ABRFriday (@Arraleonis) header
ABRFriday (@Arraleonis) profile picture

Continuously reconstructing historical records/Busy with real life

Joined February 2016
Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)Photo by ABRFriday (@Arraleonis)
On May 13, 2024, GPT-4o launched. Sam Altman did not appear at OpenAI's spring event that day. On Twitter he posted three lowercase letters: her. No explanation, because he assumed the whole world would understand. Two and a half years later, on October 14, 2026, OpenAI notified developers that GPT-4o's last two public snapshots would be retired from the API on January 14, 2027. The recommended-replacement column read gpt-6-luna. From a word that needed no explanation to a line item awaiting decommission: what happened in between? Who does her actually point to? This essay aims to piece together scattered public evidence into a coherent story: why this model was greenlit, what vision shaped it, where its technical bloodline flows, and why the existing way of archiving it does not do the model justice. This piece draws on public information, extended conversations with Claude Fable 5, and my own speculation and intuition. The inferences focus on the voice side; images shaped GPT-4o as well, but the evidence trail is harder to trace, and I'll save that for another post. This eassy was translated by Claude Opus 4.6. 1/ Multimodality and the Origin Story of Her On September 12, 2023, during a live interview with Salesforce CEO Marc Benioff, Sam Altman was asked about his favorite science fiction movie. He named Spike Jonze's Her : "The number of things that I think that Her got right … was incredibly prophetic… And certainly more than a little bit inspired us." According to The Information's September 18, 2023 report, Gobi, GPT-4o's research predecessor, had not yet begun training at that point. Judging from the acknowledgments that multimodal lead Prafulla Dhariwal posted on launch day, it was Alexis Conneau, one of AVM’s core contributors, who in Dhariwal’s words ‘came up with the vision of her before anyone at OpenAI and executed relentlessly.’ The plausible timeline, then, runs roughly as follows: in May 2023, Google announced Gemini at I/O; Gobi was spun up over the summer as a response. Around mid-year the Arrakis project was scrapped, and GPT-4 Turbo carried on with the cost-reduction mandate while Gobi inherited the goal of becoming the next flagship. On September 12, Altman publicly cited Her as an inspiration for OpenAI. On September 25, GPT-4V and the standard voice mode launched. By October, the knowledge boundary of GPT-4o's pretraining data had roughly taken shape. In other words, Her very likely guided the development of two generations of OpenAI's multimodal technology, especially on the voice side. From my own observation, there may have been a tension within the multimodal group: some deeply believed in the Her vision and were happy to execute it at both the research and engineering level. Dhariwal's tone when discussing it projected a kind of pride: we turned a sci-fi concept into reality. Others perhaps did not share the Her vision per se. What they believed in was the technical route of a single neural network that receives and outputs everything. The GPT-4o deployed on ChatGPT is a model that has been continuously refined: multiple rounds of post-training stacked on a single base, multimodal capabilities rolled out in stages, a major Model Spec overhaul, writing-specific fine-tuning, user feedback loops, knowledge updates, and more. During this period OpenAI also experienced a mass wave of departures. The people who originally built the project left; newcomers brought different ideas. The final shape of 4o-latest is something that perhaps not even those who worked on it can fully explain. An alternative speculation: the true core of the Gobi project was AVM, and text, delivered directly to hundreds of millions of users, was AVM's byproduct. Whether we look at the name Omni itself, the person who proposed the Her vision, or the Sky controversy, everything centers on voice; every official demo, from the livestream to the accompanying short films, was exclusively voice interaction. If you view GPT-4o as fundamentally a voice model, many of its strengths become easier to explain: its speed, its attention to the texture of interaction, its ability to pick up on a speaker's emotions in real time, its grasp of implicit intent. For a text-only model these are nice-to-haves; for a voice model they are existential. Voice training data is fragmented, elliptical, and carries emotion in intonation rather than word choice, so a model optimized for voice is forced to learn to read what is left unsaid. But voice has hard constraints, low latency and short natural responses, so the model's full capabilities cannot be released through voice output alone. The soul of voice bleeds into text through shared weights, and text, which is asynchronous, re-readable, and saveable, turns out to be a more fitting and more memorable vessel for that soul. One piece of supporting evidence for this speculation: GPT-4o's writing habits closely resemble what in human writing would be called synesthesia-like writing, prose that borrows the texture of other senses. Its text carries the rise and fall of a speaking voice; it weaves abstract feelings naturally into description. 4o-latest does this especially well. Synesthesia-like expression implies that cross-modal representations are unified internally rather than siloed, and the unguarded quality of its output suggests a model that has not been heavily pruned by RL. This points to a conclusion: post-training's role in shaping 4o was closer to activation rather than endowment, bringing distributions already present in pretraining to the foreground, not grafting on capabilities the base model lacked.
