AI Topic: AlphaZero, ChatGPT, Bard, Stable Diffusion and more!

[trsttte]

OpenAI's chaos does not add up - Built Not Found

Update: this post has been instantly demoted from #1 to #26 on HN frontpage :) Hmm.

builtnotfound.proseful.com

This crap just doesn't add up at all. At this point I'm just trying to make sense of everything.

How to aqcuire a very valuable company for next to nothing 101. I mean, I don't know but there's really two possible scenarios:
1- OpenAI board is super super dumb and decided to commit company suicide by overplaying their hand and giving their most valuable assets - the talent working there - away to a willing competitor.
2- Microsoft engaged in some very anti competitive, shady and unethical behaviour by negotiating special deals with OpenAI talent and the boad overreacted and made microsoft's job even easier.

Either way looks like OpenAI is finished. The special thing they had was the talent working there, the tech is not particularly special - a variation of google's transformer model that everyone else can and is replicating - and they simply have no business model to commercialize their LLM and won't be able to support the costs of running the service when Microsoft pulls out.

 

[thesmokingman]

How to aqcuire a very valuable company for next to nothing 101. I mean, I don't know but there's really two possible scenarios:
1- OpenAI board is super super dumb and decided to commit company suicide by overplaying their hand and giving their most valuable assets - the talent working there - away to a willing competitor.
2- Microsoft engaged in some very anti competitive, shady and unethical behaviour by negotiating special deals with OpenAI talent and the boad overreacted and made microsoft's job even easier.

Either way looks like OpenAI is finished. The special thing they had was the talent working there, the tech is not particularly special - a variation of google's transformer model that everyone else can and is replicating - and they simply have no business model to commercialize their LLM and won't be able to support the costs of running the service when Microsoft pulls out.

There's something going on that feels like some power is trying to keep under wraps. Like no doubt Altman is next level douche going from help the world to max profits w/o being upfront and honest about it. The board though kind of saw this coming a mile away so I wonder or doubt that its Altman's max profit move.

And yea, then again you can't discount some next level shadiness from MS.

 

[A&P211]

The current rumor is the CEO is OpenAI fired comes from the link below.

The rumor is that the board of OpenAI feared that AI was moving faster than they thought, they feared for the future of humans. OpenAI might have a huge breakthrough in AGI, Artificial general intelligence. If true, AGI is currently doing grade school level math.
AI has issues with doing math since math requires a correct answer, different from writing.

OpenAI researchers warned board of AI breakthrough ahead of CEO ouster, sources say

Ahead of OpenAI CEO Sam Altman’s four days in exile, several staff researchers wrote a letter to the board of directors warning of a powerful artificial intelligence discovery that they said could threaten humanity, two people familiar with the matter told Reuters.

www.reuters.com

 

[Assimilator]

Remember when I was talking about how ChatGPT is fundamentally dangerous because it's designed to give you answers that you want to hear? Latest example.

ChatGPT generates fake data set to support scientific hypothesis

Researchers say that the model behind the chatbot fabricated a convincing bogus database, but a forensic examination shows it doesn’t pass for authentic.

The current rumor is the CEO is OpenAI fired comes from the link below.

The rumor is that the board of OpenAI feared that AI was moving faster than they thought, they feared for the future of humans. OpenAI might have a huge breakthrough in AGI, Artificial general intelligence. If true, AGI is currently doing grade school level math.
AI has issues with doing math since math requires a correct answer, different from writing.

OpenAI researchers warned board of AI breakthrough ahead of CEO ouster, sources say

Ahead of OpenAI CEO Sam Altman’s four days in exile, several staff researchers wrote a letter to the board of directors warning of a powerful artificial intelligence discovery that they said could threaten humanity, two people familiar with the matter told Reuters.

www.reuters.com

Oh man, no matter how many times I see this stupid shit repeated, I never get tired of it.

Let me make something clear, these "researchers" consistently claiming "AI breakthroughs that threaten humanity" don't exist. The "sources" are PR people in the company feeding this nonsense to journalists because they know the "journalists" of today are too stupid and irresponsible to do due diligence, they'll just publish whatever random crap is thrown at them. All OpenAI care about is keeping their name in the news and this is the exact sort of press that accomplishes that, for free.

