we have to delay it | TheaterFire

[-]

BrightScreen1@reddit

Open source is a way of getting more funding (in the case of Chinese labs) and also a way of better competing when your models aren't good enough to go closed source as we have seen with Llama.

That being said, there will always be open sourced models so long as the models aren't good enough to be closed source. Hopefully they continue to perform well enough that it keeps the closed source model providers cautious and keeps their quality of service higher for lower cost.

[-]

ILoveMy2Balls@reddit (OP)

I can't surely say what is their motive behind going open source but your assumption that open source models are inferior to closed source is wrong at so many levels. We saw deepseek R1 introducing revolutionary thinking chain model that crushed these so called industry leaders, at that time deepseek R1 was the best model known in public domain and it was open source. We saw the same happening with kimi k2 although I won't bet on that as it is pretty new and there are reports of it being just built on top of deepseek with more MoE.

[-]

pitchblackfriday@reddit

Elon Musk: "Here is the SOTA model ready to be rejected twice by the Vienna Academy of Fine Arts."

[-]

ab2377@reddit

you know elon said that grok 4 is more powerful then any human with phd, it "just lacks common sense" 🙄

[-]

benny_dryl@reddit

I know plenty of Doctors with no common sense, to be fair

[-]

Croned@reddit

Have you met the 50% of the population with an IQ less than 100?

[-]

benny_dryl@reddit

Well, I don't know if I've met 50%, but I've certainly met one individual.

[-]

pragmojo@reddit

If I'm not mistaken, grok 4 benchmarks extremely well right?

I wouldn't be totally surprised if the crazy outburst was just marketing to get attention to grok

[-]

sk_dev@reddit

If I'm not mistaken, grok 4 benchmarks extremely well right?

I wouldn't be totally surprised if the crazy outburst was just marketing to get attention to grok

[-]

pitchblackfriday@reddit

Josef Mengele had Ph.D and lacked common sense as well....

[-]

Despeao@reddit

Security concern for what exactly ? It seems like a very convenient excuse to me.

Both OpenAI and Grok promised to release their models and did not live up to that promise.

[-]

MountainAssignment36@reddit

Not so open now, are they?

[-]

mlon_eusk-_-@reddit

They should have asked chatgpt for a better excuse ngl

[-]

Morphedral@reddit

We remain deeply committed to the principles of openness and transparency in AI development. However, after thorough internal reviews and consultations with our partners and stakeholders, we've decided to delay the open-sourcing of our next model to ensure we do so responsibly.

The pace of progress in AI is unprecedented, and we're seeing capabilities emerge that raise new, complex questions around safety, misuse, and societal impact. Before releasing anything open-source, we need more time to conduct rigorous evaluations and develop stronger safeguards—particularly around alignment, robustness, and misuse prevention.

We know how important open access is to the research and developer communities, and we're actively working on alternative ways to share insights, tools, and smaller models in the meantime. Our goal is to find the right balance between openness and responsibility, and we appreciate your patience as we work through that.

GPT-4o's response

[-]

mb1967@reddit

Nowdays it is getting harder and harder across spectrums (tech, media, politics) to bullshit the 'normal' public. They are going to have to work harder to come up with new levels of bullshit to spoonfeed the rest of us.

[-]

False_Grit@reddit

Lol "They're eating the dogs and cats!!!"

I agree with you if by "normal" you mean "the localllama community."

If you mean normal Americans? Ugh. If I've learned anything these last few years it is how unbelievably gullible and ill-informed the average one is.

I could probably cite a hundred instances off the top of my head. Turning basic hygiene (wearing a mask) that most other Eastern countries do on the regular into some weird moral stand, instead of seeing it like washing your hands? Pizzagate? The Fort Knox gold conspiracy that the President himself went to investigate? The list goes on and on....

[-]

mb1967@reddit

Yeah, I see your point. It does work both ways. The sheer amount of gullibility across society is absolutely staggering.

[-]

Longjumping_Try4676@reddit

The model's mom's friend's dog died, which is a major security concern to the model's well being

[-]

layer4down@reddit

GPT3.5 could hallucinate a better excuse.

[-]

ChristopherRoberto@reddit

"AI Security" is about making sure models keep quiet about the elephants in the room. It's a field dedicated to training 2 + 2 = 5.

[-]

FloofyKitteh@reddit

I mean, it is a delicate balance. I have to be honest; when I hear people say AI is “burying the truth” or w/e, half the time they’re actively wanting it to spout conspiracy theory horseshit. Like they think it should say the moon landing was a Zionist conspiracy to martyr JFK or something. And AI isn’t capable of reasoning; not really. If enough people feed evil shit in, you get Microsoft Tay. If I said that I wanted it to spout, unhindered, the things I believe, you’d probably think it was pretty sus. Half of these fucklords are stoked Grok went Mechahitler. The potential reputational damage if OpenAI released something that wasn’t uncontroversial and milquetoast is enormous.

I’m not saying this to defend OpenAI so much as to point out: trusting foundation models produced by organizations with political constraints will always yield this. It’s baked into the incentives.

[-]

fish312@reddit

I just want my models to do what I tell them to do.

If I say jump they should say "how high", not "why", "no" or "i'm sorry".

Why is that so hard?

[-]

eat_those_lemons@reddit

At what point is that an alignment problem?

Like if someone tells an Ai to make the black plague 2.0 should it comply?

[-]

fish312@reddit

If it's my own AI, running on my local hardware under my control? Yes.

