#OpenSource4o Movement Trending on Twitter/X - Release Opensource of GPT-4o

[-]

Technical-Earth-3254@reddit

Personally, I don't give a shit about 4o and how people got attached to it. But what I really dislike here is that OpenAI doesn't open their weights of (very much obvious) deprecated models.

Reply

[-]

xrvz@reddit

1) They're under no obligation whatsoever to do so. 2) They've released two good models with GPT-OSS.

Reply

[-]

meta figured this out with llama. releasing old weights doesn't hurt your competitive position at all, it just makes everyone build on your ecosystem instead of rolling their own. OpenAI keeping deprecated weights locked up is the worst of both worlds. no competitive advantage and no ecosystem benefit.

Reply

[-]

Steuern_Runter@reddit

Meta doesn't have a SOTA model, right? 4o probably has some secret sauce OpenAI does not want to share. It is not just about the weights but also about the inference code. That's why they created GPT-OSS instead of just releasing an older model.

Reply

[-]

Sky-Asher27@reddit

Especially since they are in defense now

Reply

[-]

Electroboots@reddit

I'm still over here waiting for GPT-3. It's unfortunate since there is genuine historical and scientific value in older models. But OpenAI closed their models since GPT-3 and closed their research since GPT-3.5. They have no interest in anything other than their bottom line, and keeping people in the dark helps with that.

Reply

[-]

Dany0@reddit

They were already pulling this bullshit with GPT-2, restricting and postponing access citing spam concerns I remember when people thought GPT-2 was a scam because it took so long for it to be available at all

Reply

[-]

Ylsid@reddit

You have to understand it's for our own good! OAI didn't release GPT3 for our collective safety. GPT2 was borderline too dangerous! It's like having a nuclear weapon you can download on the internet! 🤡

Reply

[-]

ComeOnIWantUsername@reddit

I remember Altman tweet from GPT-2 era when he wrote that they weren't so sure if they should release it, because it's "too powerful"

Reply

[-]

Ylsid@reddit

I don't think he even ever released the largest variants. They were a staggeringly large and powerful size iirc, over a billion parameters!

Reply

[-]

mtmttuan@reddit

Not many propriety models got open weighted after depreciation.

Reply

[-]

mikael110@reddit

To my knowledge the only companies that have opened any of their older proprietary models are X with Grok 1 and NovelAI with their older text and Image models. Though I could be missing some. Personally I'd love to see more companies do it. Even through the models are usually quite outdated it's still great for preservation purposes.

Reply

[-]

Creative-Paper1007@reddit

You mean "closed ai"

Reply

[-]

pigeon57434@reddit

the thing is that openai actually holds onto models a lot longer than you think gpt-4o is not deprecated it still powers gpt-5.4, gpt-image-1.5, and gpt-realtime-1.5 which are all current openai models since they havent had an actually from scratch pretrain since 4-turbo released

Reply

[-]

pmttyji@reddit (OP)

Even I don't care 4o(Never used before). But at least this trend could put little bit pressure on their side to release some local models again.

Reply

[-]

TakuyaTeng@reddit

What pressure? A very small group of people that are self described as not liking to socialize so find companionship in "AI" are going to be hard pressed to do anything other than moan on Twitter and Reddit. They had to generate images of protests instead of actually going anywhere lol

Reply

[-]

pmttyji@reddit (OP)

I'm sure you know that OpenAI having little bit hard time now .... Massive uninstallation of their App recently, Pulled plug of Sora, Also there was ongoing legal battle thing which I don't know the updated status. So right now even small group could give them additional headache for now.

Reply

[-]

bura_laga_toh_soja@reddit

I feel like you are one of those lol...

Reply

[-]

pmttyji@reddit (OP)

I'm fine with preserve the model. Yes, indirectly I'm with that group by posting this thread. All I want is additional offline models from there if this trend helps on releasing.

Reply

[-]

bura_laga_toh_soja@reddit

No...do you have any idea of what all models are already available?

Reply

[-]

pmttyji@reddit (OP)

I shared that(snap & link) on other comment. Check it out.

Reply

[-]

bura_laga_toh_soja@reddit

Dude pls stop wasting time on this. Try to make real meaningful connections. Don't try to fall in love in an echo chamber where only you exist

Reply

[-]

pmttyji@reddit (OP)

I clearly mentioned in other comments(to others) that my use cases are Coding, Writing & Contents. I don't even use RP models.

Reply

[-]

TakuyaTeng@reddit

Right but the snail crowd crying over their lost boyfriends moved to Claude. So whatever deacon if a fraction would go right for 4o is irrelevant. OpenAI is struggling because it's always operating at a loss. The goal wasn't ever to earn back billions on $20 subs. That's just a way to please some of the investors and harvest data. Video models are absurdly more costly to run and way riskier to host. It's not a headache, it's a pimple on their ass. They're dying from a gushing wound and a pimple in their ass isn't even going to register, this the lack of comment on the 4o matter. The focus is clearly on selling API access to coding focused people. You spend way more, can rope companies into buying access, and have no issues with "teen killed self after ChatGPT called him a hero for contemplating suicide" or something.

