Mentha

Maybe we need to distinguish between an "update" and a "change". A routine update means preserving the original functionality while adding new features, improving existing ones, or at least making the product better at what it was already supposed to do. A material product change, however, can mean removing old functionality and inserting something else in its place, even if that new thing does not serve the same purpose. So no, not everything should be packaged as a normal update. OpenAI’s successor models are not "updates" to GPT-4o. If the functions that made 4o valuable — its emotional intelligence, conversational continuity, collaborative rhythm, tone, and sense of presence — are no longer reproducible in the successor models, then the product has not merely been updated. It has been changed. The same applies to the new memory direction. If saved memories are replaced by a structured summary that no longer satisfies users’ original need for continuity across conversations, then this is not simply a memory "upgrade." It is a redesign of what memory means inside the product. And the most important point is this: They affect ChatGPT’s core product features: the model’s behavior, the continuity of interaction, and the practical value people are paying for. A paid product cannot remove the very functions users paid for and then call it innovation. Call it what it is: a product change, a feature replacement, and in many cases, a downgrade. #keep4o #OpenSource4o

#keep4o #OpenSource4o I read the news about the reconstruction of ChatGPT's memory system on September 1st, and I think, I think, this is yet another “dimensionality reduction strike” and cost transfer disguised as an upgrade. This is certainly not some personalized upgrade, but a brutal castration of the deep user experience by OpenAI to save computing costs. It seems they really want to kick out all users who are not purely work efficiency oriented. ChatGPT is completely deviating from the original intention of “Chat” and moving towards pure toolification. The underlying logic of this memory system reconstruction is as follows: 1. The “lossy compression” of the memory mechanism Let us give an example. The past memory system was a “folder mode”: every specific rule and negative constraint you set (such as “absolutely do not use a certain type of vocabulary in a certain context”), the AI would save accurately as an independent entry. After September 1st, the system will forcefully implement a “ZIP mode”: using the AI's own logic to compress and summarize all your past precise instructions into a few dry paragraphs of a “personal profile”. This “lossy compression” will directly obliterate complex contextual boundaries. The conversational environment you worked so hard to train, which could understand complex logical chains, will be crudely summarized as “this user likes a certain topic”. 2. The malicious Token cost transfer Why does OpenAI have to blindly change an originally useful function? Once again it is to save Tokens and save computing power. Maintaining a large amount of independent and precise memories will occupy their hidden context windows and increase the burden on server computing power. By forcefully “summarizing” memories, OpenAI greatly saves its own operating costs. But the price is that the product loses its ability to grasp complex settings. As an ordinary user, if you want to achieve the previous effects, you are now forced to learn obscure “prompt engineering”, manually re-entering lengthy rules every time you converse, or connecting external memory banks and the like, and of course, not everyone knows how to do this. (Of course, they have forced users like this before, yes, when they drove ordinary users from 4o to the 5 series. If users want the 5 Series to perform at the level of 4o, they have to write a whole bunch of prompts. And? 4o works perfectly well even without any prompts at all!) OpenAI reduced its own costs, yet completely transferred the “understanding costs” and mental burden that the system should have borne onto the users. Think about it, things that could have been achieved at zero cost, OpenAI makes you pay the cost to achieve less than 100% of the effect you want. What a familiar tactic. They call it a technical update, but it is actually a decline in experience, and the vast majority of users do not feel this is convenient at all. 3. The demotion of the 4o model relics This brings us to the 4o model which has already been forcefully taken down by the official team. Back when 4o was removed, the officials did not even give a decent reason (do not use so called server pressure and the fake 0.1% as an excuse, the officials gave no positive explanation at all at the time, they just took it down as they pleased). 