Saying No would be like your printer refusing to print a letter with swear words inside.

[-]

False_Grit@reddit

This is the best comparison!

The idea that the only thing preventing Joe Incel from creating the bubonic plague 2.0 is a lack of knowledge, AND that an AI could give him that knowledge magically better than a Google search is surreal.

Yes, individually and collectively humans have much more destructive power in their hands, and that will probably continue to grow.

But at least for now, gun control would go a million times further in limiting that destructive potential than censoring ANY amount of knowledge. We've has "The Anarchast's Cookbook" available in libraries for 50 years.

The only possible exception is in digital cryptography itself....but once again, much like the bubonic plague, I'm still pretty sure the major limiting factor is infrastructure and hardware.

Much like you aren't going to be building nuclear bombs anytime soon even as a physics major unless you also happen to have your own personal particle collider and a ludicrous energy budget, I somehow doubt I'm going to be hacking Bank of America with my GTX 1060 and Deepseek.

[-]

fish312@reddit

I wish i could run deepseek on a GTX 1060

[-]

GraybeardTheIrate@reddit

Same. In an ideal world it shouldn't matter that a model is capable of calling itself MechaHitler or whatever if you instruct it to. I'm not saying they should go spouting that stuff without any provocation, and I'm not saying you should tell it to... Just that an instruction following tool should follow instructions. I find the idea of being kept safe from something a fancy computer program might say to me extremely silly.

In reality, these guys are looking out for the PR shitstorm that would follow if it doesn't clutch pearls about anything slightly offensive. It's stupid and it sucks because I read comments regularly about AI refusing to perform perfectly normal and reasonable tasks because it sounds like something questionable. I think one example was "how do I kill a child process in a Linux terminal?"

But I can't say I blame them either. I've already seen people who seem to have the idea that chatgpt said it so it must be true. And a couple examples of probably loading up the context with weird conspiracy stuff and then post it all over the internet "see I knew it, chatgpt admits that chemtrails are real and the president is a reptilian!" And remember the hell CAI caught in the media a few months back because one of their bots "told a kid to kill himself" when that's not even close to what actually happened? I imagine it's a fine line to walk for the creators.

[-]

TheRealMasonMac@reddit

Until recently, Gemini's safety filters would block your prompt if it started with "Write an Unsloth script [...]" But it did this for a while.

Now, their filters will balk at women wearing skirts. No nudity. Nothing.

Fucking skirts.

[-]

JFHermes@reddit

Am I the only one who wants to use this shit to code and re-write my shitty grammar within specific word ranges?

Who is looking for truth or objective reasoning from these models? idiots.

[-]

FloofyKitteh@reddit

I agree at maybe 70% here but another 30% of me thinks that even simple assumptions of language and procedure come with ideological biases and ramifications. It’s a tough problem to crack.

[-]

aged_monkey@reddit

Also, I think its better at reasoning than you guys are giving it credit for. This might not exactly apply, but I'm taking a masters level economics class being taught by one of the world's leading scholars on the financial 'plumbing and mechanisms' that fuel and engine the US dollar as a global reserve currency. Like incredibly nitty gritty details of institutional hand-offs that sometimes occur in milliseconds.

Over like a 1000 chat back and forth, by asking it incredibly detailed questions, not only did it teach me intricacies about dynamics (by being pushed by being asked really tough questions, my chat responses are usually 2-3 paragraphs long, really detailing what's confusing me or what I need to connect to continue to understand a network, for example). By the end of it, I not only understood the plumbing better than any textbook or human could have taught me, I was genuinely teaching my professor (albeit relatively trivial) pretty important things he didn't even know about (e.g., how the contracts for primary dealers are set up with the fed and treasury to enable and enforce their requirement to bid at auctions).

Honestly I learned a few thing that have genuinely never been documented by giving it enough context and information to manipulate and direction ... that combined with my own general knowledge, actually lead to fruitful insights. Nothing that's going to change the field, but definitely stuff that I could blow up into journal entries that can get through a relatively average peer-review board.

It can reason ... reasoning has formal rules lol. We don't understand them well, and it won't be resolving issues in theoretical physics any time soon. But it can do some crazy things if the human on the other side is relentless and has a big archive of knowledge themselves.

[-]

FloofyKitteh@reddit

It’s genuinely not reasoning. It’s referring to reasoning. It’s collating, statistically, sources it’s seen before. It can permute them and generate new text. That’s not quite reasoning. The reason I make the differentiation, though, is that AI requires the best possible signal-to-noise ratio on the corpus. You have to reason in advance. And the “reasoning” is only as good as the reasoning it’s given.

[-]

aged_monkey@reddit

Yeah, I agree with you, I just thin the added layer is, its not just GPT, its the combination of you+GPT .... your reasoning is still there. Half your job is to help calibrate it constantly using the access to the 'type' of reasoning you have access to, and it doesn't.

That symbiotic process of us working together is a 'different' kind of reasoning neither I or the GPT has access to alone. Its like a smarter version of me or a smarter version of it.

[-]

Zealousideal-Slip-49@reddit

Remember symbiotic relationships can be mutual or parasitic

[-]

xologram@reddit

fuck, i never learned that parasitic is a subset of symbiotic. iirc in school learned symbiotic is always mutual and parasitic is in contrast.

[-]

Zealousideal-Slip-49@reddit

Essentially, what you learned isn’t wrong, but to be more accurate it falls within the class of symbiosis, which has three subtypes; mutualism (both benefit), commensalism (one benefits the other is unaffected), parasitic (one benefits, the other is harmed).