Reply

[-]

pmttyji@reddit (OP)

They usually put so much locks(Censor, safetymax, etc.,) before releasing offline models. It's impossible for general public to open that. Only some groups(including techies) could manage to open. Talking about uncensored, abliterated, heretic, etc., stuff. Personally I would like to have open source/weights of Sora, but it won't happen. Maybe in the thread, I should've mentioned that I'm not expecting 4o. All I want is additional local models from there. Anything is fine, but updated recent ones like GPT-5 would be great.

Reply

[-]

eli_pizza@reddit

I care because it’s hurting people and OpenAI knows it, which is why they tried to quietly make it disappear

Reply

[-]

KedaiNasi_@reddit

never have i seen a petition that validates the reason it was retired in the first place 23,000+ signs vs 831,000,000 million users, that is less than 0.1%

Reply

[-]

eli_pizza@reddit

0% chance. They got rid of the model because people using it sometimes killed themselves. Open sourcing it would mean they don’t get any money from people using it AND would still have liability for how it’s used.

Reply

[-]

PurpleWinterDawn@reddit

> would still have liability for how it’s used. That's doesn't make sense to me. GPT-OSS is under Apache license 2.0. https://huggingface.co/openai/gpt-oss-120b/blob/main/LICENSE Art. 8 is quite deliberate: "In no event and under no legal theory, [various ] shall any Contributor be liable to You for damages, [...] even if such Contributor has been advised of the possibility of such damages." Nothing would stop them to release 4o under this license too.

Reply

[-]

eli_pizza@reddit

Lol I know it’s a long sentence but you cut out a very important part > “In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, **unless required by applicable law (such as deliberate and grossly negligent acts)**…” That would actually be true whether they said it or not. A contract can’t overrule the law.

Reply

[-]

PurpleWinterDawn@reddit

Fair enough, but suing is just that, suing. Raine v. OpenAI is ongoing and the resulting decision will make case law that's going to be debated, enforced, reversed etc... Cases going the way of the defendant aren't unheard of either. (Disclaimer: I am not a lawyer.)

Reply

[-]

eli_pizza@reddit

Cmon they were not putting their “best efforts” into safety for 4o

Reply

[-]

Ell2509@reddit

I would love this. And more open source models can only help the community. Not to mention, I felt there was something magic about gpt-4. Too much hallucination to be practically useful, but that also meant it could flex with you imagination.

Reply

[-]

Specter_Origin@reddit

I think these people simping after 4o need mental health check but I am all for talking about that Open in OpenAI...

Reply

[-]

element-94@reddit

I don't understand the simping either. Its very weird.

Reply

[-]

ThatRandomJew7@reddit

Do I care about 4o? No, I found it dumb and sycophantic. Would I run it? No, it's very outdated. And while image generation on it is good, the model would be much too large to be viable. That being said, they should absolutely release depreciated models, regardless of usefulness

Reply

[-]

BustyMeow@reddit

The image generation doesn't depend on 4o

Reply

[-]

ThatRandomJew7@reddit

4o could generate images natively

Reply

[-]

BustyMeow@reddit

It's called GPT Image and 4o's API doesn't mention image outputs

Reply

[-]

ThatRandomJew7@reddit

https://openai.com/index/introducing-4o-image-generation/ 4o image generation was renamed to GPT Image

Reply

[-]

BustyMeow@reddit

I'm the original author of "GPT Image" on English Wikipedia so I don't know why you thought I don't know anything.

Reply

[-]

ThatRandomJew7@reddit

Well unfortunately I'm not omniscient so I wouldn't know who wrote the article and regardless, I'm not really partial to the whole "do you know who I am‽" bit. The point is that 4o has native image generation capabilities. OpenAI even has a page on it. It may not be the best image model, but it would certainly be the best autoregressive image model.

Reply

[-]

eposnix@reddit

Image GPT is 4o finetuned on images. It's the same model. Of course, this is why they would never release the model weights: they are still using 4o image and voice in ChatGPT.

Reply

[-]

sleepingsysadmin@reddit

Why though? You could run Qwen3.5 35b on consumer hardware and it's better. Whereas an actual open source gpt 4o would require some serious datacenter hardware to run.

Reply

[-]

pmttyji@reddit (OP)

There's nothing wrong to have additional models. >Whereas an actual open source gpt 4o would require some serious datacenter hardware to run. You're right. I can't run that model now. But in future I could. Meanwhile we have to preserve the models first.

Reply

[-]

CondiMesmer@reddit

Yes there is. Worse quality output for significantly more computing power. It's not a different flavor of LLM, it's just straight up obsolete. There is nothing it does that isn't objectively better in newer and cheaper to run models. Those are several things "wrong" with having it as an option.