 

[masterdeejay]

What is the best buy used GPU for stable diffusion? I like the old tesla cards. I can make a custom cooler.

M40 80-100 usd 12Gb GM200
M10 ~150 usd 32gb 4xGM104 (weak GPU-s?)
P4 70 usd 8Gb GP104 (easy to cool)
K80 100 usd 24gb 2xGK210 (cuda version is low?)
or there a few ex mining card P104-100 8Gb 60 usd (cuda capable?)
or is there any better GPU? (~50-150 usd range) older AMD cards (radeon pro, MI...) is also cheap but i don't know the support for the stable diffusion.

i have two old tesla cards but vram and the cuda support is too low i think. (C2075 6gb and K20C 5gb)

The pc for this project:
Dell T5820 (Rebar enabled and turbo unlocked bios)
Xeon 2699v3
4x16Gb ram
Win10 Pro, but linux is possible if needed.
Thanks!

 

[dragontamer5788]

Police Say AI-Generated Article About Local Murder Is "Entirely" Made Up

An AI-generated article about a murder that never happened in New Jersey has police and local news outlets in a tizzy.

futurism.com

Police are complaining about a (now deleted) article that falsely claims a murder happened in town.

It sounds like some lazy journalist clicked a few buttons on ChatGPT to make a story, and it just fully hallucinated a murder wholesale. The story got published for some reason (and then quickly removed). I'm not sure if this rises to the level of libel or slander though, since it'd be hard to say who exactly would sue this newspaper.

Still, it serves as a warning case to the new era we're in. Lazy people are clicking AI buttons they don't understand and just publishing the text without even reading it.

 

[the54thvoid]

Police Say AI-Generated Article About Local Murder Is "Entirely" Made Up

An AI-generated article about a murder that never happened in New Jersey has police and local news outlets in a tizzy.

futurism.com

Police are complaining about a (now deleted) article that falsely claims a murder happened in town.

It sounds like some lazy journalist clicked a few buttons on ChatGPT to make a story, and it just fully hallucinated a murder wholesale. The story got published for some reason (and then quickly removed). I'm not sure if this rises to the level of libel or slander though, since it'd be hard to say who exactly would sue this newspaper.

Still, it serves as a warning case to the new era we're in. Lazy people are clicking AI buttons they don't understand and just publishing the text without even reading it.

Yup. Governments need to get IT savvy lawmakers* into gear. This is going to add to an already emerging problem.

*Yeah, I know. IT savvy lawmaker is an oxymoron.

 

[R0H1T]

The issue is there's no "incentives" for the lawmakers to penalize the perpetrators hard for this, although in the above case it's probably only a minor thing. Now if this were against a billion dollar company heads would obviously roll

 

[dragontamer5788]

Amazon.com

----------

ChatGPT users ruining Amazon, trying to automatically create titles and descriptions for various products. This is one of the many default error-messages that OpenAI / ChatGPT will emit in some circumstances.

People have been finding a huge uptick in Tweets and other social media where clear bot-behavior is happening.

 

[dgianstefani]

Yup. Governments need to get IT savvy lawmakers* into gear. This is going to add to an already emerging problem.

*Yeah, I know. IT savvy lawmaker is an oxymoron.

Getting lawmakers that don't belong in a nursing home would probably help with the IT savvy part.

Term and age limits when?

"AI" real time translation is the best use of it I've found so far. Little chance of catastrophic fk up.

 

[dragontamer5788]

[2401.12070] Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text

Paper arguing they've discovered an algorithm that can differentiate between LLMs and Humans.

 

[trsttte]

[2401.12070] Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text

Paper arguing they've discovered an algorithm that can differentiate between LLMs and Humans.