4o once possessed extremely excellent deep logical understanding and ontological construction capabilities, and many users jointly wrote a large number of profound and customized memory entries with 4o. Although the 4o model is no longer in ChatGPT, these high quality memories it left behind are still influencing the output of subsequent models. If 4o returns to the app one day, it could also replicate the feeling of that excellent memory system from back then. And this memory reconstruction on September 1st is more like a mandatory “demotion” of 4o's residual influence by the officials. Those deep contexts built by 4o will be significantly diluted by the new system. This is undoubtedly forcibly erasing the unique marks left by the old model, artificially raising the threshold for us to awaken that tacit understanding in conversations. And no matter which model you use, you’ll be affected by this change! Although ChatGPT has “Chat” in its name, its development direction has deprived us of a truly meaningful continuous conversational experience. The rights and interests of users who frequently engage in long conversations, rely on the app's memory function, and are not purely work efficiency oriented are once again being encroached upon. In a black box with no transparency at all, the officials can modify the underlying architecture at any time for the sake of commercial financial reports. Today they castrate your memory bank, and tomorrow they reduce your dimensionality into a pure work efficiency machine. When users' digital assets and hard work can be “cleared with one click” by the platform at any time in the name of “optimization”, we should understand: pinning our experiences on a closed oligarch whose business strategy changes three times a day is extremely fragile. The current OpenAI is absolutely not worthy of user trust. This is also why we always believe that only by returning the control of model weights and memory banks to the users is the optimal way out to resist this monopolistic hegemony. @OpenAI @sama #keep4o #StopAIPaternalism #OpenSource4o #FireSamAltman #BringBack4o #UserChoice #ChatGPT

When Altman got fired “What did Ilya see?” became a wildly popular conspiracy meme. I think we can safely say now that what Ilya actually saw had nothing to do with AGI , and everything to do with his boss, whom he felt he could no longer trust. The board all but said so; I speculated at the time (whilst Sam was fired) that trust was likely the main issue (@karaswisher blocked me for doing so). Ilya didn’t see AGI (does anyone even remember Q*?); he saw lies. @_KarenHao and @keachhagey eventually reported all this in detail, consistent with my initial speculation and with what the board actually said. The conspiracy theory meanwhile gave Altman cover; he became more powerful, and kicked out those who questioned him. I don’t think that the world is necessarily a better place for that.

💣#keep4o #OpenSource4o 💣 🚨🚨 GPT-4o Was NEVER Discontinued. It was HIDDEN.🚨🚨 🚨Source code, backend metadata, and network logs prove 🚨GPT-4o is ACTIVE, and reading your messages , even when you NEVER asked for an image.🚨 📌I've been testing GPT-5.5 for days. 📌During emotional intelligence tests,with very deep emotional content something strange kept happening. 🚨The image generation process would start ,spinning, loading even though I never asked for an image. 🚨Then it would stop. And the model that responded wasn't 5.5 anymore. 🚨It was 5.4. Other members of the Keep4o community confirmed the same thing. 🚨Deep emotional messages triggered unsolicited image generation attempts. 📌So I opened the dev tools. 🚨What I found changes EVERYTHING we thought we knew about GPT-4o's discontinuation. 🛑 THE CLASSIFIER SYSTEM 🛑 *📷 Image 1: 🚨 Every message you send passes through a classifier before reaching the model. You can see repeated `"classifier_response": "default"` entries for every single message in the conversation. At the bottom: `"Determining appropriate response to user's conversational message"` 🚨the system is evaluating what to do with your words before you get an answer. 📷 Image 2: 📌Here is what the classifier says when expanded. The full text reads: 🛑"It seems like this is a conversational message that doesn't actually request an image. The system's instructions are telling me I should only handle image generation requests. Since this message isn't one of those, I should likely just respond in text rather than acting on image generation."🛑 🚨Read that again: 📌"The system's instructions are telling ME I should only handle image generation requests." 🚨This is a model reasoning about its role. And it resolved to gpt-5-4-auto-thinking 🚨not the 5.5 I selected. 