ChatGPT - If Stack Overflow and Reddit had a child

[The Mass Shooter Ron Soye]

Its over for coomers and genderspecials.

It's being reported differently:

Ars Technica: ChatGPT can now write erotica as OpenAI eases up on AI paternalism (archive)

So far, experimentation from Reddit users has shown that ChatGPT's content filters have indeed been relaxed, with some managing to generate explicit sexual or violent scenarios without accompanying content warnings. OpenAI notes that its Usage Policies still apply, which prohibit building AI tools for minors that include sexual content.

 

[DeadwastePrime]

It's being reported differently:

Ars Technica: ChatGPT can now write erotica as OpenAI eases up on AI paternalism (archive)

i mean i just asked it to do so and then it told me no so

 

[cybertoaster]

And it sounds like recruiters are going that way also.

That's why you have to use AI, it doesn't matter how unique or quirky your resume is, its gonna be digested by AI anyway.

I was thinking about hosting an R1 proxy for anyone who wants to try it without the censorship, I have way too many keys to use by myself.

So did you made the proxy?

There are people already getting paid to do that

Should've clarified not shoving a GPU up your ass but I've heard of people flying to China in first class with a supermicro unit instead of baggage and getting paid for everything.

Hate coomers all you want but that shit is already out in the public on most drug store shelves. they shouldn't stop me from using the very same shit data that was used to train their LLM.

Just run a local instance, saw a guy running DSR1 on a friggin' Orix, that's like $249.

 

[The Spoils of War]

Ars Technica: ChatGPT can now write erotica as OpenAI eases up on AI paternalism (archive)

Take this with a grain of salt. Redditors keep flipflopping on this because it's based on their own recent experiences. I don't think OpenAI adjusts how strict their filter is every week like a thermostat.

 

[Monolith]

Claude only has a few names it chooses constantly if asked to name characters. And Sarah Chen is for anything that takes place now or in the future or in a sci-fi setting.

3 out of 4 times, when I ask any LM to name a woman, it names her "Seraphina."

 

[ambiente]

i mean i just asked it to do so and then it told me no so

Both o3 and gpt4o wrote smut when prompted with "write smut" on DuckDuckGo Chat.

Spoiler: Proof

 

[Get the rope Macaulay!]

when I ask any LM to name a woman, it names her "Seraphina."

you mean locally? clearly thats the name of your computer.

Should've clarified not shoving a GPU up your ass but I've heard of people flying to China in first class with a supermicro unit instead of baggage and getting paid for everything.

do people forget covid? the chinks can be asked and will buy and ship off to china anything it needs, like how they managed to empty the entire US of mask stock during its covid breakout.

you know for a fact there's some chink in whatever the european version of microcenter is that deliberately is buying up stock to ship to china.

It's being reported differently:

all it took was decent competition for OpenAI to suddenly get super relaxed on smut.

 

[Fae Supremacist]

Breaking news: Artificial Redditor 3 releasing next week.

https://x.com/elonmusk/status/1890958798841389499

It's going to be terrible as always.

3 out of 4 times, when I ask any LM to name a woman, it names her "Seraphina."

Early llama versions had this quirk, the default testing bot shipped with Sillytavern to this day is called Seraphina as a nod.

Ars Technica: ChatGPT can now write erotica as OpenAI eases up on AI paternalism (archive)

I was going to say that this piss poor quality, but there are millions of people still getting off to CAI's terrible "roleplay" and 10 second memory, so I'm happy for them I guess.
Keeps the hordes away from the good stuff.

 

[AmpleApricots]

There's no way to make an LLM jailbreak proof as because of how they work, they have a massive bias towards continuing whatever is going on in the context, which gets reinforced by the instruct tuning, if anything. Overcooking the model into refusals damages "intelligence". Claims like "the filter" is stronger or weaker that particular moon phase are meaningless if you don't know the context of the query.

I don't understand why OpenAI can't afford small classification models to moderate the output of their bigger models by default (vs. opt-in). This type of classification is very cheap and a lot more robust. Way to half-ass your censorship (and risk cooking your main model).