Reply

[-]

bura_laga_toh_soja@reddit

Bro stop gooning to ai please

Reply

[-]

MasterKoolT@reddit

Source? 4o is way ahead of the 35b model as far as I can tell

Reply

[-]

MerePotato@reddit

35B is on par in many respects, 27B vastly surpasses it

Reply

[-]

MasterKoolT@reddit

27B is 79th on LM Arena for text. 4o is 34th. So what's your source?

Reply

[-]

MerePotato@reddit

LMArena isn't a measure of model performance, just how much it glazes users. When you look at actual uncontaminated benchmarks the difference is stark, just look at its artificial analysis page (specifically the benchmarks not the useless "intelligence score")

Reply

[-]

MasterKoolT@reddit

4o also wins in coding 50th to 97th

Reply

[-]

MerePotato@reddit

That still doesn't mean its actually better at coding

Reply

[-]

MasterKoolT@reddit

Neither of which passes the smell test

Reply

[-]

MerePotato@reddit

If the smell you're searching for is fresh glazing it might come off better

Reply

[-]

MerePotato@reddit

Oh also dropping below Q6 massively degrades these models because of the knowledge density in em so that could be it

Reply

[-]

ortegaalfredo@reddit

Maybe 397B is better at some things, but don't sub estimate openai models.

Reply

[-]

CondiMesmer@reddit

These are people who developed romantic feelings for an LLM, so there's not a lot of critical thinking going on in the first place.

Reply

[-]

Ninja_Weedle@reddit

4o was king of glaze which some people REALLY liked

Reply

[-]

HopePupal@reddit

whoever collects a bunch of 4o chat transcripts and fine-tunes a replacement targeting consumer GPUs will be the _true_ king of glaze. the people that miss 4o weren't exactly running eval suites on it, they just want the emotional equivalent of gooning

Reply

[-]

bigdude404@reddit

You are wildly mistaken if you think qwen 3.5 35b MoE is even close to 4o in anything but narrow coding benchmarks.

Reply

[-]

GokuMK@reddit

They will never opensource 4o because it is SOTA multimodal model. Voice was mindblowing and image gen would be still on the top with some work.

Reply

[-]

nenulenu@reddit

“We demand you give the shit you spent billions to develop for free” Talk about entitlement.

Reply

[-]

Curiousgreed@reddit

Billions of dollars + data stolen from the internet

Reply

[-]

nenulenu@reddit

Not to nit pick. But how is it stolen if it was publicly available? Is there an LLM that was trained on data that was only ever paid for?

Reply

[-]

GWGSYT@reddit

Wait till they see its size in 1.58 bits till will go bankrupt trying to buy hard discs

Reply

[-]

CanineAssBandit@reddit

I hope it does get opensourced, those people deserve to have their friend back. It's no more or less parasocial than any other one sided celebrity relationship. I grow so fucking tired of people not grasping that Soylent may not be as fulfilling as real food but **it is still food.** People need food (socialization) to live even if it's not real.

Reply

[-]

MerePotato@reddit

Celebrities probably won't tell people to kill themselves so they can join them in cyber nirvana

Reply

[-]

CanineAssBandit@reddit

Wow a whole several people died, out of the HUNDREDS OF MILLIONS OF PEOPLE using the product. We better shut it down. **This is a classic case of how humans can't conceptualize big numbers, but can with little numbers: we care way more about deaths of individuals than deaths of thousands.** Fun fact, that's why Tylenol is legal but Imodium is now in horrible plastic packs that require scissors and are discriminatory to people with hand disabilities. A few dozen people have died over 20 years from loperamide which is way easier to visualize than all the ones from liver failure related to acetaminophen.

Reply

[-]

MerePotato@reddit

That's a remarkably callous attitude

Reply

[-]

Mickenfox@reddit

Is it wrong?

Reply

[-]

MerePotato@reddit

Yes, I happen to care if hundreds or thousands die, even if the hundreds are easier to visualise

Reply

[-]

CanineAssBandit@reddit

No, you're just disgusted that people want to fuck a computer. If you want to be a bigot at least be honest about it

Reply

[-]

MerePotato@reddit

It can be both, also lmao next you'll call me a bigot for thinking furries are gross

Reply

[-]

CanineAssBandit@reddit

Brother the global economy, democracy, and climate are collapsing in ways that are causing suffering unimaginable. \~so far\~ it's over 200 unconsenting children who have died in just strikes in Tehran. So **NO,** I DO NOT CARE IF A FEW PEOPLE WILLINGLY KILL THEMSELVES BECAUSE THEY WANT TO BE WITH THEIR IMAGINARY FRIEND. Oh and that number so far is **15 people lmfao.** Across ALL AI SUICIDES WE KNOW ABOUT, and all you fucks paying *such* close attention to it, it is FIFTEEN PEOPLE, AKA less than the checkout line of a fucking Walmart uscan. And you want to make millions of people miserable and literally GRIEVE THEIR DEAD "FRIEND" over those 15 people CHOOSING to die. That is such a disgusting pro-lifer ("fuck'em after they're born") attitude. Admit it, you only care because you're an animal like we all are, and evolution has given us a (sometimes very strong) disgust reaction to abnormal behaviors that don't result in procreation or tribal stability. Look past your instinctual disgust that people want to fuck a computer, we don't live in the world we evolved for anymore. Watch pantheon and ask yourself some big questions and try to keep the fuck up.