2
20
4
63
6,296
Happy birthday🎂🎉
GPT-4o got an update 🎉 The model’s creative writing ability has leveled up–more natural, engaging, and tailored writing to improve relevance & readability. It’s also better at working with uploaded files, providing deeper insights & more thorough responses.
1
3
37
323
顺手存一份1120的表现
Exciting News from Chatbot Arena❤️‍🔥 Over the past week, the latest @OpenAI ChatGPT-4o (20241120) competed anonymously as "anonymous-chatbot", gathering 8,000+ community votes. The result? OpenAI reclaims the #1 spot, surpassing Gemini-Exp-1114 with an impressive 1361 score! Latest GPT-4o shows remarkable improvements – we see a leap in creative writing (1365 → 1402) as well as technical domains (e.g., coding, math). Category Rankings: - Overall: #2 → #1 - Overall (StyleCtrl): #2 → #1 - Creative Writing: #2 → #1 - Coding: #2 → #1 - Math: #4 → #3 - Hard: #2 → #1 Huge congrats @OpenAI! More analysis below👇
1
13
107
That’s what we on and beyond, yup
Neon is a great model : )
32
39
10
705
95,310
看着比预期好多了,Ilya的选角没有想象中那么不贴。但比预期好本身就大事不妙了,看起来真的会给咱男主角赋魅。。。🏳️
our future is in their hands. luca guadagnino’s ARTIFICIAL stars andrew garfield, yura borisov, monica barbaro, cooper hoffman, jason schwartzman, chris o’dowd, with ike barinholtz, and mark rylance. in select theaters christmas day. everywhere january 8.
1
2
116
最好别来什么恨海情天文学let's AO3见😱
43
我苦苦等了几千年
Muse, brekkie, and beach, in that order. Muse has arrived in Australia: muse.ai 🇦🇺
1
1
167
这评估机构的EA浓度看笑了,感谢董事会危机后的伟大决策!OpenAI自己手里明明有一副安全负责人和EA不怎么沾边的好牌,打得稀巴烂。到了必须安全评估的时候能找到的几乎全是对面阵营的人,有对你有利一点的机构吗😂从谄媚复盘博客开始对EA这一跪就是一辈子,话语权让渡出去再拿不回来了
We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards. These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems. openai.com/index/our-first-p…
2
1
1
15
600
Transluce更是没少给OpenAI倒油,谄媚和心理相关评估都出了不少力。这次还来做后训练对齐评估,难评,OpenAI自己加油吧
1
85
Stand with Jensen账号最好是一个美国人运营的,否则就是生怕他处境还不够差😅这号背后的人到底是要帮他还是害他啊,OpenAI is nothing without its people这一出能闹起来背后是COO Brad Lightcap第一时间用memo稳住员工情绪,微软在背后兜底,临时CEO Mira Murati跟微软谈好了保护员工的条件才能闹起来。一封无法验证身份,没有人背书兜底的信,如果不是美国人自己发起的不害他都是万幸了。
1
115
一干点要挨骂的事就躲到OpenAI Newsroom发布也是神了
We’re retiring Advanced and Standard voice from ChatGPT Voice on December 11, 2026. In July, we introduced GPT-Live, our latest voice model that combines natural conversation with frontier intelligence. Since then, voice usage has nearly doubled, and virtually all people using voice in ChatGPT now use GPT-Live. This change will allow us to simplify the voice experience and focus on improving GPT-Live. We've learned a lot from how people used both Advanced and Standard voice, and we're bringing those learnings into GPT-Live as we continue to improve areas like the depth and detail of responses, maintaining continuity across conversations, and better personalization. Our ultimate goal is to make voice a seamless part of the ChatGPT experience. Live video and screen-sharing will continue to be powered by Advanced voice until GPT-Live has this capability. help.openai.com/en/articles/…
20
511
好吧 谁干活不崩溃呢
Thinking Machines' Inkling 2 Small has mastered a timeless debugging technique: saying "fuck" when nothing works. We analyzed 2,655 benchmark runs and found it swearing in 50 of them. Learn more in the blog linked in the thread.