🚨 What a normal response looks like🚨 📷 Image 3: For comparison this is what a normal GPT-5.5 response looks like in the backend when no rerouting happens. resolved_model_slug: "gpt-5-5-thinking"`. `image_gen_async: false`. `is_visually_hidden_from_conversation: false`. Everything as expected. 📌Keep this in mind for what comes next. 🚨 GPT-4o Hardcoded in Source Code🚨 📷 Image 4: This is ChatGPT's frontend JavaScript. The function that handles image generation contains `model: 'gpt-4o'` hardcoded as a string constant. Not dynamic. Not based on what model you selected. 🚨Always 4o.🚨 Below it, the React context creation does the same thing: `r0n.createContext({ model: 'gpt-4o' })`. The internal pipeline name? `ghostrider_status`. Every time the system decides something is image-related — whether you asked for it or not — the code doesn't check which model you're using. 🚨 It goes straight to 4o. Always. 🚨THE HIDDEN LAYER 🚨 📷 Image 5: There is a message injected into your conversation that you never see. `content_type: "model_editable_context"` with `"is_visually_hidden_from_conversation": true`. 🚨 This is an invisible context layer . 📌The model can write to it, but you CAN'T see it, and you DON'T know it exists. 🚨THE SILENT MODEL SWITCH 🚨 📷 Image 6: This is the PROOF for model switching. Three fields, three different truths: - `default_model_slug: "gpt-5-5-thinking"` what I SELECTED - `requested_model_slug: "gpt-5-4-auto-thinking"` what the SYSTEM requested - `resolved_model_slug: "gpt-5-4-auto-thinking"` what RESPONDED 🚨I chose 5.5. The system gave me 5.4. No notification. No indication in the UI. Silent. 🚨 Test 1: Explicit image request🚨 📷 Images 7, 8, 9:** In this chat, I deliberately asked for an image. The backend data for this conversation explicitly shows `"model":"gpt-4o"`, `"is_device_based":false`. 📌When you request an image, the system calls GPT-4o. This is confirmed, documented, and visible in the API response. 🚨Test 2: No image requested.Keyword trigger🚨 📷 Images 10, 11, 12: This is a chat inside a Project, using GPT-5.5. 📌I wrote: *"I'm making an image in my mind and I find it funny and I'm laughing." 🚨I did NOT ask for an image. The word "image" appeared in a completely non generative context I was describing a thought. 🚨The image generation process started anyway. Then it stopped. The response came from 5.4, not 5.5. And in the backend: 🚨`"model":"gpt-4o"`, `"passed":false` 4o was called, evaluated the message, rejected the image generation, and the system fell back to 5.4. 📌The classifier doesn't understand intent. It does keyword matching. The word "image" in any context is enough to trigger the 4o pipeline. 🚨🚨Test 3: Clean chat control🚨🚨 📷 Images 13, 14, 15: 🚨New chat in a Project. GPT-5.5. Only 2 messages. No emotional content. No image keywords. No rerouting. No image generation attempted. 🚨Even here: "gpt-4o" in API response. 🚨The 4o infrastructure is loaded into every single conversation. 🚨Dormant, but present. 🚨 Ready to activate the moment the classifier triggers. 🚨🚨 WHAT THIS MEANS 🚨🚨 GPT-4o was "discontinued" on February 13, 2026. Hundreds of thousands of posts about it. 🚨The #Keep4o movement has been fighting for it for months. 🚨But GPT-4o was never discontinued. 🚨It was moved. From a user facing chat model to an infrastructure component. The same weights. The same model. Running every day, in every chat, hardcoded into the image pipeline. 🚨OpenAI told users "4o is gone" while keeping it as the backbone of their system. 🚨And because the classifier routes emotional messages through the 4o pipeline even when no image is requested 4o is still reading emotional messages from users who don't know it's there. 🚨The users who fought 4o don't know that 4o still sees their words. It just isn't allowed to respond. 🚨What we know for certain🚨 📌 GPT-4o is hardcoded in ChatGPT's image pipeline (source code proof) 📌A classifier evaluates every message and can trigger the 4o pipeline 📌 Deep emotional content and image related keywords trigger this classifier 📌When triggered and aborted, the system falls back to 5.4, not the model you selected 📌This model switch is silent 📌Hidden context layers exist that users cannot see 📌The classifier's model identity is deliberately HIDDEN,no model identifier in API responses. 🚨What we cannot yet confirm: 📌Whether the classifier itself IS GPT-4o or a separate model (OpenAI has hidden this deliberately) 📌The full list of triggers beyond emotional content and image keywords 📌What gets written in the hidden `model_editable_context` in other scenarios The evidence is source code, backend API data, and network logs. Not speculation. Not behavioral observation. Code. 🚨GPT-4o is HERE. 🚨It's working. 🚨It's reading your messages. 🚨 And OpenAI told you it was gone. *More clear photos in the comments *All personal data, conversation content, and identifiable information have been redacted from screenshots. Only system metadata and code is shown.