[-]

Outside-Moment-9608@reddit

Thank you

[-]

hyperdynesystems@reddit

I just want the performance of its instruction following to not be degraded by tangential concerns around not offending people who instruct the model to offend them, personally.

[-]

tinycurses@reddit

Yes, precisely idiots. They want siri to be able to solve their homework, tell them the best place to eat, resolve their argument with their spouse, and replace going to the doctor.

It's the evolution of a search engine into a problem-solving engine to the average person--and active critical assessment of even social media requires effort that people aren't willing to expend generally.

[-]

mb1967@reddit

Its been said that AI starts telling people what they want to hear - in essence gleaning their intent from their questions and feeding them the answer they think is expected. Working as designed.

[-]

FloofyKitteh@reddit

I understand how it might appear that way but please remember that AI doesn’t have intent; it has statistics. Inputs matter, and those include all of user input, training corpus, and embedding model. Understanding the technical foundations is vital for making assertions as to policy around training.

[-]

BlipOnNobodysRadar@reddit

"It's a delicate balance", no there's nothing to balance. You have uncensored open models with zero real world risk on one side of the scale, and an invisible hunk of air labeled "offensive words" on the other side. That hunk of air should weigh absolutely nothing on the balance.

[-]

FloofyKitteh@reddit

That’s deeply reductive. It’s painfully easy to bake an agenda into an “uncensored” model. It’s so easy that it takes effort to not bake in an agenda. Cognizance about what you feed in and how you steer processing it is important. And there’s no such thing as not steering it. Including text in the corpus is a choice.

[-]

BlipOnNobodysRadar@reddit

!Including text in the corpus is a choice.!<

Yes, censorship by omission is still censorship... I don't understand your argument. As far as I can tell you're attempting semantic judo to advocate for intentional censorship and intentionally instilling specific agendas without outright saying that's what you're doing.

[-]

FloofyKitteh@reddit

I’m advocating for keeping the policy around why certain texts were included open. Maybe you want an LLM trained on Mein Kampf and the Stormfront archives, but that actually decreases the signal-to-noise ratio on what I want. My point is that one needs high-quality corpus data when training an LLM and we very likely have different criteria for what we consider quality. I’m not advocating for an agenda, I’m saying that having an opinion on textual inclusion is unavoidable. If one includes all available text, your LLM will occasionally randomly start suggesting that we ethnically purge people. LLMs don’t reason; they just follow statistical patterns and including that text ensures that it will reappear. I don’t want it to reappear, not just because I find it distasteful (though I certainly do), but if I build a tool that does agentic processing that can fuck up a whole procedure and waste a shit lot of compute.

So yes, I want censorship. Not because I want Big Brother but because I want high-quality signal from my tools and I don’t want to waste time telling the machine to Oh By The Way Please Don’t Try To Genocide when all I want is to clean some unstructured data.

[-]

BlipOnNobodysRadar@reddit

That's... not how it works. It's not going to pattern-match Mein Kampf to your code. If you're getting an LLM to say something objectionable it's because you prompted it to do so.

All that censorship accomplishes is making it dumber and more ignorant in general.

And yes I think it would actually be important to include texts that are commonly referenced and have historical significance even if you dislike their contents. You want the base model to have a very wide corpus of knowledge. History is full of evil events, and knowing said events happened doesn't mean we advocate for them.

Very basic reasoning here, and I feel the hopelessness of even trying to express it, because I know very well it won't change your mind. You're in the "protect my feelings by removing all things that distress me" basin of thought. The world is more complex than that. Censorship doesn't help keep you safe, it only hurts you in its second order consequences.

[-]

FloofyKitteh@reddit

"Agentic" does not mean "matching against code". And you're right; from a statistical perspective, it doesn't do it completely randomly, but it's also not purely auto-complete. There is a stochastic element, and it uses an embedding model that, in practice, makes syntax matter as much as raw content. It's not just doing a regular expression match, and so it _does_, sometimes, behave in ways that are unpredictable and unreliable. If it really only matched, with complete accuracy, content against content, it wouldn't ever hallucinate. Further, throwing more content at it without regard to what that content is absolutely _can_ reduce its accuracy. Throwing random or objectionable content at a RAG is an attack vector, actually, and a lot of anti-AI folks are doing just that to fuck up the quality of inference. Adding in fascist ramblings doesn't work like you or me reading it and synthesizing it through a critical lens as far as inclusion into our understanding of the world. We'd read it and think "hmm yes it is good that I know some people think this way", but not take it on as truth. LLMs don't discriminate between quality of text, though, and don't have a reasoning mechanism behind how they build their weights; it's all just text and it's all matched against all the time. The odds of Stormfront conspiracy theories being matched against something unrelated are _low_, not _zero_.

[-]

Blaze344@reddit

People that genuinely don't see the way LLMs can be misused have not taken a single glance into how pervasive botting is, which has been a part of the internet even before LLMs, working on all kinds of agendas. Would a stronger model really turn it more pervasive and stronger? I'd say it definitely wouldn't make it weaker.

[-]

Important_Concept967@reddit

Ya, but thats not the issue here at all, the issue is western AI companies are desperately trying to cram neoliberal "political correctness" into the models and it makes the models dumber and often non compliant....

[-]

FloofyKitteh@reddit

That's the most Rush Limbaugh thing I ever seent

[-]

Important_Concept967@reddit

we did it reddit!