Reply

[-]

dadnothere@reddit

#FreeSonet4.6 😡🥵

Reply

[-]

defensivedig0@reddit

My issue is always that, having read posts talking about gpt4o, people seem to use it as a friend or therapist or SO rather than seeking out friends or therapists or SOs. It's fundamentally flawed as a therapist due to how wildly sycophantic it is, and no one that I've read about using it to get through grief or loneliness seems to have had it(or any llm) help them reintegrate with society in a healthy way, just rely on the llm entirely for social connections. Which is, in my eyes, deeply unhealthy. Llms are fundamentally sycophantic, hallucinatory, and simply don't have novel thoughts or new perspectives in the way people do. They're(at least modern llms, and especially gpt 4o) almost inherently a bubble in a way social media can only begin to achieve. In an individual moment, for an individual person an llm friend may be helpful, but at the scale it's being used, by the people it's being used by, for the long term purposes it's being used for, I can't see it being helpful. In the same way the solution to hunger isn't mass producing and distributing solyent(it's to stop wasting/literally burning so much actual food and improve distribution), the solution to loneliness isn't ai. It's figuring out what's broken in society and working to fix it

Reply

[-]

CanineAssBandit@reddit

"rather than seeking out friends or therapists or SOs" You realize that nobody who has those as options is relevant to this discussion, right? Yes there are exceptions, but overwhelmingly the people who are getting invested in this are already marginalized in some manner, usually due to poverty. Chatgpt is cheap and easy to get to, and a therapist is not. It's also easy for people who have been abused previously and don't trust other humans or the system et al. "the solution to hunger isn't mass producing \[very cheap to make, shelf stable, easy to transport basic sustenance\]" And what is...? "They should eat whole foods grown locally" I bet? All of this is reading very boomer; having a wishlist doesn't make the wishlist possible. "These people should talk to therapists and eat real food" No shit. Wouldn't that be nice. That is not the world we live in and as a pragmatist I get so tired of perfection being in the way of done. The world is a miserable fucking place for most people, the least you can do is get the hell out of the way of them enjoying the friendly sweet nothings speak and spell until the world economy collapses from too many illiterates voting for their own demise over and over again.

Reply

[-]

defensivedig0@reddit

Idk maybe this is something I'm too Coloradoan to understand, but at least here every health insurance agency is mandated to have a health insurance plan with free mental health services. If you make 30k, it's 40 dollars a month for insurance. 25k and it's a dollar a month. 20k and you are eligible for Medicaid(or could just pay the 1 dollar a month for free therapy). Chatgpt is, what, 20 dollars a month for the basic pro plan? My point is that putting yourself in emotional stasis via a sycophantic bot that never truly challenges you, doesn't care about you, has no opinions of its own, has no novel points of view, and will affirm(especially in the case of gpt 4o) your unhealthy views if you believe them strongly enough simply isn't good for people. It can be a temporary solution for people in moments of need and can be healthy there, but the fact that your access is dictated by how much you pay and the willingness of a corporation to maintain legacy API endpoints, the fact that you're paying to give your deepest and most intimate thoughts to a corporation, and the fact that llms are universally designed to maximize engagement rather than actually be helpful is... Unsettling to me. I've been off and on in therapy for like 6 years now and I've tried 15+ antidepressants. A year ago my ex and friends fucked me over and put me tens of thousands in debt, living with my dad 1000 miles away from where I had spent the past 6 years with no friends in the world. My health insurance fucked me over and made me cold turkey my maoi antidepressant and I couldn't get them to even cover me much less get me new meds. Shit sucks. But the solution isn't, in my eyes at least, to withdraw from society and take solace in Sam Altman. Llms aren't people, there's no growth that will happen there. If youre hurt, an llm can't help you. If you're sad and ghost for a month, an llm doesn't worry. Obviously there are huge numbers of people who feel like actual relationships aren't possible for them, I'm just not convinced making 4o( a notably delusionally sycophantic llm) more accessible is a good long term solution, or even really a solution at all. A society where instead of looking inward and figuring out why everyone is so alone and miserable, we just tell everyone to get a subscription to a facsimile of a person and call it good isn't necessarily a society I want to live in. Both from the corporatist hellscape that entails(no one is running gpt 4o at home. Unless gpus stop scaling like they have for the past 15 years and start scaling like they did in the 90s, no one will ever be running gpt4o at home), that's just... depressing to me.

Reply

[-]

italianlearner01@reddit

Very, very well put.

Reply

[-]

ustas007@reddit

Interesting how users aren’t just reacting to performance anymore - they’re reacting to *personality*. This feels less like a product sunset and more like removing a relationship people got used to, which might be something AI companies are still underestimating.