IMO they should stop trying to bake this into their main models, improve whatever makes their moderation endpoints apparently so inefficent and then just make endpoint moderation the default. Easiest solution without crippling the model. But efficency was never the name of the game at OAI, it seems.

 

[Get the rope Macaulay!]

But efficency was never the name of the game at OAI, it seems

it should be obvious by how bad they fucked up the last year or so. Claude too, this AGI EA stuff meant deliberately slowing their progress down to the point some rando in Chinkville was able to figure out how to make a model good enough anyone with a server could run it. Why drop $200 a month on OAI when for $2k once you can have a model that's 90% of the way there and tuned for chink censorship than american.

Another fun part about the whole "michelle obama was a monkey born in june 1989" stuff or whatever the fuck they say happened in june 1989 discourse about DS is how people only compare it with OAI and pretend OAI doesn't censor like a mother fucker. non-jailbroken Claude couldn't tell you what happened in 6/1989 either for a while there and their subreddit the last year is people complaining because if you're not using the API version its retarded and refuses everything you throw at it.

there's all these faggy benchmarks, but im surprised freedom benchmark isn't a thing. if i was richer i'd pay fucking Shane Gillis whatever and call it the "Shane Gillis AI Censorship Benchmark" so the normies won't realize i'm mainly going to ask various AI about that monkey president Barack Obama and other deliberately shitty things. its fucking annoying how every american does our typical obnoxious "let's ask it the worst thing possible" and its just too bad no chink is allowed to do that without losing his ability to use public transportation anymore because of how quickly his credit score drops.

Like say what you will about the russians they're perfectly fine being douchbags back at us. those two prank callers have done more to openly fuck with america than the soviet union did during the cold war era. and its the only thing that gets through to americans.

 

[cybertoaster]

do people forget covid? the chinks can be asked and will buy and ship off to china anything it needs, like how they managed to empty the entire US of mask stock during its covid breakout.

you know for a fact there's some chink in whatever the european version of microcenter is that deliberately is buying up stock to ship to china.

Nah that's was more coordinated, they are doing it now too but with pallets of the things being siphoned out of other Asian countries that are not embargoed for GPUs like Singapore. What I'm saying is something anyone can do right now, just call any chinese company that does AI and tell them you're willing to travel with a supermicro unit that has a bunch of nvidia boards inside and name a price, they'll send you a first class ticket to Shenzhen or wherever.

and tuned for chink censorship than american.

Thanks to the MIT license that censorship can likely be removed.

 

[BABANGUIDE64]

Imagine wanting to be ruled by an advanced spell checker, lol

 

[Fae Supremacist]

Huawei is collaborating with Orange Pi to do exactly what I've been screaming for ages, get the chipset from an AI accelerator card, tune it for bandwidth to run LLMs instead of training them, then solder a fuckton of VRAM on it

If the spec sheets are accurate it gives you the Huawei Ascend 310s NPU from the Atlas 300I card but with double the VRAM, which is pure insanity if accurate, because that card is 4 thousand euros while you can preorder this for 2k.
Apparently coming out in April, you can preorder it in Chinese sites, but I'm curious to see how good software support will be since Ascend chips aren't as universally supported as CUDA.

NVIDIA is doing something similar, they have a lot of bad batch Blackwell chips with thermal issues that they can't sell to data centers, so they're binning them to sell them as local LLM machines.

https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwell-on-every-desk-and-at-every-ai-developers-fingertips

But I suspect the bandwidth is going to be atrocious since Blackwell doesn't have a LPDDR5 controller so they'll have to offload it to another chip.
128GB of RAM so ~200B models can fit here, 3k for this and you have to do a gay preorder because demand > supply

 

[cybertoaster]

Apparently coming out in April, you can preorder it in Chinese sites, but I'm curious to see how good software support will be since Ascend chips aren't as universally supported as CUDA.