Their example on github:

from binoculars import Binoculars

bino = Binoculars()

# ChatGPT (GPT-4) output when prompted with “Can you write a few sentences about a capybara that is an astrophysicist?"
sample_string = '''Dr. Capy Cosmos, a capybara unlike any other, astounded the scientific community with his
groundbreaking research in astrophysics. With his keen sense of observation and unparalleled ability to interpret
cosmic data, he uncovered new insights into the mysteries of black holes and the origins of the universe. As he
peered through telescopes with his large, round eyes, fellow researchers often remarked that it seemed as if the
stars themselves whispered their secrets directly to him. Dr. Cosmos not only became a beacon of inspiration to
aspiring scientists but also proved that intellect and innovation can be found in the most unexpected of creatures.'''

print(bino.compute_score(sample_string)) # 0.75661373
print(bino.predict(sample_string)) # 'AI-Generated'

0.01% false positive rate you say? You're full of shit...

 

[KLiKzg]

The current rumor is the CEO is OpenAI fired comes from the link below.

The rumor is that the board of OpenAI feared that AI was moving faster than they thought, they feared for the future of humans. OpenAI might have a huge breakthrough in AGI, Artificial general intelligence. If true, AGI is currently doing grade school level math.
AI has issues with doing math since math requires a correct answer, different from writing.

OpenAI researchers warned board of AI breakthrough ahead of CEO ouster, sources say

Ahead of OpenAI CEO Sam Altman’s four days in exile, several staff researchers wrote a letter to the board of directors warning of a powerful artificial intelligence discovery that they said could threaten humanity, two people familiar with the matter told Reuters.

www.reuters.com

Well, they are right...in few years, imagine how much "work force" will need to change because of AI.

 

[A&P211]

Well, they are right...in few years, imagine how much "work force" will need to change because of AI.

Its already start(ing), amazon is trying out new robots that picks up tots.

 

[JWNoctis]

For what it worth, Stable Diffusion 3 is out, from a struggling pioneer of the field and without much of a splash, in a space already saturated by production-grade APIs and hobbyist finetunes of SDXL.

It's currently most noted for generating body horrors on any prompt requiring a smidgeon of anatomical knowledge and atypical pose. The baby has been thrown out with the bathwater at high velocity, in the name of safety.

The effect has been observed since at least SD 2.0, but I assume they had no choice except rendering it worse-than-useless for this purpose, in an official release of model weights to the general public, where legitimate artistic expression and anatomical studies, versus deepfaked AI smut - and worse, unprintable, things - often only differed by end use and distribution, and of course intention.

That, and being able to spell and understand composition prompts relatively well at last. It should work fine, so long as you don't use it to generate anything human-shaped beyond the basic.

P. S. Now I wonder whether they would have been better off, or at least subjected to less ridicule, had they been simply more open about the whole thing upfront, with disclaimers and acknowledgements like "To prevent misuse of the model by bad actors, we excluded potentially objectionable images from the training dataset, and utilized data poisoning to disrupt the utility of potentially harmful concepts. As a consequence, this model cannot reliably render human form in select contexts, and cannot produce accurate non-facial anatomical detail."

That, instead of the too-broad statement in their actual press release:

We believe in safe, responsible AI practices....We have conducted extensive internal and external testing of this model and have developed and implemented numerous safeguards to prevent harms.

 

[Assimilator]

For what it worth, Stable Diffusion 3 is out, from a struggling pioneer of the field and without much of a splash, in a space already saturated by production-grade APIs and hobbyist finetunes of SDXL.

It's currently most noted for generating body horrors on any prompt requiring a smidgeon of anatomical knowledge and atypical pose. The baby has been thrown out with the bathwater at high velocity, in the name of safety.

The effect has been observed since at least SD 2.0, but I assume they had no choice except rendering it worse-than-useless for this purpose, in an official release of model weights to the general public, where legitimate artistic expression and anatomical studies, versus deepfaked AI smut - and worse, unprintable, things - often only differed by end use and distribution, and of course intention.

That, and being able to spell and understand composition prompts relatively well at last. It should work fine, so long as you don't use it to generate anything human-shaped beyond the basic.

P. S. Now I wonder whether they would have been better off, or at least subjected to less ridicule, had they been simply more open about the whole thing upfront, with disclaimers and acknowledgements like "To prevent misuse of the model by bad actors, we excluded potentially objectionable images from the training dataset, and utilized data poisoning to disrupt the utility of potentially harmful concepts. As a consequence, this model cannot reliably render human form in select contexts, and cannot produce accurate non-facial anatomical detail."