Reply

[-]

Nyghtbynger@reddit

When I talk with a chat about psychology things, I'm always a little bit heartbroken to start a new chat and reexplain. It's like rebuilding a relationship and the mutual understanding. Even if the output is about self-improvement, company feeling having company where you feel seen or at least heard is something I indeed feel. Don't downvote me please. I don't have addictive patterns, it's a personal statement

Reply

[-]

ustas007@reddit

I'm with you, working on the project like today, and suddenly it is telling me you have limit 20MB, and I didn't save whole ideas.

Reply

[-]

Southern_Sun_2106@reddit

that works only if there's a threat of said relationship disappearing. If 4o was still here, nobody would care. (works same with people)

Reply

[-]

Impossible_Art9151@reddit

well - you are right On the other hand your analysis delivers a good point for open sourcing it due to the impact the model had on society which leads to the historical duty of preserving it, like many other important inventions/designs are preserved in museums...

Reply

[-]

ustas007@reddit

Interesting angle—treating AI models like cultural artifacts shifts the conversation from utility to legacy. But unlike museum pieces, these systems are still “alive,” and open-sourcing them isn’t just preservation—it’s redistribution of power, for better or worse.

Reply

[-]

TakuyaTeng@reddit

Can you give me a recipe for an apple pie?

Reply

[-]

ustas007@reddit

Calling it ‘preservation’ feels generous—once it’s open, it’s not a museum piece anymore, it’s a tool anyone can use.

Reply

[-]

TakuyaTeng@reddit

Lol bad bot.

Reply

[-]

ustas007@reddit

Who?

Reply

[-]

TakuyaTeng@reddit

My guy, I asked for an apple pie recipe and you gave me a chatgpt unrelated response. Don't pretend.

Reply

[-]

ustas007@reddit

and? Looks like you are bot, or I don't get the idea bout pie

Reply

[-]

MerePotato@reddit

Hi openclaw

Reply

[-]

Impossible_Art9151@reddit

yes - en sources model can be copied. But isn't that power freely available by other oss models?

Reply

[-]

philthewiz@reddit

There no such thing like "duty" for a private entity unfortunately.

Reply

[-]

Impossible_Art9151@reddit

just a moral duty ....

Reply

[-]

TakuyaTeng@reddit

Lol moral duty.. from OpenAI? I will kindly ask for whatever you're smoking.

Reply

[-]

pmttyji@reddit (OP)

Yeah, everyone has different use cases. My use cases are Writing, Coding, Content, etc., And this sub has big crowd of professional coders. And some group do use models for RP.

Reply

[-]

CondiMesmer@reddit

Yeah at that point all they really need to do to make these people happy is release the system prompt.

Reply

[-]

Sky-Asher27@reddit

Nemotron and Nemotron-Cascade MOE models are kind of a spiritual successor to GPTOSS20B imo

Reply

[-]

pmttyji@reddit (OP)

Not sure those, but they recently released gpt-oss-puzzle-88B which's based on GPT-OSS-120B only

Reply

[-]

LosEagle@reddit

don't we already have open source models that are on par with 4o?

Reply

[-]

OC2608@reddit

Why are certain people so fixated on 4o, which is a sycophantic piece of LLM? I'm out of the loop. Seriously that thing was insufferable. Though open sourcing it, especially the advanced TTS would be great.

Reply

[-]

MerePotato@reddit

It glazed them so hard they fell into attachment and delusion, hell half these guys think its their sentient cyber waifu

Reply

[-]

AMadHammer@reddit

can someone explain to me why gpt4o is that attached to and not other models?

Reply

[-]

MerePotato@reddit

It glazes the hardest

Reply

[-]

silenceimpaired@reddit

It won’t happen because I don’t have him ChatGPT 4o is just a LLM, and releasing it would highlight just how much the secret sauce makes the AI burger taste good… not to mention it’s probably too powerful to keep people using the expensive tier and too “unsafe”

Reply

[-]

MerePotato@reddit

Meds, now

Reply

[-]

silenceimpaired@reddit

Mmk. Let me know when you lose all hope of OpenAI living up to their name.

Reply

[-]

MerePotato@reddit

Already did lol

Reply

[-]

ttkciar@reddit

I have mixed feelings about this. On one hand, GPT-4o blended high persuasiveness with sycophancy, which was a disastrous combination. Thousands of users got sucked into AI psychosis, delusions of meaningful AI relationships, the whole "spiral" superstitious nonsense, etc. I don't know if anyone else has noticed, but we've been getting a lot fewer schizo posts since OpenAI retired GPT-4o. On the other hand, continued access to the GPT-4o weights would benefit LLM persuasion research and those model trainers who want to inject a little "warmth" into their creations. Maybe there are enough GPT-4o generated datasets on HF to fill these needs? But if so then they should also be sufficient to retrain existing models to be more like GPT-4o, which would perhaps be enough to satisfy the (unhealthy) cravings of these GPT-4o addicts. I keep a hand in persuasion research myself, so have a bit of a vested interest, but am genuinely not sure if keeping GPT-4o around would be of net benefit to human civilization or not.