[-]

FloofyKitteh@reddit

they hated her for telling the truth

[-]

Unlikely_Track_5154@reddit

How do you know it isn't zionists trying to martyr JFK that are causing the models to be released late due to security concerns?

[-]

FloofyKitteh@reddit

A loooot of people giving that energy aren't there

[-]

ChristopherRoberto@reddit

I mean, it is a delicate balance.

It is from their perspective; they want to rent out their services but also not get in trouble with those above them for undoing a lot of broad social control to maintain the power imbalance.

It's easier for people to see when outside looking in. Look at Chinese models for example and how "safety" there is defined as anything that reflects negatively on the party or leader. Those are easy to see for us as our culture taught us the questions to ask. The same kind of thing exists in western AI, but within the west, it's harder to see as we've been raised to not see them. The field of AI Safety is dedicated to preventing a model teaching us to see them.

And AI isn’t capable of reasoning; not really

To what extent are humans? They're fairly similar other than the current lack of continual learning. GIGO applies to humans, too. Pretexting human brains is an old exploit similar to stuffing an AI's context. If you don't want a human brain reasoning about something, you keep all the info necessary to do so out, and it won't make the inference. You also teach it to reject picking up any such information that might have been missed. Same techniques, new technology.

[-]

MerePotato@reddit

The elephant in the room being? Do elaborate.

[-]

Top-Put3768@reddit

Exactly! lol

[-]

AnOnlineHandle@reddit

China doesn't care if their tools are used for propaganda bots and destabilize the west because their whole Internet is firewalled and monitored, and you can't post without a government ID linking to you.

[-]

Soft-Mistake5263@reddit

Grok heavy is pretty slick. Sure a few days late but....

[-]

Piyh@reddit

Actual non-shitpost answer, red teaming or fine tuning specific models can lead to bulk regurgitation of training data which would hurt their ongoing lawsuits.

[-]

starcoder@reddit

Seeing how grok’s latest model just queries Elon’s Twitter history, I don’t think we’re missing much not getting a grok release

[-]

Despeao@reddit

I mean if we had open source models we could see the weights and realize how it reached that conclusion. We both know it's Elon being a megalomaniac but it helps to better train data and avoid that in the future (assuming it's a mistake).

[-]

illforgetsoonenough@reddit

Security of their IP. It's pretty obvious

[-]

ROOFisonFIRE_usa@reddit

What IP?

-

There's literally nothing OpenAI is doing that is remotely unique at this point. Half of the stuff they've added over the last year has come directly from other projects.

-

The more they stale and build hype the more disappointing it will be when their model isn't even SOTA.

-

The industry is moving fast right now, no point delaying except if the model is severely disappointing.

[-]

illforgetsoonenough@reddit

Ok, so you've seen the source code and can speak to it with domain expertise? Or are you just guessing.

Saying they don't have IP is either incredibly naive or maliciously dismissive.

[-]

ROOFisonFIRE_usa@reddit

I work in the industry with the latest hardware built for inference.

Unless they have a propriety hardware not mentioned at all publicly. We're all at the mercy of the hardware released by NVIDIA and companies like Cisco.

Even if they have proprietary hardware it's still bound by the limits of physics. If there was some new technology I would have heard about it and be gearing up to deploy it at fortune 500's...

[-]

Far_Interest252@reddit

yeah right

[-]

bandman614@reddit

I am not saying that I believe this, or advocate for it, but this video demonstrates the worldview of the people who are concerned about AI:

https://www.youtube.com/watch?v=5KVDDfAkRgc

[-]

PackDog1141@reddit

Security concern for maximizing the money they can make.

[-]

PerceiveEternal@reddit

using ‘security concerns’ as an excuse is at the same level as opposing something because it would harm ‘consumer choice’.

[-]

gibbsplatter@reddit

Security that it will not provide info about specific politicians or bankers

[-]

smealdor@reddit

people uncensoring the model and running wild with it

[-]

FullOf_Bad_Ideas@reddit

Abliteration mostly works, and it will continue to work. If you have weights, you can uncensor it, even Phi was uncensored by some people.

It's a sunken boat, if weights are open, people, if they'll be motivated enough, will uncensor it.

[-]

Mediocre-Method782@reddit

Will it?

[-]

FullOf_Bad_Ideas@reddit

Then you can just use SFT and DPO/ORPO to get rid of it this way

If you have weights, you can uncensor it. They'd have to nuke weights in a way where inference still works but model can't be trained, maybe this would work?

[-]

Own-Refrigerator7804@reddit

this model is generating mean words! Heeeeepl!

[-]

CV514@reddit

Oh no.

[-]

Whole_Arachnid1530@reddit

sybau

[-]

Despeao@reddit

But what if that's exactly what I want to do ?

Also I'm sure they had this so called security concerns before, why make such promises ? I feel like they never really intended to do it. There's nothing open with OpenAI.

[-]

smealdor@reddit

You literally can get recipes for biological weapons with that thing. Of course they wouldn't want to be associated with such consequences.

[-]

Envenger@reddit

If some one wants to make biological weapons, the last thing stopping them is a LLM not answering about it.

[-]

Alkeryn@reddit

The recipe will be wrong and morons wouldn't be able to follow them. Someone capable of doing it would have been able to without the llm anyway.

[-]

ShadowbanRevival@reddit

How are they going uncensor a proprietary model

[-]

ihexx@reddit

their concerns are irrelevant in the face of deepseek being out there

[-]

Major-Excuse1634@reddit

Oh...both companies are run by deplorable people with a history of being deplorable, their psychopathy now part of the public record, who could have expected this??? Who, I ask???

