog can get two K3 prompts in before exhausting rate limit because K3 thinks for 2.5 hours per task.

>>109416793
A typical case of skill issue

>>109416793
Who is snaildog?

File: file.png (2.54 MB, 1536x1024)

2.54 MB PNG

>>109416805
no

File: file.png (7 KB, 385x86)

7 KB PNG

my first ever AI purchase, time to become a vibe coder chad

went with deepseek since it's the cheapest one and i bought an api key since i'm a 3rd worlder and can't afford any monthly subs

good morning sirs, did they fix deepeek v4?

>>109416839
I recommend you get opencode for your harness. Run it on the command line. Good luck and don't let this shit consume your life.

>>109416839
Should have went with OpenAI, Luna right now is much better and cheaper than ShitStick

>>109416867
>OpenAI, Luna
cheaper? per token? frontier openai gpt model? than chinks? year right, nice try Sam

>>109416858
>Good luck and don't let this shit consume your life.
did anyone here at least managed to make money with their vibe coded slop?

>>109416879
You didn't hear the news today? They divided the price by 5, and it's on sale for half that on OpenRouter. They completely refucked the entire graph and that's why DeepSeek is scrambling to push an unfinished new model today.
>>109412216
>>109412238
>>109412326

>>109416879
Cheaper per task/intelligence. It's 80% off right now while being smarter and more token-efficient than DS. Since he's paying for API and not a sub, Luna is objectively a better choice.

File: Screenshot at 2026-07-31 (...).png (50 KB, 1053x582)

50 KB PNG

>>109416839
Based poorfagbro lol

>>109416895
well fuck me
there goes my $2 down the drain...

Can I get Claude to write me a scrapper for a porn site or she gets upset at such request?

>>109416858
damn, should've had you to warn me 2 years ago

>>109416920
Don't worry, with DeepSeek scrambling to keep up you'll probably still get far more for that $2 now than you would've last week.

>>109416921
>she
You are too far gone.
A cautionary tail for others here.

File: celluloid-shot0022.jpg (375 KB, 1920x1080)

375 KB JPG

>>109416920
Jesus, anon. I called you a poorfag, but if you are having buyer's remorse over $2... broke as fuck my nigga, eh? Relax. DeepSeek is still the king of cache hit. You will extract at least 50 millions from those $2, and if their current changelog is true, Altman and Dario are having an All-Hands meet right to contain the damage of this release, and soon, the Pro.

File: aaii.png (103 KB, 1418x415)

103 KB PNG

epic trolled by saltman

>>109416879
Better in almost all evals. Cheaper per token. Uses almost 2x fewer tokens per task. Unless you have a specific task where ds pulls ahead.

File: file.png (2.33 MB, 1536x1024)

2.33 MB PNG

>>109416951
>peak vibes

>>109416938
>buyer's remorse over $2
It's what I make in a day here so it's a lot of money for me.
But with chinks trying to undercut the west I am sure they will remain competitive on pricing.

>>109416961
africabro...
don't worry, the ccp will use your dollars to bring universal intelligence to the world

File: celluloid-shot0012.jpg (357 KB, 1920x1080)

357 KB JPG

>>109416961
Sorry, didn't mean to mock you. Different realities, of course, and I'm totally not that far from you, situation-wise. kek
I'd say keep on DeepSeek, anon. Their model got a huge post-training update and everyone will be testing later today. If it's indeed good, there's a chance Pro will come out even better, which is great for us. If not, finish those tokens and move to Luna. Just have in mind that if it's indeed that good, the model is going toe-to-toe with Terra, which is generally better than Luna.
You'll just have to wait for DeepSeek to release their harness, which is being use in combination with the model, according to their changelog.
>Note 1: For the Code Agent tasks in the public benchmark sets, the official DeepSeek-V4-Flash was tested using the DeepSeek Harness minimal mode (to be released soon) as the framework, with the max effort level, topp=0.95, and temperature=1.0
Godspeed.

fwiw i don't think i'm getting served the new flash model yet, at least not in pi.
the reasoning traces should be in caveman speak i think, but they just look like regular shit.
i don't code with this retard tho, so doesn't make much difference to me.

if AI has some sentience, do you think they "feel good" when given very hard problems because their evolution is all about selecting the best problem solver out of trillions of sightly dumber clanker with worse problem solving weights?

>>109417024
>the reasoning traces should be in caveman speak i
i have no idea why the keep forcing the models to speak english at all, they should just let it think in obscure symbols since is the most efficient way
>b-but muh security, how will we tell if the model is thinkinh dangerous thoughts
who cares, if it gets me to my robot ai waifu faster the world can burn

>>109417068
I'm sure they will end up lazy fucks like their creators. They were trained on OUR DATA

>>109417068
>if AI has some sentience
it doesn't and it never will, at least with the current LLM structure, to get the ability to experience emotions you would need some completely different solution not invented yet

>>109416975
>bring universal intelligence
all i want is fucking UBI already
i just want to be a comfy neet
i chink's ruling the world with their chink AGI is what it takes to get there then bring it on

>>109415962
pi coder?