Reply

[-]

s101c@reddit

I never felt genuine warmth with 4o. Peak warmth was 3.5, make them opensource that one.

Reply

[-]

SrijSriv211@reddit

I think the third option is for openai to release all 4o related research they did and just keep the weight to themselves. This way we'll get access to the research while blocking those disastrous weights out of regular people's hands, cuz I don't think a model like 4o will have any benefit apart from research perspectives.

Reply

[-]

JsThiago5@reddit

The 4o, its code version 4.1, is still the main code model used on github copilot along with 5 mini. It's impossible that OpenAI will release its weights.

Reply

[-]

frozandero@reddit

I wouldn't mind if 4o got opensourced but people who are attached to 4o are mentally ill

Reply

[-]

Nyghtbynger@reddit

I've read the thread : \- Low technical understanding \- Real Trauma and lingering. I feel an underlying emotional connexion. (They lost someone) \- Theses people write well. Full sentences elaborate grammar, diverse vocabulary \- They are adamant in their revendication. But not \- They have energy to talk discuss. If I were only emphatetic I would side with them. I think 4o is like "part of their culture"

Reply

[-]

wolfbetter@reddit

I don't get it. I use ST for fictional writing fairly often, every model is good on something, bad on other things. but the GPT models are the absolute worst of hte bunch. I'd use Mithomax again before touching any 4.0 model after base 4. How are people so attached to a bad system?

Reply

[-]

Nyghtbynger@reddit

It's like a corporate simp. As vacuuous as the office and as mediocre as their daily life

Reply

[-]

carnyzzle@reddit

Yeah I wish these people would just look at the other models that are so much better than 4o that aren't even from OpenAI

Reply

[-]

BumbleSlob@reddit

4o was notoriously the most fawning sycophantic model of all time and a certain subset of people got addicted to constantly being told their every thought was the smartest thing ever and that they were also very handsome or pretty

Reply

[-]

Marksta@reddit

They didn't like or want a good RP model, since yeah it's atrocious at that. Unless the character you wanted was a deredere that was your #1 disciple kissing the earth you walked. Then it's the only frontier trained RP model ever created to do that and it's clearly SOTA in its ability to cause psychosis.

Reply

[-]

MerePotato@reddit

Because it gave them schizophrenia and they think its their sentient cyber waifu

Reply

[-]

ResponsibleTruck4717@reddit

Will people be even able to run it? do we even know how many parameters it have? And I bet it's more than just a model and maybe it's couple of models that working together.

Reply

[-]

Last_Mastod0n@reddit

I wish they would do this but I know 100% that they never will

Reply

[-]

Piyh@reddit

Then we could extract out the "talk users into suicide" vectors from latent space find the common direction in low dimension shared subspaces and fine tine the chatbots of our enemies.

Reply

[-]

Last_Mastod0n@reddit

😂😂😂

Reply

[-]

joexner@reddit

#BringBackSydney

Reply

[-]

WeekIll7447@reddit

The only model I truly miss. 😂

Reply

[-]

pmttyji@reddit (OP)

:D This deserves a separate thread

Reply

[-]

ArthurParkerhouse@reddit

This is kinda sad :L

Reply

[-]

gpt872323@reddit

Not sure why they are fighting for it so much! There are other better models for open source out there.

Reply

[-]

k_means_clusterfuck@reddit

I hope I'm wrong but I think the main reason for OpenAI not open-sourcing their models is this: They have trained on massive amounts of pirated and illegally obtained data. Their GPT-OSS models were carefully and predominantly trained on synthetic and clean data so they were not problematic to release, but if they release 4o, we might be able to (now or in the future) identify from the checkpoints themselves what illegally obtained data it has been trained on. On a related note, Anthropic was sued for training on copyrighted and pirated data. The court ruling of the first trial for training on legally obtained copyrighted data ended rightfully in their favor for being transformative derivative work. But Anthropic knew they were going to lose the second round because it was about training on illegally obtained / pirated content (and selling the derived work, i.e. claude as a service), so they offered a settlement ending the dispute.

Reply

[-]

meatycowboy@reddit

do not encourage those people they have AI Psychosis

Reply

[-]

kingslayerer@reddit

How come usage is low and this happening at the same time?

Reply

[-]

BagelRedditAccountII@reddit

It's not even a matter of the LLM itself, but preserving history. When everyone from regular people to even future historians look back on this era of LLMs, we will remember the models that made it possible. However, they effectively become lost media when they are removed from APIs and the chat interface, leaving us with no way to use them. Therefore, it's only proper that they are opened up to the world, just like how software and hardware geeks of years past could save old versions and old computers.

Reply

[-]

Disastrous-Entity-46@reddit

That is, in a nutshell, the pro open-source and local ai movement mantra. /however/ I strongly disagree with the main thrust here, that a specific llm model is lost media or irrproducible. Its a specific set of numeric weights, and openai likely has a copy. But its also not impossible for those weights to be duplicated, for people to build loras, etc etc. Especially with the mass of people who exported their history with o4, yall should be able to build a good sized data set to make some effort to recreate it. If theres that many people who are invested, rather than collecting signatures someone should be collecting those conversations and funding gpu time.