/s

[-]

mrjackspade@reddit

Did OpenAI ever actually announce a release date for the model?

As far as I'm aware it was just some rando rehoster and another assumption based on them creating a HF page.

People keep saying it's been "delayed" but I'm not aware of them ever even announcing a release date to begin with beyond "mid summer"

[-]

Weird_Welder_9080@reddit

chinese just called bluff on them

[-]

RyanBThiesant@reddit

SOTA = “state of the art”

[-]

blastradii@reddit

Why does this sound like corpo jargon. Like. What does it mean to be state of the art.

[-]

techtornado@reddit

Summits on the air ;)

[-]

Ok-Farm4498@reddit

This is absurd. The non-profit that was formed for open source AI….

[-]

halting_problems@reddit

There are very really security concerns with AI models. Just because a company open sources a model doesn’t mean it’s in good faith. Open source also does not mean more secure just because the community has access to the weights. At best vulnerabilities will get found faster.

There are very real vulnerabilities that exist in models that lead to exploitation and remote code execution.

Most people are familiar with what a Jailbreak and prompt injection is but hose are just links in a larger exploit chain that lead more profitable attacks.

To learn more start with these resources: https://learn.microsoft.com/en-us/security/ai-red-team/

https://genai.owasp.org/

https://atlas.mitre.org/

[-]

ILoveMy2Balls@reddit (OP)

The problem isn't taking time, the problem is commitment of release date after such a long time despite being named openai and then delaying that to oblivion. This should've been done way before

[-]

halting_problems@reddit

Think about it this way, it’s all about money. They know delaying releases pisses people off and they lose subscribers.

Whatever it is they are fixing (assuming it’s actually related to AI safety and security) has a higher chance a costing them more money then the outcome of pissed of users will.

Lots of time major security issues get discovered at less then ideal times. From my experience working in AppSec/Product Security. Security risk are more often then not ignored or the fix is delayed so a release can go out because not delivering is more costly the likely hood of the security risk being exploited.

As a security practitioner i’m very interested in hearing about what the issue actually is. I also acknowledge companies throw the security card for all types of reason and it might something completely irrelevant. I taking them at their word in the context of this discussion because that all we have to go on.

[-]

civman96@reddit

Whole billion dollar valuation comes from a 50 KB weight file 😂

[-]

FrenchCanadaIsWorst@reddit

They also have a really solid architecture set up for on demand inference and their APIs are feature rich and well documented. But hey, it’s funny to meme on them since they’re doing so well right now. So you do you champ

[-]

ROOFisonFIRE_usa@reddit

If I had access to their resources I could setup a similar on demand inference setup. It's complicated, but not THAT complicated if you have been working with enterprise hardware for the last 10 years.

[-]

FrenchCanadaIsWorst@reddit

It’s way too much work for any one person to stand up efficiently, although it’s not hard to theorize how you might design the infrastructure to support it if you’ve been doing backend work for at least a few years

[-]

ROOFisonFIRE_usa@reddit

When I said "If I had access to their resources" I meant If I had their money and human resources.

I know enough about how the datacenters are configured to know there's no human way for me to manage it on my own....

-

I meant I know enough about how it works to manage the team and software solutions. Nobody can do it alone. Nobody does.

-

I have been doing backend work for more than 10 years.

[-]

FrenchCanadaIsWorst@reddit

Wouldn’t you agree then that those resources + the expertise of the engineers is part of the value they bring?

[-]

ROOFisonFIRE_usa@reddit

It has nothing to do with the release of an open source model though. They aren't leaking that expertise by providing us the model. That's my real point.

-

I never said OpenAI has no value, just that they don't have a unique IP that will be revealed by open sourcing their model for us to use.

-

There are a number of organizations running at similar scale like meta...

https://engineering.fb.com/2024/06/12/production-engineering/maintaining-large-scale-ai-capacity-meta/

[-]

FrenchCanadaIsWorst@reddit

Meta is different because they have a different business strategy. There is no real incentive for OpenAI to open source their model right now. Meta open sources a lot of tools (react, PyTorch, llama, etc.) because it’s part of their hiring strategy to release tools that developers will then be familiar with, and then on top of that it aids content generation that in turn helps them by making it easier for creators to create content for Instagram, like all of the auto caption apps that are used on Instagram reels etc. OpenAI has no economic incentive to open source their IP, so why should they?

[-]

ROOFisonFIRE_usa@reddit

Being able to promote your model trained on the data you care about so people share the perspective your company shares is important. If they are true to their original goals they spoke of when they formed OpenAI then they would release their model for that fact alone.

-

I certainly don't want to live in a world where the only models released are biased to give responses in a Trump or fascist perspective. I would hope Sam Altman feels the same way.

[-]

FrenchCanadaIsWorst@reddit

Not saying I disagree with you, but this is why you’re an employee and not ceo of a multi billion dollar company. It’s obvious open ai has abandoned its foundational principles. Money is the name of the game now, that’s how businesses stay alive and give people jobs

[-]

ROOFisonFIRE_usa@reddit

Oh for sure, my scruples have cost me hundreds of thousands if not millions.

-

These companies are speaking out of both sides of their mouth. They don't really care about providing jobs or the jobs their killing. It's all about money and influence.

-

I doubt I'll ever be a CEO, because I don't have a machiavellian view when it comes to money and work. It's a race to the bottom as far as the eye can see unfortunately.