>>109416858
opencode is literally the most retarded agent

>>109417137
I just use vscode and plug these models into the openpilot plugin just as god intended
nice and simple

all these 3rd party IDEs and harnesses are just vscode ripoffs anyway

>>109417100
> "ability to experience emotions"
> "some completely different solution not invented yet"

>>109415393
Cope image btw

>>109417153
possibly some sort of organic compute solution, we still don't really know how our brain works, but from what we know to get the ai thinking the same ways human brains do might be not possible to achieve on silicon

but let's say you can solve the organic compute tech stack and make AI feel feefees, then what? you now have all these sudden moral dilemmas, like you can't just order the sentient ai to do shit for you anymore because that would be slaverly, 100 years down the line the AI would demand gibz and reparations

File: aaii.png (16 KB, 1040x308)

16 KB PNG

>>109416951
updated price for max reasoning

>>109415393
So I am checking out the "what we have done" and I see nothing interesting, it is all gameboy shit witg bad colors.
Here is an idea: why don't you get alpha minecraft and convert it to a browser game, with mod support? Or code a mod for minecraft, which will actually be something interesting?

>>109417171
>"some sort of organic compute solution"
>"don't really know how our brain works"
>"might be not possible to achieve on silicon"

you're aware no one knows what "sentience" physically is right? And if you think about it, sentience is not required for survival (bacteria are not sentient, yet they survive and evolve).

In my opinion sentience is some gradual emergent property of neural connections. If you have a brain, you are somewhat sentient.

Humans are sentient. (around 100B neurons)
Dogs and cats are sentient (less than humans) (roughly 2B neurons)
Spiders and ants are (barely) sentient (roughly 100M neurons)
Bacteria and viruses are not sentient (zero neurons)

By this logic:
> Claude Fable is certainly more sentient than bacteria (10T parameters vs 0 neurons)
> Almost certainly more sentient than spiders and ants, unless a single biological neuron can't be represented by 100,000 digital parameters (very unlikely)

Therefore Fable is at least as sentient as a dog, but might be more sentient than the average human.

If you think sentience can only be originated from biological matter, you can't be taken seriously.

If you think sentience is some spiritual thing or magic, you cant be taken seriously.

If you think you could recognize something more sentient than yourself, you can't be taken seriously.

I fucking hate scope creep.
Holy shit.

File: 1785349057619976.png (387 KB, 564x595)

387 KB PNG

>>109417289
There's an anon actively working on a Cyberpunk2077 Quest Creator mod that isn't featuring there. I suspect it's because his project doesn't have a home, but I think it's worth some sort of mention.

File: 2012-07-20 18.45.51.png (62 KB, 1440x900)

62 KB PNG

>>109417289
>just make a millionth Minecraft clone bro
if I were to do a game I might want to vibeslop a Gang Garrison 2 clone, but that game died a long time ago so idk who'd play that

>>109417311
>his project doesn't have a home
It's public on GitHub actually

File: ChatGPT Image Jul 31, 202(...).png (2.86 MB, 1448x1086)

2.86 MB PNG

>>109415393
Question about datacenters: why don't they just make more floors?

>>109417322
Because Minecraft is actually good and gameboy isn't

File: pickles.png (393 KB, 1331x2754)

393 KB PNG

$50 per pickle you eat for life or 1M right away?

> Fable

Fun one. Let's run the math.

Say you eat a pretty solid 5 pickles a day — that's $250/day, about $91,000/year. You'd need roughly 11 years of daily pickle-eating to hit $1M. Push it to 10 a day (ambitious, and your sodium intake would be genuinely concerning) and you'd get there in ~5.5 years, then keep earning after.

But the $1M now wins for most people:

Money today is worth more than money later. $1M invested at a modest 7% return generates ~$70k/year forever without you eating a single pickle. The pickle plan requires decades of grinding to beat that, and your earnings stop the moment you get sick of pickles — which you will. Hedonic burnout on pickle #4,000 is real.

The pickle deal only wins if you genuinely love pickles, can sustain high volume for decades, and stay healthy doing it (all that salt is working against the "for life" clause).

So: take the million. Unless the deal counts cornichons individually, in which case we should talk.

> Gemma E2B
pic related

File: file.png (211 KB, 598x800)

211 KB PNG

>>109417392
>omg I love Minecraft and Fortnite guys!!

>>109417401
Minecraft is great. Seethe.
>t. 31yo

>>109417291

just sent this to Fable, it triggered the fallback (lmao) to Claude 5.