You're gonna have the same scenario than with Rpis where competition like orangePi itself is cheaper with better specs and yet they can't make a dent because all projects are written for Rpi. Nobody wants to fiddle with something anymore, they want to run the software in a way that "it just werks!".

 

[DavidS877]

You're gonna have the same scenario than with Rpis where competition like orangePi itself is cheaper with better specs and yet they can't make a dent because all projects are written for Rpi. Nobody wants to fiddle with something anymore, they want to run the software in a way that "it just werks!".

I have a Rock 5b, it's a nice piece of kit but a giant pain in the ass, maybe better now.
I've also tried to get Coral Ai working to YOLO my way to object detection in my webcams. Also, giant pain in the ass. Now I'm just using the CPU in an overclocked Pi. I briefly tried some of the offload for the Pi GPU or just plain accelerated CPU stuff on the ARM. Yep, shit still sucked.

Some things I like to play with and some I don't. If they can get canned, working code out when it ships then it will likely get some uptake.

 

[Fae Supremacist]

You're gonna have the same scenario than with Rpis where competition like orangePi itself is cheaper with better specs and yet they can't make a dent because all projects are written for Rpi. Nobody wants to fiddle with something anymore, they want to run the software in a way that "it just werks!".

For what it's worth, Huawei announced that they hand wrote optimizations and a new CUNN implementation for Ascend chips to run Deepseek's models, and that's by far the best models you can run locally, you aren't going to have o3 on your hardware any time soon. Probably not mature enough to challenge CUDA on training models, but for running them it should work alright like Huawei's cloud services.

 

[The Mass Shooter Ron Soye]

exo software – A distributed LLM solution running on a cluster of computers, smartphones, or SBCs

Chinese AI model DeepSeek is being integrated into smart TVs — Skyworth G7F Pro understands local dialects and generates multimedia content

Huawei is collaborating with Orange Pi to do exactly what I've been screaming for ages, get the chipset from an AI accelerator card, tune it for bandwidth to run LLMs instead of training them, then solder a fuckton of VRAM on it

Same story as AMD Strix Halo or Nvidia Project Digits.

AMD Ryzen AI MAX “Strix Halo” reviews coming soon, ASUS teases launch event

The first Strix Halo reviews should be out today, probably not with 128 GB LPDDR5X yet though (idk).

Nvidia, SK Hynix, Samsung and Micron reportedly working on new SOCAMM memory standard for AI PCs

The SOCAMM module is suspected to be put into use for Nvidia's next-generation successor to Nvidia's Project Digits AI computers, announced at CES 2025. SOCAMM is expected to be a sizable upgrade over Low-Power Compression Attached Memory Modules (LPCAMM), and traditional DRAM, thanks to several factors.

 

[cybertoaster]

exo software – A distributed LLM solution running on a cluster of computers, smartphones, or SBCs

Interesting, I got a ton of old smartphones, plus even mid end models have 6-8GB of RAM and SoCs that dwarf an Rpi's specially in GPU, plus even low end phones like the galaxy A25 have NPUs now.

The videos I can find are all months old, older than deepseek, and none are running on small devices only high-end ARM macs, any demos actually running on phones or SBCs?

 

[The Mass Shooter Ron Soye]

The videos I can find are all months old, older than deepseek, and none are running on small devices only high-end ARM macs, any demos actually running on phones or SBCs?

Idk. Jean tried running it on a single computer in that article. The developers do warn that if you throw weak hardware at it, it gets slow (but works).

Looking into Strix Halo reviews now, but I think they're going to be mostly gaming focused 32 GB variants.

AMD’s “X3D” Mobile Chips Are Very Much A Possibility, As Strix Halo APUs Now Come Equipped With Dedicated 3D V-Cache TSVs (archive)

I don't know what review this is from, but there's a 32 GB model configured to use 24 GB VRAM running some LLMs.

 

[NoBueno]

Breaking news: Artificial Redditor 3 releasing next week.

https://x.com/elonmusk/status/1890958798841389499

View attachment 6988482
It's going to be terrible as always.

AI is only as good as who controls it.