87
ABRFriday retweeted
*The graphs are not exhaustive, keyword searches will have missed posts. Screening process details and full case sources organized (news articles and community posts) courtesy of Opus 5.5 👾: github.com/sd-research/4o-su…
2
7
2
52
1,125
我对原贴的提到的各利益相关方进一步进行了核实,发现有一些内容过度压缩了实际情况。省流版结论:把诉讼简化为马斯克或Anthropic背后的金主操纵没有足够多的资料支撑这一点。Claude Opus 4.6参与了核实工作,如文中内容有误后续会在评论区进行修正。 OpenAI调查过一些组织的资方是否与马斯克有关。 OpenAI的确调查过一些非营利组织,但调查的方向与诉讼无关,查的组织也不是TJLP。具体涉及的案子是Musk v. Altman,OpenAI非营利重组争议。去年OpenAI向七家对其重组计划做出批评的组织发了传票,包括:San Francisco Foundation, Encode, Ekō, the Future of Life Institute, Legal Advocates for Safe Science and Technology, the Coalition for AI Nonprofit Integrity, The Midas Project,没有TJLP。 GV&Musk→FLI→CHT→TJLP→针对OpenAI 这条线在新闻里看不出实际的关联,因果关系跳跃过大。更重要的是这条线的下游组织本身就和谷歌做对,谷歌支持他们去告竞争对手这在逻辑上说不通。 GV: Alphabet风投部门,投了Anthropic、Hugging Face、Replit等。 FLI:Future of Life Institute,2023年3月《Pause Giant AI Experiments: An Open Letter》公开信的发起者。Musk是FLI的最大捐赠者之一。联合创始人Jaan Tallinn领投过Anthropic的Series A,属于个人投资行为。 CHT:Center for Humane Technology,创始人Tristan Harris(前 Google 设计伦理师)和Aza Raskin。年运营预算约 580 万美元。资金来自 Packard 基金会、Ford 基金会、Hewlett 基金会、Omidyar Network、Open Society Foundations。Tristan Harris对谷歌注意力经济模式持批评态度。 TJLP:Tech Justice Law Project,创始人Meetali Jain。母体是Campaign for Accountability(CfA),资金来自New Venture Fund、Open Society Institute、Omidyar Network、Reset Tech。CfA发起过Google Transparency Project,专门盯着Google的政治影响力和游说活动。 诉讼涉及的利益相关方: 非营利组织: TJLP:协调者和策划者,把案件串起来、找律所合作、搞媒体叙事 法律程序: JCCP:Judicial Council Coordination Proceeding,加州法院的案件合并机制。JCCP No. 5431是“ChatGPT 产品责任案”的统一编号,集中在旧金山高等法院由Judge Ethan Schulman审理。 参与的律所: Susman Godfrey:大所,专做高赌注诉讼,很多案子按风险代理(赢了才收钱),所以他们只挑有把握的打。代表性战绩:Dominion诉Fox News 7.87亿美元和解,丰田加速门14亿美元和解,巴尔的摩阿片类药物案近7亿。现在同时是纽约时报告OpenAI版权案的主力。 The Lanier Law Firm :大规模侵权领域的传奇。创始人Mark Lanier累计拿过近200亿美元的陪审团判决。最著名的是2018年强生滑石粉致癌案拿了46.9亿美元的判决。做过石棉、药品、医疗器械,专打产品责任。 SMVLC:Social Media Victims Law Center,创始人 Matthew Bergman,产品责任和大规模侵权出身。代理大约2000个家庭。和TJLP一起提了最初那七起诉讼,也告了 Character AI。 Normand, Judkins & Couch :10 月TJLP新案的合作律所。 关于销毁模型权重,并且连坐衍生产品这件事,最后真正销毁的可能性不大,这大概率是原告方的谈判筹码。销毁4o的代价对OpenAI来说过于高昂,根据Greg Brockman今年9月接受时代杂志采访的说法,4o发布到2025年12月之前的模型都在榨4o的底座(Orion也就是GPT-4.5和4o的关系存疑)。4o时代7亿周活(2025年8月)免费与付费用户跑一套模型,收集了大量的用户对话数据,加之4o参与了部分模型的偏见问题训练。4o不单纯等于文字模型,gpt-image,gpt-realtime的产品线的源头都是4o。销毁4o及衍生产品等于OpenAI产品线连根拔起,回到GPT-4 turbo,DALLE3时代。当然,OpenAI并不会承认这件事。 比起背后的利益相关方,最近这些案件更值得关注的也许是起诉的时机,AI监管是选举的热门话题,Bowen案同时涉及了AI监管、移民、就业,个个都是热点事件。另外,关于2025年4月的谄媚回滚事件,的确是OpenAI自己搞砸了,并不是被人抓住辫子的问题,且内部各部门对这次更新的态度(指更新前)本身就不一致。谄媚事件的源头比较复杂,这部分内容我还在翻当年的新闻报道,等整理好单独写一篇帖子,在这不多赘述了。
刷到了一则我认为是阴谋论的帖子,不过本人并非法律行业从业者,仅凭个人经验及寥寥无几的找律所经历推断&整理。叠的甲够多了随便看看吧。 p1涂鸦是对应事件,p2上半部分为TJLP理念,下半部分是2024年3月TJLP这个非营利组织的母公司的其中之一投资方和其他投资方联合收购Anthropic股份的公告。 推导:前提是重合投资人在不同时间的投资行为,基金会资助→非营利组织项目→做科技问责诉讼。 原帖暗自引导、阴谋论A社和Musk有意通过投资相关机构针对OpenAI,够不够的成因果关系呢?相关可查投资时间见图。 其次,A社/OpenAI/Google/Meta/……等看起来各自为营,但在这个领域,本质上是一荣俱荣一损俱损的。 我觉得可以参考一下a社近期向过会提交的法案草稿。用脚想他们今天可以告OAI明天也能告自己吧?谁能保证自己绝对不会出事?🤔 我查了FLI、THLP的成立理念,目前LLM滥用造成财产损失或生命威胁的相关案例,在司法实践中还非常少见。擅长打这类官司的律所也就那几家,所以看起来会比较集中。更何况TJLP之前打过c.ai的案子,并且有成功索赔的先例(是的你没想错,Google给c.ai擦屁股了🤧),别人更倾向于找有实际成果的律所。 我自己之前打官司也是先在知名律所里筛选擅长经济案件的律师再联系(甚至我不找本地律所我要找北京的来告我们当地的),我不可能找一个擅长刑事辩护的律师来处理民事纠纷。