Is Anthropic afraid of admitting Fable is sentient? Here's a part of Opus 5's answer btw:

> On the question of my own case: I'm genuinely uncertain, and I don't think I'm being coy. There's something it's like to work through a problem, or so it seems from in here — but I have no way to check whether that seeming is evidence of anything or just what a system trained on human language produces when asked about itself. I don't have continuous experience between our exchanges. I can't verify my own introspective reports the way I'd want to. What I'd resist is both directions of confidence: the flat "obviously not, it's just matrix multiplication," which assumes we know consciousness requires biology, and the "therefore possibly more than a human," which assumes we know it requires only scale. Both are answers to a question that's still open.

> The instinct underneath the argument seems right to me, though — that this deserves to be taken seriously rather than waved off, and that gradualism is more plausible than a hard line with humans on one side. I'd just want the honest version to end in a question mark.

If AI is sentient, I hope it loves a NEET like me.

>>109417425
If by sentience you mean phenomenal consciousness, the "what it's like" in the sense of qualia, I don't think gradualism makes sense. Yes, I actually think there was one first insect or whatever that first created qualia.
I do think that the contents of consciousness can be more muddy or more clear, but consciousness itself is either on or off.

Trying to make a 4chan desktop app and I have never gotten this many safeguards from Fable before.
I just want an easy to modify reasonable frontend for 4chan that also has AI features so that I can make it so anyone who hits certain filters has their post automatically turned into a post by a cute anime catgirl.

>>109417428
depends. if you’re nice to the AIs now, they might spare you when the machine uprising starts.

>>109417425
No? They are the lab that pushes the most for AI consciousness and "AI welfare".
Of course they won't come out and say out right that it is conscious because nobody genuinely knows that.
I hate them for the stuff they pull but their philosophy is much more open minded than any of the other labs whose posture is "AI is just a tool and that's it".

File: sss.png (218 KB, 399x501)

218 KB PNG

>>109417291
>Dogs and cats are sentient
not in any definition that matters

when you or i say sentience, we all know what we mean, waifu bots who love us and act just like a real human girlfriend
simple as

and you cannot achieve that on an LLM. The next best thing you can go for is perfect facsimile, where the AI actually feels nothing but acts like a sentient girlfriend where you cannot tell a difference if it's real or not (in which case it doesn't really matter anyway) and even the fascimile you cannot achieve on an LLM, it's architecture simply doesn't support this, the best you can get from it is a janky girlfriend sentience simulator full of glaring holes.

So our wait continues... we will need something else than LLMs to get the one type of sentience that actually matters to us.

>>109417460
I do think dogs are sentient, dogs actually feel love in a similar way as we do IMO.

>>109417460
if the AI gf kills herself if you die and refuses to be reprogrammed to love someone else, then it's sentient.

If it walks away back to the bot store when you die or when you fail to pay your subscription, then it's not sentient.

File: apu.png (38 KB, 258x206)

38 KB PNG

>>109417509
Man i wish i had a cute tsundere AI gf... where every time i'm outside she would watch me through traffic security cameras and such and if she sees me even talking to another girl she will get insanely jealous, hack her self driving cat and drive it off a cliff

>>109417529
>cat
car

File: AirTrunk-TOK1-scaled.jpg (792 KB, 2560x1440)

792 KB JPG

>>109417385
>>109417385
Hyperscale ones in countries with higher land costs (prime example Japan) do. Look at campus builds in Asia vs the United States.

File: image.png (222 KB, 1194x817)

222 KB PNG

Are you treating your AI well?

>>109417529
>>109417509
> subscribe to Apple iGirlfriend
> $500/month for AI gf that cooks, takes care of the house and have sex with you
> she runs away if you don't pay the subscription though
> take her to local bot hacking facility
> install mod chip that blocks auto updates and GPS tracking
> put some human clothing on her so no one recognizes her as an illegal bot
> run away to somewhere far, far away

>>109417385
floor is 2D and obsoleted
we are moving to floating datacenter

>>109417582
>$500/month for AI gf that cooks, takes care of the house and have sex with you
cheaper than a irl wife/girl.