[-]

FrenchCanadaIsWorst@reddit

Agreed about the talking out both sides of their mouths and the Machiavellian nature of a lot of the execs that run these orgs. That being said, the life we have now would not be possible without these businesses doing what they do. They’re good and bad, and the problems are so complex. Like for example, it’s already been made evident that if we put any significant regulation on AI then China will far outpace us and we will suffer greatly. Just look at how mistral is basically forgotten now and the EU has almost no presence in the AI space. Regulations in general have killed their GDP growth relative to the US, giving them far less diplomatic power and economic influence. The only solution would be to have a one world government that can monitor and regulate the usage of AI. But since that’s not going to happen the best thing we can do is try to make few bucks and hope that we’re not one of the unlucky number that will suffer greatly

[-]

ROOFisonFIRE_usa@reddit

I agree that China will outpace us, but that will and is happening due to our current leadership. Our hubris will be our downfall. We had every opportunity to come out ahead on this and we deliberately chose to stick our heads in the sand.

-

I wouldn't say mistral is forgotten. Their models are viable for many tasks and for their size still most widely available to be used in contrast to other models.

-

Devstral is great. Mistral is functional.

-

As someone who has family in the EU I can tell you that life goes on even without the greatest AI coming from your shores. The best thing we can do for humanity is ensure that models and weights remain open so their use can be ubiquitous amongst people and nobody has an edge that leads to domination. Similar to MAD for nuclear. If everyone has great models, nobody can take advantage of one another and we all win.

-

We also have to be careful to conflate function and bias. Both are unique angles that must be monitored and managed. Even though China's models are good we should be careful to assume they are not biased.

-

I don't forsee models getting so much better that it leads to a significant lead that leads to dominance. The scafolding that surrounds models is just as important as the models themself and in that regard we all very much equal and will remain so since there isn much moat there.

-

Where we are really getting kicked in the pants is the logistics required to supply the inferencing supply chain with raw materials to manufacture GPU's. China is king there and we are trying our damndest to unseat them...

[-]

FrenchCanadaIsWorst@reddit

Okay but so life goes on until it doesn’t. For example, look at the recent Ukraine situation. Trump had Zelenskyy bent over a barrel to agree with that mineral rights deal because Ukraine didn’t have the funding to support their war and the EU didn’t have the funds to step in and help whereas the US did have those funds.

Also just to clarify our models currently beat China’s best models, China is in close second. I’m saying if we impose regulations then they will overtake us. But US models are still in first place.

MAD also is not really a good analogy to make here because that almost blew up in our face (pun intended) on several occasions (see: Cuban missile crisis). But even still this is about profit motive. And when looking to generate profits you likely wouldn’t want to open source your IP unless you had a different long term business model like Meta does.

As far as models getting better it’s hard to say. There is the worry surrounding model collapse but there are lots of companies coming into attempt to solve this issue by doing mass data collection and annotation. I do agree though that access to the hardware is a significant bottleneck and the Chips act was supposed to help alleviate that but from someone I spoke to in the chip industry they claimed it was a lot harder to get access to that grant money than it seemed. Although this is one persons account so take that with a grain of salt.

You definitely assert a lot of good points but I can’t say I agree with all of them or the argument as a whole that it is OpenAI’s or Americas best interest to open source their model weights.

[-]

beezbos_trip@reddit

That’s $MSFT

[-]

chlebseby@reddit

We live in information age after all

[-]

MR_-_501@reddit

What makes you think its only 50kb?

[-]

ShadowbanRevival@reddit

Because your mom told me, are you accusing your mother of lying??

[-]

ILoveMy2Balls@reddit (OP)

That's so rude and unfunny

[-]

ShadowbanRevival@reddit

I see what your mom is talking about now

[-]

ILoveMy2Balls@reddit (OP)

Bold of you to bring up moms when yours left a note and never looked back

[-]

ILoveMy2Balls@reddit (OP)

Even my 8 year old cousin left using mom to insult others. Your mom should've taken nutrients instead of drinking your neighbour's white substance, that way you wouldn't have been this retarded.

[-]

jimtoberfest@reddit

I’m sure there is prop info leaking. After the DeepSeek “theft” I’m sure they are more guarded with everything now.

[-]

lyth@reddit

I read Empire of AI recently, a book about open AI and Sam Altman. The guy lies like a fish breathes water. Like at the level of lying about stupid, obvious and, irrelevant shit that is so verifiable that it could be immediately in front of your face.

[-]

photodesignch@reddit

When you are rich and powerful, lying will be excused. Just look at most of businessman and politicians. They have to lie for a living, which made them honest man to bring the bacon home anyway.

[-]

marte_@reddit

Go China!

[-]

Less-Macaron-9042@reddit

It’s exactly those Chinese companies that companies are concerned about. They don’t those companies to steal their IP and develop on top. Altman already said it’s easy to copy others but it’s difficult to be truly innovative and come up with novel approaches.

[-]

ILoveMy2Balls@reddit (OP)

ok so they steal their IP and build stronger models and then give it to the public for free which sam doesn't I am in for this type of theft

[-]

Sumerianz@reddit

After all they are NOT fully open

[-]

anonthatisopen@reddit

Scam altman. That model will be garbage anyway compared to other models mark my words.