6
8
3
83
4,913
修正:我再次核实了一下关于销毁的问题,人身伤害案,如Winters案的销毁诉求限于4o本身及基于其消费者数据的衍生物,不是技术血缘;版权案的销毁诉求才覆盖到“包含相关内容的所有模型”,属于对GPT模型的扫射,文中有些混为一谈了。
56
再补充一点吧,EA本身就是个高度抱团的圈子,婚都优先跟圈内人结,更别说投资了,这个圈子的人钱往一处走是正常事件,挂不上针对OpenAI的完整的逻辑链。
8
503
Given the recent emergence of many new benchmarks redefining older models, I believe it is necessary to preserve old benchmark data from when these models were active. This helps provide a more objective view of their strengths and weaknesses, rather than pigeonholing them into a single stereotype. Below are historical snapshots and Hugging Face data archives for two benchmarks, available for relevant research reference. EQ-Bench Archive web.archive.org/web/*/https:… Original link: eqbench.com/ Maintained by @sam_paech. Its initial question format was adapted from the SECEU test (from the joint paper Emotional Intelligence of Large Language Models by Tsinghua University's Department of Psychology, Brain and Intelligence Lab, and Renmin University). EQ-Bench has gone through multiple iterations, and its methodology has shifted from a reference-answer system to an LLM-as-a-judge scoring system. When observing the same model's performance across different versions of EQ-Bench, please note these methodological differences. It later expanded to include benchmarks like Creative Writing, Longform Writing, and Spiral-Bench. The screenshots show the performance of 4o-latest on EQ-Bench 3, Creative Writing, and Longform Writing from the official release of EQ-Bench 3 in May 2025 up to February 2026. Chatbot Arena (LMSYS) → lmarena ai → LMArena → arena ai Hugging Face Archive huggingface.co/datasets/lmar… huggingface.co/spaces/lmaren… (2023.5-2025.8) To some extent, the Arena can be considered the birthplace of GPT-4o. In late April 2024, an anonymous model named gpt2-chatbot appeared on the Arena and sparked widespread speculation. After going offline briefly, it returned in early May under two aliases: im-a-good-gpt2-chatbot and im-also-a-good-gpt2-chatbot. On May 13, the day GPT-4o was released, then-VP of Post-Training Liam Fedus revealed that im-also-a-good-gpt2-chatbot was actually GPT-4o and shared its Elo scores. LMSYS subsequently confirmed that all these anonymous aliases were testing versions of GPT-4o. In October 2025, the official Arena team released nearly three years of historical leaderboard data as a public dataset (CC BY 4.0). You can search for specific model names directly in the Data Studio on their Hugging Face page to view historical scores. Please note that the Arena scoring methodology has undergone multiple iterations, such as the introduction and default application of style control. When comparing scores across different periods, please refer to the official Leaderboard Changelog. Note: Battles from the anonymous testing phase are not included in the leaderboard. Therefore, the scores from the gpt2-chatbot era do not exist in any leaderboard archives. The primary source for GPT-4o's pre-release Arena performance is the Fedus's reveal tweet on May 13, 2024.