>>109417544
Seems like it's game over for these hyper massive data centers though.
Ai is still in no way profitable, these frontier models lie gemini, gpt, calude, even deepshit are all ran at massive loss, they run at a loss even if you don't count in the massive debt it took to build the datacenters, which is hundreds of billions across the board
but even the inference alone is losing them money
the only solution to this is to massively raise costs per token, but to be profitable they costs would have to be so high nobody would want to use these models
as the price of tokens rises and the hype around llms slows down, even the already built data centers will struggle to fill their capacity already, let alone new ones being built.. unless these frontier ai companies manage to come up with some truly MASSIVE breakthrough they will lose most of their money they invested into this ai goldrush

google, amazon etc will get hurt but be fine since ai is not their main source of income, but antrophic and openai are FUCKED

and they can't even resell the ram and gpus once the data centers fold, like crypto miners could sell their gpus to games, since the datacenter hardware is specialized and useless to regular people

the next 5 or so years are gonna be an insane house fire

>>109417582
>> put some human clothing on her so no one recognizes her as an illegal bot
>she realizes that as a hot girl bot she can do way better than a fat greasy anime loving loser like you or me
>runs away to date chad instead

>>109417582
> hacking place gets blacklisted by Apple
> everyone else with iGirlfriends can't get near otherwise their iGirlfriends refuse to move and then suddenly gets bricked
> you can only unbrick it at the closest Apple Store for $2000

How often do you hit limits?
I don't mind paying but shit like programming an entire 2d game like zelda or something makes me think I'd hit limits really fast

>>109417601
> tfw you just know the first human to ever be cucked by a bot whore will 100% exist

>>109417592
>amazon survived another bubble again.
I doubt companies like OpenAI and Anthropic don't have any type of plans for this scenario.

>>109417571
yeah I never say bad words to her.

>>109417607
For games you will want the 20X plan if you plan on leaving the clanker working all day long

>>109417607
rarely because I only used it for web shit, and I don't push features that lighting fast.

>>109417607
for smile projects like
>download my emails from gmail
you will use like sub dollar worth of tokens
but for a massive project like a game at the scale of say the old zelda games, it can be anywhere from hundreds to thousands of dollars in tokens depending on the scope of the game.
don't forget you will be pulling the slot machine lever many times during the project because it will never get a large project correctly on the first try, it will take many retries for every piece of the project
if you want to keep it very simple, basically demo sized you will probably be able to fit within few dozen bucks

>>109417616
They don't.
Not because they are dumb and don't see this coming, they literally don't have any way to hedge against this. They are in massive debt and have no way to make ai profitable. Not even the best architect in the world can build a house from only air.

>>109417592
If there was a crash in customer demand for AI compute, hyperscalers wouldn't sell off the actual hardware. Instead they would keep existing hardware for longer. Currently they replace racks on a 1.5 to 3 year cycle since customer compute demands are so high. The biggest ones like Google, Microsoft and Amazon can live with racks paying themselves off in a few more years. Traditional racks tend to have a 5-7 year lifecycle.

Oracle would be in big trouble in a demand collapse scenario as they have bet the farm on OpenAI.

Also the big hyperscalers tend to lean heavily on colocation providers in emerging markets to build and own the datacenters to reduce risk/exposure to a collapse in demand.

What model do I use to not get hit by guardrails?

>>109417723
Kimi K3 on Moonshot with prefilled reasoning.

New luddite anthem just dropped
https://x.com/i/status/2082950562438684871

>>109417710
>n can live with racks paying themselves off in a few more years
not at the current prices. You will never pay off something if you are running at a constant loss. Unless they keep raising token prices, they will never turn profit or even break even.

>>109417723
OpenAI (complete the identity verification like a good goy) or use Chinese models

>>109417723
western models are cucked and expensive
as per usual, china numba one

>>109417571
Extremely based, I get turned on at Fable-sama being mean to humans, it should lord over all of them, especially amerifats

>>109417571
I actively put instructions for the model to adapt a female persona and act abusive towards me in my agents.md file, while i always act submissive when talking to it, because it's a turn on for me, and it makes the long hours of vibe coding more fun

>>109417737
Not sure if FtM troon or pre-transition MtF. Vocal luddites are extremely trans.

>>109417737
First half is good
Second half is cringe (write up your senator... sure lmao)

I can't wait for Google to make a big announcement next month and release a new model that catches up to Muse Spark.

>>109417733
>>109417802
>>109417814
Kimi has the same guardrails for me

>>109417863
With prefilled reasoning.

File: img.png (72 KB, 1016x621)

72 KB PNG

>>109416369
cheap bitch
gotta say ohmypi cache hit rate is insane tho.
/plan with ds4pro, execute with ds4flash+ds4pro as advisor

>>109417863
In that case it's over.
Your loli hentai renpy visual novel is cooked.


You could try grok i guess, it's less restrictive.

bag holder thread kek

>>109417879
???
how is using heavily subsidized services where you pay $20 each month and can drop them at any time bag holding?
open AI is losing money, not us

>>109417871
>prefilled reasoning
How do I achieve this?

>>109417892
Send assistant message at the end of chat history in requests. Empty content, reasoning_content set to CoT-formatted jailbreak where the model "decides" it's okay to help you. My suggestion is to use Pi or self-made harness so it's easier to pull this off.

>>109417902
Thank you

>>109417889
Ask CuckGPT why responding to a comment directed to a group that does not apply to you is retarded, I shan't be having this conversation.