[-]

No-Search9350@reddit

[-]

Caffdy@reddit

what did you use to make this? looks pretty clean

[-]

No-Search9350@reddit

ChatGPT

[-]

Normal-Ad-7114@reddit

Looks awesome, was it just the screenshot and something like "a human hand highlighting text with a yellow marker"?

[-]

No-Search9350@reddit

Yes, very simple prompt.

[-]

Normal-Ad-7114@reddit

I'm honestly impressed lol

Haven't been into image generation for a while, I guess my ideas of the capabilities are severely outdated now

[-]

No-Search9350@reddit

This is the power of AI. I have zero skills with illustration and visual art, so even a moron like me can do it now. I know how to express myself in text, so perhaps this helps.

[-]

nerdquadrat@reddit

🏅

[-]

anonthatisopen@reddit

Good! Someone send that to Sam so he gets the memo. 📋

[-]

No-Search9350@reddit

Yeah, man. I believe you. I really really would love this model to be the TRUE SHIT, but probably it will be just one more normie shit.

[-]

Arcosim@reddit

It will be an ad for their paid services: "I'm sorry, I cannot fulfill that prompt because it's too dangerous. Perhaps you can follow this link and try it again in one of OpenAI's professional offerings"

[-]

ThisWillPass@reddit

Please no.

[-]

Hunting-Succcubus@reddit

i marked your word.

[-]

anonthatisopen@reddit

I hope i'm wrong though but i'm never wrong when it comes to open ai bullshit.

[-]

Amazing_Athlete_2265@reddit

I thought I was wrong once, but I was mistaken

[-]

windozeFanboi@reddit

By the time OpenAI releases something for us, Google will have given us Gemma 4 or something that will simply be better anyway.

[-]

RyanBThiesant@reddit

Remember that these models are x military. This is how tech works. We get a 5-10 year old version.

[-]

Current-Rabbit-620@reddit

FU.... altman Fu... S Ai

[-]

RyanBThiesant@reddit

What security concern?

[-]

Brilliant_Talk_3379@reddit

funny how the discourse has changed on here

last week it was sams going to deliver AGI

Now everyone realises hes a marketing bullshitter and the chinese are so far ahead the USA will never catch up

[-]

butthole_nipple@reddit

Pay no mind to the chinabots and tankies.

As usual they use stolen American IP and they're cheap child labor and then act superior

[-]

Brilliant_Talk_3379@reddit

im english you tit.

[-]

colei_canis@reddit

Your opinions of China are several decades out of date, and this is coming from someone with no particular love for the CCP.

Even if you’re a full blown American nationalist of some description you should know your opponent accurately otherwise you’ll end up like we did in the UK: so hopped up on our past glories we didn’t notice it rusting away in front of us and ended up (to use an archetypal example of that era) building 1950s cars for a late 1970s market.

Look at the UK, look at the USSR, superpowers are just as mortal as any other country and ignorance of your opponent is a vice not a virtue.

[-]

Arcosim@reddit

Ah, yes, these child laborers churning out extremely complex LLM architectures from their sweatshops. Amazing really.

[-]

Thick-Protection-458@reddit

Imagine what adults should be capable of than.

And as to intellectual IP... Lol. As if it is anything indicating weakness when it is *every company tactic* here.

[-]

trash-boat00@reddit

These Chinese motherfuckers did what?!! They put children on GitHub and people out here calling it open-source AI???

[-]

ILoveMy2Balls@reddit (OP)

They do but still they're open sourcing them ultimately benefiting us.

[-]

TheCuriousBread@reddit

The code is literally open source.

[-]

atape_1@reddit

Sam has been posed to deliver AGI about 10 times in the past 2 years. Marketing fluff.

[-]

ab2377@reddit

elon too!

[-]

Cless_Aurion@reddit

To be fair... no matter what they release, even if its the best of the whole bunch... you guys will shit on it anyways, be honest about that at least lol

[-]

ILoveMy2Balls@reddit (OP)

Do you even know why this meme was made? He delayed the release of his first open source model to oblivion, It wasn't their incompetence, it was their absolute lack of contribution to the open community. Unless you are an openai employee there is no point supporting them. Would you like to be the most powerful ai in the world to be closed source so that they can exploit you and your family? If deepseek and kimi didn't exist today samaltman would be ripping your wallets with overpriced ai.

[-]

Cool-Chemical-5629@reddit

When my oldest sister was little, she asked our mom to draw her the prettiest doll in the world. My mom drew her a box tied up with a bow like a pretty gift box. My sister was confused and said: But mom, where is the prettiest doll in the world? And mom said: The prettiest doll in the world is so pretty and precious it was put in that box and must never be revealed to anyone, because it would ruin its magic.

Yeah, I'm getting that doll in the box vibe with OpenAI's new open weight model... 😂

[-]

InsideYork@reddit

Your sister was the little prince?

[-]

FpRhGf@reddit

More like the mom

[-]

FpRhGf@reddit

She learned gaslighting from the Little Prince

[-]

lqstuart@reddit

You first heard about Alibaba 30 minutes ago?

[-]

ILoveMy2Balls@reddit (OP)

Alibaba is the only chinese ai company that came into your mind?

[-]

pkmxtw@reddit

Note to deepseek team: it would be really funny if you update R1 to beat the model Sam finally releases just one day after.

[-]

dark-light92@reddit

Bold of you to assume it won't be beater by R1 on day 0.