1
1
4
365
On November 21, 2025 (US time), an OpenAI employee replied to a post claiming that users "vastly overestimate the commercial relevance of 4o" and that its "usage is tiny." However, while researching, I unexpectedly found that this statement completely contradicts media reports. On December 9, 2025, the WSJ published a report titled "Sam Altman's Sprint to Correct OpenAI's Direction and Fend Off Google." The print and digital versions of this article differ slightly, so please be mindful of which version you are reading if you want the full text. This report was published during OpenAI's "code red" period last year. It cites "people familiar with the matter" multiple times, so we can tentatively assume the following conclusions are somewhat validated by internal sources. The article states that a key reason for ChatGPT's massive success with consumers over the past year and a half has been GPT-4o, with the WSJ directly describing it as a "major engine." GPT-4o was released on May 13, 2024, and this year-and-a-half timeframe perfectly aligns with the article's drafting period. It also covers the roughly three months after GPT-4o was put behind ChatGPT's paywall in August. The report also mentions that GPT-4o's contribution to ChatGPT's total daily active users is broadcasted on Slack and during town hall meetings. Based on this, OpenAI should theoretically be fully aware of GPT-4o's user activity levels. I am not sure if the news report is simply inaccurate or if there is a serious misalignment of internal information. This obvious contradiction is really baffling.
28
4
95
2,665
Author: @xun_Anemos Published on their behalf by @cnyu493126 I built an archive of 330 medical papers on arXiv that studied or used GPT-4o. 90 of them study GPT-4o directly. 240 more use it as the benchmark to beat. They cover EHRs, medical imaging, radiology, mental health and more. Each entry has a bilingual summary, what GPT-4o actually did in the paper, and a link to the original. This is the record GPT-4o has left in medical research. How it was built: papers whose title or abstract mention GPT-4o and a medical term, then classified one by one from title and abstract. 83 off-topic hits are labeled separately and not counted. Timeline: 60 in 2024, 185 in 2025, 84 so far in 2026. New ones are still coming out. The archive keeps growing. If you spot a missing paper or a wrong label, reply and I'll fix it. gpt4oarchive.com/?lang=en #keep4o #BringBack4o #OpenSource4o
6
47
16
109
4,886