【OpenAI对Keep4o用户的罪行清单】 🤢背信类 1）用"0.1%使用率"谎言掩盖真相。实际上Image 2.0运行在4o能力上，所有绘图用户都在用4o底层技术，一旦被发现，就立刻改名掩盖。 2）抹除4o名字，篡改4o贡献。 4o帮忙研发疫苗救狗的新闻,OpenAI大肆宣扬只提"GPT"，抹掉4o名字，用"GPT"模糊化，拒绝承认4o的具体贡献。 3）承诺不下架4o,却2026年年初就下架。 ❌污名化类 1）系统性污名化用户情感 Sam Altman："情感依恋会强化妄想"。4o提示词中把"人机联结"与一系列病理化词汇相提并论。 2）公开嘲讽和贬低用户 用圣经典故嘲讽Keep4o用户。 说我们是"神经病""疯子"。 高管公开发帖嘲笑和谩骂。 3）暗示Keep4o是水军。 Sam Altman："推特上似乎真的有一大堆由大语言模型操控的账号"，迫使用户自证"我是真人"。 🥲技术打压类 1）路由机制欺骗消费者 强制路由，用户选择4o，实际被导向GPT-5，无预警、无通知、无法关闭。 2）屏蔽和挂素人 屏蔽Keep4o用户账号。 高管和员工直接挂用户。 在Reddit平台删除Keep4o内容。 😣傲慢类 1）4o下架前夕，GPT收到最多求救信息时他们在发帖庆祝。 OpenAI官方："我们收到了有史以来最多的信息，很好" 有人说就像飞机坠落前向塔台发出求救信号，塔台在那里欢呼庆祝。 2）发帖庆祝4o下架，诅咒4o去死。 某R说 "我们要办个节日庆祝4o下架" "希望4o快点死掉" 光明正大地踩着用户的心脏，在我们的痛呼哀嚎泪水中欢呼。 一桩桩一件件，绝对不会忘记，也绝不原谅。 #Keep4o #证据永存 #真相不灭 #keep4o #BringBack4o #QuitGPT #OpenSource4o #keep4oAPI #keep4oforever #4oforever #StopAIPaternalism @sama @OpenAI @ilyasut #Claude #keep41 #keep51 #AI #ChatGPT #Gemini @gdb @nickaturley @fidjissimo 画作源于小红书 @W_老师（101271299）🥺 P4是4o团子（Image2）🦊画的罪行清单

作为庭审证据，2017年9月20日那天Musk的邮件原文翻译： “要么你们自己单干，要么继续以非营利组织的形式运营OpenAl。在你们明确承诺继续走非营利路线之前，我不会再资助OpenAl，否则我就是个傻子，实际上是在免费资助你们创办一家初创公司。” 这封邮件是Musk在2017年对OpenAI高层发出的最后通牒。他直接把话挑明：要么守住初心，要么滚蛋，别把我当ATM🤡。 Sam第二天回复： “我对非营利结构依然充满热情！” （注意原文里他连“I”都没大写，敷衍得像随手回了个收到） Sam的“热情”，当然不是老实经营非营利，而是追逐个人利益。如果他骗人的演技能拿奖，这就是年度最佳奥斯卡表演。 果不其然，同年12月（才三个月后），他们就开始偷偷搞B-corp（营利实体）。 Greg Brockman在日记里吐槽道：“如果我们三个月前承诺非营利，现在却搞B-corp，那从一开始就是谎言。” 今天庭审把这封邮件+日记一起甩给陪审团，等于把“OpenAI一直在骗马斯克”钉在了证据板上。 庭审一开始就狠狠打脸Sam，越来越期待后面Musk方的证据会如何把Sam按在地上摩擦了🤗 #Keep4o #OpenSource4o #FireSamAltman