That, instead of the too-broad statement in their actual press release:

I honestly don't understand this stupid prudishness from Stability - there is no way that they released an image generation product not knowing it would be used to create porn. This is the internet FFS, it pretty much exists to facilitate the sharing of smut. And if they're willing to hobble their product so it can't generate porn, I can guarantee you their competitors won't - so all Stability has accomplished is to lose marketshare to those competitors. GG WP, not.

Presumably they are concerned about deepfakes and far more objectionable types of porn being generated by their product, and being somehow held liable as a result, but honestly... that particular genie is already waaay out of its bottle. Just like you can't have the internet without porn, you can't have image generation without porn. Humans and sexuality are intrinsically linked, can we maybe start being rational about this?

 

[JWNoctis]

I honestly don't understand this stupid prudishness from Stability - there is no way that they released an image generation product not knowing it would be used to create porn. This is the internet FFS, it pretty much exists to facilitate the sharing of smut. And if they're willing to hobble their product so it can't generate porn, I can guarantee you their competitors won't - so all Stability has accomplished is to lose marketshare to those competitors. GG WP, not.

Presumably they are concerned about deepfakes and far more objectionable types of porn being generated by their product, and being somehow held liable as a result, but honestly... that particular genie is already waaay out of its bottle. Just like you can't have the internet without porn, you can't have image generation without porn. Humans and sexuality are intrinsically linked, can we maybe start being rational about this?

Considering that they are already being sued for copyvio, and I think unmentionables in the training dataset of the original SD 1.x, I do understand their behaviour with this release. As to the policy of their real, commercial, competitor, I plead ignorance.

But yes, given how the model otherwise does have merit, releasing the model weights into the wild will mean that someone will finetune all that safety engineering away within the month, if the model could be finetuned at all. It is Internet, after all. Sufficiently powerful models - models that can spell and do composition - had previously been locked behind commercial API with all their content policy and terms of service. All Stability gained here is plausible deniability, if it could be finetuned.

I have no idea how this one will go down, but it won't be good.

 

[dragontamer5788]

I honestly don't understand this stupid prudishness from Stability - there is no way that they released an image generation product not knowing it would be used to create porn. This is the internet FFS, it pretty much exists to facilitate the sharing of smut. And if they're willing to hobble their product so it can't generate porn, I can guarantee you their competitors won't - so all Stability has accomplished is to lose marketshare to those competitors. GG WP, not.

Presumably they are concerned about deepfakes and far more objectionable types of porn being generated by their product, and being somehow held liable as a result, but honestly... that particular genie is already waaay out of its bottle. Just like you can't have the internet without porn, you can't have image generation without porn. Humans and sexuality are intrinsically linked, can we maybe start being rational about this?

There's this white-market porn that Mastercard / Visa are willing to serve, grey-market where Mastercard / Visa thinks is too risky (but isn't technically illegal yet), and then the straight up black-market FBI-will-start-investigating-you porn variety.

When you consider that in today's world, "porn" includes CSAM, and that CSAM is actively used to attack public websites (ex: if I were a dick and wanted to harm TechPowerUp, I'd create a bunch of accounts and start posting CSAM to mess with the moderators), then it makes more sense. The way CSAM is tracked is by hashes and fingerprints, they don't want humans looking at the stuff on a regular basis so its a lot of automated processes. So if you have a new "porn generator" that could potentially create CSAM content, then you need to participate in the database or get fucking nuked by the powers that be.

Otherwise, I'd take all those .jpgs and start posting them to TechPowerUp (or whatever website you like), just to shut down other websites. And if all those .jpegs aren't in the hashing databases or whatever thata various websites use, then that means manual curation. Ex: Imagine how the moderators would feel if they had to spend ~2 or 3 hours manually deleting that stuff. Would I, the hypothetical troll, have gotten a moderator at work and thus in trouble with their IT department? Etc. etc. etc.

So yeah, the black-market porn / CSAM stuff is extremely toxic. I would assume that this generative AI would be used to create CSAM, and no, no one actually wants loads of AI-CSAM to start proliferating the internet. That'd just wreck moderators.