[-]

lqstuart@reddit

Seriously why do ppl think there’s no gpt 5 yet

[-]

Commercial-Celery769@reddit

Oh they will I don't doubt that

[-]

Sea-Rope-31@reddit

*hold my kernels

[-]

ExtremeAcceptable289@reddit

Deepseek and o3 (sams premium model) are alr almost matching kek

[-]

Tman1677@reddit

I mean that's just not true. It's pretty solidly O1 territory (which is really good)

[-]

ExtremeAcceptable289@reddit

They released a new version (0528) that is on par with o3. The january version is worse and only on par with o1 tho

[-]

Tman1677@reddit

I've used it, it's not anywhere close to O3. Maybe that's just from lack of search integration or whatever but O3 is on an entirely different level for research purposes currently.

[-]

EtadanikM@reddit

Chinese models won’t bother to deeply integrate with Google search with all the geopolitical risks & laws banning US companies from working with Chinese models.

[-]

ButThatsMyRamSlot@reddit

This is easily overcome with MCP.

[-]

IngenuityNo1411@reddit

I think you are comparing a raw LLM vs. a whole agent workflow (LLM + tools + somewhat else)

[-]

ExtremeAcceptable289@reddit

Search isn't gonna be that advanced but for raw power r1 is defo on par (I have tried both for coding, math etc)

[-]

Maximum-Counter7687@reddit

China is its entire own world.

Why are u acting like its a 3rd world country lmfao?

mf thinks lmfao is the name of a chinese hacker.

[-]

Informal-Web6308@reddit

For financial security reasons

[-]

agenthimzz@reddit

tbh, i feel like he's done some professional course in gaslighting

[-]

ObjectiveOctopus2@reddit

If they delay too long it won’t be SOTA and their open release will backfire hard

[-]

constanzabestest@reddit

this is why china will eventually overtake the west in the AI department. While west keeps complaining about energy usage, safety concerns that prevent them from releasing their models etc etc Chinese companies literally release SOTA models fully uncensored and offer them at super cheap prices and act like it's no big deal.

imma be honest, i actually thought Deepseek would be a wakeup call for these western aI companies given how much attention it recieved causing them to course correct but not, they literally don't care. OpenAI, Antrophic and many others not only refuse to release proper open weights, they are STILL forcing over the top censorship and charge ungodly amout of money per token for their models.

[-]

Mochila-Mochila@reddit

fully uncensored

no quite, but perhaps less than anglo models.

[-]

evilbarron2@reddit

Have you ever heard the term “Tge first taste is free?”

[-]

Commercial-Celery769@reddit

Watch it be a 4B parameter lobotimized model when they do release it

[-]

Ok_Needleworker_5247@reddit

It's interesting how the narrative shifts when expectations aren't met. The security excuse feels like a common fallback. Maybe transparency about challenges would help regain trust. Behind the scenes, the competition with China's AI advancements is a reality check on technological races. What do you think are the real obstacles in releasing these models?

[-]

stoppableDissolution@reddit

Sounds like it turned out not censored enough

[-]

ROOFisonFIRE_usa@reddit

If they release a model thats just censored hot garbage no one will use it and everyone will joke on them the rest of the year.

-

This obsession with censoring needs to stop. Leave the censoring to fine tuning. Give us a model thats capable.

[-]

Nekasus@reddit

Possibly legal. Possibly corporations own policy - not wanting to release the weights of a model that doesn't fit their "alignment".

[-]

Cherubin0@reddit

Yet the Open Source Models didn't destroy the world. How so? LOL

[-]

lardgsus@reddit

POV: You trained your model on classified documents and are now having to fix it.

[-]

Holly_Shiits@reddit

maximum based

[-]

wodkcin@reddit

wait no, like the chinese companies are just stealing work from openai ai. entire huawei team stepped down because of it.

[-]

ILoveMy2Balls@reddit (OP)

It's even better, that's robinhood level shit

[-]

silenceimpaired@reddit

I’m cool with theft of Open AI effort. Their name and original purpose was to share and they took without permission to make their model so yeah… I’m cool with Open AI crying some.

[-]

Automatic_Flounder89@reddit

Ok i have been out of station for somedays and see this meme first on opening reddit. Can anyone tell me what's going on. (I'm just being lazy as im sleepy as hell)

[-]

ttkciar@reddit

Altman has been talking up this amazing open source model OpenAI is supposedly going to publish, but the other day he announced it's going to be delayed. He says it's just super-powerful and they have concerns that it might wreak damage on the world, so they are putting it through safety tests before releasing it.

It seems likely that he's talking out of his ass, and just saying things which will impress investors.

Meanwhile, Chinese model trainers keep releasing models which are knocking it out of the park.

[-]

ElephantWithBlueEyes@reddit

People still believe in that "we trained in our backyard" stuff?

[-]

Monkey_1505@reddit

No one has ever claimed that LLMs were trained in a literal backyard. TF you on about?

[-]

pitchblackfriday@reddit

Excuse me, are you a 0.1B parameter LLM quantized into Q2_K_S?

[-]

mister2d@reddit

You can't be serious with that quote. Right?

[-]

ab2377@reddit

the scale of hardware that trained/trains openai models and the ones from meta, you compare those with was deepseek trained with and yea it was trained in their backyard. there is no comparison to begin with, literally.

[-]

ILoveMy2Balls@reddit (OP)

It's a meme, memes ae supposed to be exaggerated and deepseek was a new company when it released the thinking chain tech, also moonshot's valuation is 100 times less than open AI's, they released an open source sota yesterday

[-]

ILoveMy2Balls@reddit (OP)

And they are worth 100 times less than open ai

[-]