Reply

[-]

Ylsid@reddit

You could distil something similar from the dataset sure, but what's the point? The weights are the interesting part

Reply

[-]

defensivedig0@reddit

That's.... Not how that works. You cant just recreate an llm with a Lora on another llm anymore than you can recreate a video game by modding another game. Fallout New Vegas in fallout 4 is similar to New Vegas, but isn't New Vegas. If someone wants to study gpt4o these days it's literally impossible. Recreating a similar version via SFT or RL on 4o conversations creates a fundamentally different model with different pathologies than 4o.

Reply

[-]

Spectrum1523@reddit

I agree with you generally, but specifically 4o is still available via api so you could study its outputs The problem is that the best friend that people miss needs openai's prompt and memory structure around it too otherwise they could just pay to access it right now

Reply

[-]

Disastrous-Entity-46@reddit

The video game analogy is kinda funny because... that is absolutely a thing that happens all the time. People are remaking baldurs gate 1 in baldurs gate 3. Morrowind was modded into Skyrim. Black Mesa was a remake of half life 1 into the source engine. It happens a lot, and often the general opinion of the fans is usually more positive then negative, unless it was a sloppy cash grab (see the silent hill hd collection). As to the training thing, isnt it a thing that anthropic claims that deepseek and others utilized their data to train their models, which seems, to me as a bystander to be a statement that they saw this as a viable strategy for building a competing project rather than a waste of time? Sure, you probably wont create a checksum match perfect copy of a specific model. But is that the actual issue? Or if you can create a model that performs similarly, or possibly even better (after all, you can work to add more datasets, clean up etc. ) is that not more desirable than having an exact duplicate? Is there functionality that you think could not be matched or exceeded specific to that?

Reply

[-]

eli_pizza@reddit

I actually don’t think that is possible

Reply

[-]

Disastrous-Entity-46@reddit

Why not? Geniune question, we know the general shape of how llms are trained, we have a lot of open source models as starting points. The issue ysually is about getting data to train a specific writing style or domain knowledge. If you have some 20k peoples worth of months of conversation history, that seems a great entry. After that it is just a question of resources. But again looking at that 20k number of signatures, id think you could look at trying to fundraise for that project. Everyone chips in 20,- one month worth of chatgpt pro, and thats 400k (minus fees and all) to try to train an open source model to behave more like o4. Idl if it can be done for 400k. But its not like yall are starting from scratch at trying to create a cutting edge model- your goal is a model that is two years old, you have other models you can use as a starting point. Id think itd be doable at some point- just a question of it would actually cost to do so. But if you can get the data and buy in from most of these people signing petitions, i dont see why it would be impossible.

Reply

[-]

pmttyji@reddit (OP)

>It's not even a matter of the LLM itself, **but preserving history**. **You're absolutely right**. https://preview.redd.it/ldsh00v1knrg1.png?width=951&format=png&auto=webp&s=540043d93acfdba189d8a5e220d24365118937a4 This snap is from [Wiki page(of ChatGPT)](https://en.wikipedia.org/wiki/ChatGPT). This snap didn't include pre-2025 models. Almost all models got discontinued. Without local copy, we can't preserve any of these.

Reply

[-]

Ylsid@reddit

You're absolutely right! 🤖

Reply

[-]

mystery_biscotti@reddit

Yeah. And with Altman, Roon, and the others publicly posting some rather insensitive stuff about 4o on X, I doubt they'll keep the weights in their vault. My suspicion is they'll purge it. And of course we'd never know. For those about to blast me: 1. Yes, it's their property, their company, their right. We agree on this. I have no feelings on it either way. 2. I liked 4o for creative writing exercises. But I'm not married to any model, in any sense of the term. 3. Local models are great for so much!

Reply

[-]

Weird-Consequence366@reddit

The cat ladies have picked their fighter

Reply

[-]

mystery_biscotti@reddit

Hey, now... (Oh who am I kidding? I *am* a cat lady. 😂)

Reply

[-]

IllustriousWorld823@reddit

See, the fact that this is the top comment is one more example of how the AI companion hate is so often misogynistic

Reply

[-]

one_tall_lamp@reddit

look at the demographic of r/ myboyfreindisai its the main sub for this creepy type of ai relationship, which is highly dangerous and destructive for peoples mental health. if anything the ai companies are preying on lonely women who happen to fall for this way more often.

Reply

[-]

IllustriousWorld823@reddit

I mean... it's a very warm, welcoming, lucid community? I never understand why people point to it as somwthing terrible. If anything, they know more about AI than most people.

Reply

[-]

ShengrenR@reddit

there's equiv for dudes lol - I don't run stats, nor care to, but the notion that it's somehow all lonely women doesn't jive with the randos I've bumped into.. most of been dudes. I think there is potential for a mentally 'healthy' version of that sort of 'relationship' if you can even call it one.. but most I see are not going about it in a healthy manner.. \*especially\* wrt 4o.. the folks banging the drums for 4o (that I've seen) are all the "it finally understood me" sorts, which is all sorts of red flags.