-----------

If there were assurances that the porn would "only" be of the white-market variety that everyone is cool with, then yeah, no one would have a problem with it. But **every** pornographic website has to deal with the inevitable rise of the CSAM community intruding, spreading that stuff, and then... well... things get really political real quickly.

The Grey-market stuff (ie: Visa/Mastercard doesn't like it, but FBI won't investigate you over it, because its not illegal yet) is also troublesome. Because that's sufficient to get demonetized in practice even if it isn't illegal. Visa/Mastercard has stricter standards because they want to be accepted by more than USA, but also around the world. So Visa/Mastercard pretty much plays by the "most conservative" rules across the spectrum of countries they serve. Like, illegal porn by Australia-standards would likely be banned by Visa/Mastercard (even if its legal in USA). And vice versa, illegal in the USA means its banned (by Visa/Mastercard) in Australia (even if its illegal in Australia). This grey-market stuff is a similar issue to the CSAM stuff, just led by the payment processor companies instead of any particular legal entity.

 

[Assimilator]

There's this white-market porn that Mastercard / Visa are willing to serve, grey-market where Mastercard / Visa thinks is too risky (but isn't technically illegal yet), and then the straight up black-market FBI-will-start-investigating-you porn variety.

When you consider that in today's world, "porn" includes CSAM, and that CSAM is actively used to attack public websites (ex: if I were a dick and wanted to harm TechPowerUp, I'd create a bunch of accounts and start posting CSAM to mess with the moderators), then it makes more sense. The way CSAM is tracked is by hashes and fingerprints, they don't want humans looking at the stuff on a regular basis so its a lot of automated processes. So if you have a new "porn generator" that could potentially create CSAM content, then you need to participate in the database or get fucking nuked by the powers that be.

Otherwise, I'd take all those .jpgs and start posting them to TechPowerUp (or whatever website you like), just to shut down other websites. And if all those .jpegs aren't in the hashing databases or whatever thata various websites use, then that means manual curation. Ex: Imagine how the moderators would feel if they had to spend ~2 or 3 hours manually deleting that stuff. Would I, the hypothetical troll, have gotten a moderator at work and thus in trouble with their IT department? Etc. etc. etc.

So yeah, the black-market porn / CSAM stuff is extremely toxic. I would assume that this generative AI would be used to create CSAM, and no, no one actually wants loads of AI-CSAM to start proliferating the internet. That'd just wreck moderators.

-----------

If there were assurances that the porn would "only" be of the white-market variety that everyone is cool with, then yeah, no one would have a problem with it. But **every** pornographic website has to deal with the inevitable rise of the CSAM community intruding, spreading that stuff, and then... well... things get really political real quickly.

The Grey-market stuff (ie: Visa/Mastercard doesn't like it, but FBI won't investigate you over it, because its not illegal yet) is also troublesome. Because that's sufficient to get demonetized in practice even if it isn't illegal. Visa/Mastercard has stricter standards because they want to be accepted by more than USA, but also around the world. So Visa/Mastercard pretty much plays by the "most conservative" rules across the spectrum of countries they serve. Like, illegal porn by Australia-standards would likely be banned by Visa/Mastercard (even if its legal in USA). And vice versa, illegal in the USA means its banned (by Visa/Mastercard) in Australia (even if its illegal in Australia). This grey-market stuff is a similar issue to the CSAM stuff, just led by the payment processor companies instead of any particular legal entity.

None of that's relevant, though. Stability isn't training their database on CP/CSAM (well, I bloody well hope they aren't) so the only way they could be exposed to such material is if someone uploads one of those images as a source... in which case they would match it to their CSAM DB and refuse to do anything with it (and hopefully send the cops to the uploader's house). That seems about the same level of care that any ordinary image host would be expected to provide, so I don't see how CSAM would be any more of a concern for Stability than for anyone else. In other words, I'm really not seeing how or why CSAM could be a strong reason for this company to censor their dataset.

 

[dragontamer5788]

Stability isn't training their database on CP/CSAM (well, I bloody well hope they aren't) so the only way they could be exposed to such material is if someone uploads one of those images as a source...

You don't need to have "good looking" CSAM for it to cause issues. You just need to output material that qualifies as CSAM.