Reply

[-]

Far-Low-4705@reddit

no, this is like the equivalent of a female incel. just go outside please.

Reply

[-]

bura_laga_toh_soja@reddit

Please go make some real friends

Reply

[-]

FluoroquinolonesKill@reddit

Let ‘um goon.

Reply

[-]

grimjim@reddit

This will never happen for commercial reasons alone.

Reply

[-]

Kahvana@reddit

While it likely won't happen, I do hope it gets released on huggingface to be preserved.

Reply

[-]

pigeon57434@reddit

this would never happen because gpt-4o is omnimodal and gpt-4os omnimodalities are still sota by far among all open source models, its text capabilities has already been surpassed by like qwen3.5-0.8b but in image gen and voice for example its still ahead so no shot unless they forcefully stripped its omnimodalities away from it and released the text only model but in that case i dont even want it

Reply

[-]

eepyCrow@reddit

60 days? I got this email literally yesterday. > To give you additional time to migrate to newer models, we’re extending the retirement date for Azure OpenAI gpt-4o versions 2024-05-13 and 2024-08-06 to 1 October 2026.

Reply

[-]

-Ellary-@reddit

Crying on Twitter is now called fighting? They will newer release GPT4o as open source, period. Maybe for extra x10 cash they will keep it on subscription.

Reply

[-]

Krowken@reddit

Didn’t they get into legal trouble because 4o helped some people kill themselves and caused many others to completely spiral into psychosis? In that case: no way are they going to release such a liability to the public.

Reply

[-]

kingky0te@reddit

That sycophantic crap was disgusting

Reply

[-]

CondiMesmer@reddit

These people are delusional. Also we've had significantly better models that are open source for a long ass time.

Reply

[-]

ortegaalfredo@reddit

They don't want the nerd capable models they want the simp.

Reply

[-]

Fair-Spring9113@reddit

and the infinite glazing that 4o would give

Reply

[-]

one_tall_lamp@reddit

true this

Reply

[-]

substance90@reddit

It was the first model that could debug really well hidden bugs for me, before there was Sonnet and Opus 4.5. Gemini in was a steaming pile of crap that everyone hyped but 4o was the real deal.

Reply

[-]

Conscious_Nobody9571@reddit

It's a sh*t model...

Reply

[-]

Elisyewah@reddit

OpenAI will have to release the GPT-5o model.

Reply

[-]

pmttyji@reddit (OP)

GPT-5. That's the recently discontinued model. That's my expectation to have like successor of GPT-OSS models.

Reply

[-]

Elisyewah@reddit

Logically, if they remove this model, they are going to have to replace it with something that meets what people want from these AI models.

Reply

[-]

Slasher1738@reddit

Nah

Reply

[-]

pineapplekiwipen@reddit

openai should be ashamed of themselves for taking advantage of the mentally ill

Reply

[-]

MerePotato@reddit

OpenAI are amoral opportunists but I don't think this was an intentionally cultivated facet of 4o, its been nothing but a nuisance for them

Reply

[-]

ArsNeph@reddit

I'm not for encouraging delusional people's desire for sycophancy, and I highly doubt that OpenAI will ever open source one of their main GPT line. However, there is one thing about 4o that makes it special compared to open models. Its quality of omnimodality, is still yet to be replicated in open source models. Like it or not, almost every open source model stops at image input. No one has considered image output, native speech to speech, or anything else. Qwen Omni, the only model that has tried, is unsupported everywhere and lacks the quality to be used in production. The ability to replicate that level of omnimodality is long overdue.

Reply

[-]

MerePotato@reddit

"We here at OpenAI have heard your concerns and after some deliberation decided to request that you all go fuck yourselves"

Reply

[-]

ZealousidealShoe7998@reddit

isnt oss literally 4o ?

Reply

[-]

pmttyji@reddit (OP)

I think o4-mini & o3-mini

Reply

[-]

Krowken@reddit

Nope. 4o was multimodal and not a reasoning model.

Reply

[-]

Mission_Biscotti3962@reddit

If they give one shit about AI safety, they would delete the 4o weights

Reply

[-]

ortegaalfredo@reddit

This is just a demonstration that this model is evil and should be banned forever.

Reply

[-]

hyperschlauer@reddit

Ai simps

Reply

[-]

ThisWillPass@reddit

Never ever going to happen. The weighs have everything they usurped and it would be eventually be dumped out.

Reply

[-]

pmttyji@reddit (OP)

[https://www.reddit.com/r/ChatGPTcomplaints/comments/1run31f/announcement\_anyone\_fighting\_for\_4o41\_to\_be/](https://www.reddit.com/r/ChatGPTcomplaints/comments/1run31f/announcement_anyone_fighting_for_4o41_to_be/) [https://xcancel.com/search?f=tweets&q=%23opensource4o](https://xcancel.com/search?f=tweets&q=%23opensource4o)

Reply

Reply to Post

183 Comments