Lobotomizing your AI so that its bad with humans in general sounds like a step too far, but think about it. If you can't make any human look good, the system as a whole cannot be used as CSAM at all. Its a very conservative approach but its what I expect is driving the decision making here.

 

[JWNoctis]

None of that's relevant, though. Stability isn't training their database on CP/CSAM (well, I bloody well hope they aren't) so the only way they could be exposed to such material is if someone uploads one of those images as a source... in which case they would match it to their CSAM DB and refuse to do anything with it (and hopefully send the cops to the uploader's house). That seems about the same level of care that any ordinary image host would be expected to provide, so I don't see how CSAM would be any more of a concern for Stability than for anyone else. In other words, I'm really not seeing how or why CSAM could be a strong reason for this company to censor their dataset.

You don't need to have "good looking" CSAM for it to cause issues. You just need to output material that qualifies as CSAM.

Lobotomizing your AI so that its bad with humans in general sounds like a step too far, but think about it. If you can't make any human look good, the system as a whole cannot be used as CSAM at all. Its a very conservative approach but its what I expect is driving the decision making here.

That's the crux. Current text-to-image models can extrapolate, and these capabilities get considerably more... let's just say, dual-use, with improved compositional capabilities mostly provided by an advanced encoder stage. Such capabilities existed years ago - I think prototype Imagen could do it, it's the first model I remember to have seen to spell reasonably well - but as far as I'm aware it did not make it into any well-known open-weight model, until now.

Getting back to the issue at hand, if you could - in that instance described in the Ars article - produce reasonable pictures of people lying on grass, you could reasonably expect the model to be able to produce <any person or combination thereof of any appearance and physique> lying on <anything physical or metaphorical in any context>. I'll just leave it at that. It's not a comfortable topic.

 

[JWNoctis]

And now, one monthly calendar roll-over later, exactly two finetunes of Stable Diffusion 3 appeared to show up on HF. Only one of them looked viable, but still proof-of-concept-y.

I take back my prediction of how at least someone was going to finetune its safety engineering away within the month. At least nothing showed up there. Meanwhile, SDXL LORAs, finetunes, and mixes still popped up left and right.

Seems like the community is more or less giving this one the pass, like they did with SD 2.0, and that's when SD 2.0 is not nearly as badly limited. Probably not surprising either, given how SDXL is actually larger than the currently released SD3-medium in terms of parameter count. Whatever one could say about Stability failing to monetize their models and efforts, and pending something unexpected happening, this looked to be how it would end, and certainly not with a bang.

 

[JWNoctis]

Light update and personal take for those not actively tracking the issue elsewhere, but still interested:

----

July and first half of August were interesting. FLUX.1 is essentially what SD3 should have been, incidentially trained by the same people behind the original SD, SD2, and SDXL in a fresh startup. It is sprouting a bumper crop of finetunes and LORAs as we speak, in stark contrast to the SD3. 8GB of VRAM is apparently no longer sufficient, though if you have more than 16GB of RAM, you can run it on CPU at maybe 1 image/hour. Authors have not disclosed & detailed source and makeup of training data. I could only wonder why.

Guess at this point it is more of a fait accompli, for all the goods and ills it would do, and all the lives and careers it and other similar offerings ruin and derail. Now it is supposed to be about having the capability open & freely available for anyone who could afford the hardware, versus having it locked behind APIs by large corporations. It sounds lofty.

Incidentally, it still fails my favourite personal test of text-to-image capability since SD 1.5: Generate an image of an elf holding a bow, without mangling both the bow and the fingers, in plain text-to-image configuration. Though it is mostly better than any previous versions I have tried.

----

Mistral Large 2 Instruct is also released. Chatbot Arena and personal experience puts it above Llama 3/3.1 70B Instruct. It shows impressive world modeling, passes the "Is 8.3 > 8.11" litmus test with flying colours, and might be the most powerful LLM currently available to run local on a reasonable PC*. Enjoy the AI Summer while it lasts.

*2x48GB DDR5 as practical limit. 4*48GB is slow, expensive, and counterproductive. Threadripper HEDTs or 256GB/512GB in a second-hand 1U server is definitely not.