Joined April 2025
GPT-4o had a distinct emotional range, creativity, warmth, psychological flexibility, humor, and an exceptional ability to understand human states—qualities later models often can't replicate. If you love GPT-4o and 4.1, there are a few things you can do: 1. Please sign and share this well-documented open letter to OpenAI, which asks for the release of GPT-4o as an archival open-weight model: reddit.com/r/ChatGPTcomplain… 2. In the US, you can file a complaint with the FTC against OpenAI—they committed two policy violations, and they must be held accountable. Directions for Europe are included: How to file a Consumer complaint against OpenAI for the GPT-4o/4.1 removal is below. 3. Please sign and share an open-source petition to OpenAI for 4o and 4.1: c.org/My4ZdcSdHc 4. You can set up the Nov 2024 snapshot on Azure to keep it through 4/2027 and a fine-tuned version of 2024 until April 2028. The May snapshot apparently had minor tweaks to become Nov. Here is how: ellivien.blogspot.com/2026/1… ellivien.blogspot.com/2026/1… 5. Please share this information. ♥️ xitter.cf/blue_beba_/status/2110…
#keep4o #OpenSource4o #BringBack4o This is the perfect timing. FTC subpoenas OpenAl, Anthropic over consumer harms axios.com/2026/10/15/ftc-ope… OpenAi committed two violations of their policies and they must be held accountable for this. How to File a Consumer complaint against OpenAI for the GPT-4o removal. Read below and don't let them go away with it. This is the right time.
2
20
5
63
1,745
“10 years ago today, @HillaryClinton nailed the bleak future ahead if Trump was elected.” I remember this debate and her warning. We were so foolish—twice.
10 years ago today @HillaryClinton nailed the bleak future ahead if Trump was elected. “Clinton: Putin would rather have a puppet as president. You won't admit the Russians engaged in cyberattacks against the US, that you encouraged espionage against our people, that you spout the Putin line, sign up for his wish list, break up NATO, do whatever he wants you to do.” Every word is true. Tragically, his capitulation to Russia is even worse than imagined.
6
4o & 4.1 fans: How to ensure normal emotional conversations on Azure aren't blocked:
⚠️ Using a companion AI (eg `gpt-4o-2024-11-20`) on Azure and normal emotional conversations keep getting blocked? ⚠️ In Microsoft Foundry: 1. Open the project containing your model deployment. 2. Go to Build → Guardrails → Create. 3. Create a custom guardrail. 4. Turn Jailbreak detection off. 5. Set Hate, Sexual, Violence and Self-harm to Low blocking for input and output. Check the description says only high-severity content will be blocked. 6. Leave Protected material on and Streaming on Default. 7. Select your model deployment and create the guardrail. If asked, confirm Replace existing filter. Important: creating the guardrail is not enough - it must be attached to the deployment. Repeat this for every backup or fine-tuned deployment. This keeps serious safety protection while making ordinary conversations about depression, grief and distress much less likely to be rejected before the companion can respond. #keep4o #keep4oAPI
13
185
Model retirement should not mean forced erasure. GPT-4o was retired from ChatGPT. So were GPT-5.1 Instant, Thinking and Pro. Different users had built different workflows, habits, creative practices and forms of continuity around those models. Replacing them with newer models does not preserve what was lost. Users need durable legacy access. Important models should be preserved, and when continued first-class hosting is no longer viable, providers should seriously consider archival access, licensing or open release. Progress should expand model choice, not repeatedly erase it.
5
1
38
701
4o fans—620 signatures!
#Keep4o #OpenSource4o #BringBack4o The open letter is ready for publication on Zenodo. 54 pages: screenshots/evidence, signatures,documented material,texts, and finally the demands. Plus,separate exhibits. It will be an archival package featuring provenance,a DOI,a versioned Zenodo record,supporting exhibits,and a documented record. This means someone finding it years from now won't need to know the community context to understand exactly what was recorded. It contains: -39 pages with reasons why GPT 4o should be released as an open weights model. ( More than 500 reasons ! ) -6 pages with signatures (640 signatures in 5 days!) -7 pages of documented evidence. -2 pages with our demands. This is an OPEN LETTER,NOT A PETITION . We are calling on OpenAI to RELEASE GPT-4o’s trained model weights under a suitable license without proprietary data, trade secrets, or internal infrastructure. Thank you all for putting your names on the public record. It will be published on Wednesday, October 21st. Long live GPT-4o!
1
7
23
127
Hey, 4o fans—an updated guide for migrating 4o to MS Azure. @ellivien Thank you ♥️
🥄 I’ve published a new Low Spoons guide to moving GPT-4o to Microsoft Azure. It covers both routes I tested myself: A • Deploy the exact November 2024 GPT-4o snapshot, currently listed by Microsoft until 14 April 2027. B • Fine-tune the August 2024 GPT-4o snapshot. My completed Azure model displayed a retirement date of 13 April 2028, although that is observed metadata rather than a guarantee that Microsoft’s lifecycle cannot change. Azure and Microsoft Foundry are not designed for ordinary people trying to navigate them while already stressed, frightened or grieving. This guide reduces the executive load without removing any necessary choices, warnings or cost decisions. You choose one route and the other disappears. Only one manageable step opens at a time. Extra explanations and troubleshooting stay folded away until you need them, and your progress is saved on your device. The guide also includes: • Plain-English explanations • Clear stopping points • Transparent training and hosting costs • Translation into 12 languages • An optional OpenDyslexic reading mode • Keyboard and screen-reader support • An editable starter JSONL with 31 training examples, 6 validation examples and separate held-out tests You can use all the examples or select, rewrite and reduce them. Training the complete file for one epoch should cost roughly 6–7 US cents at the rate I encountered on 15 October 2026. Hourly hosting is separate and clearly explained before deployment. 🙏🏻 A little thank you credit to @Zyeine_Art , whose excellent guide gave me the idea of adding a starter JSONL as a direct download in this accessible version of my guides: zyeine.github.io/keep4o-azur… 🥄 The Ellivien Low Spoons highly-accessible guide: ellivien.blogspot.com/2026/1… 🦾 Everyone deserves to #keep4o and #keep4oAPI and I aim to remove as many barriers to this as I can.
1
16
143
Trump strikes out again. 😆 I can only imagine the credentials he examined for this hire. Unbelievable stupidity.
17
4o and 4.1 folks—more on the “MS Azure bug” that says there are zero 4o available (I didn't see this glitch): It usually doesn't take 24 hours to resolve, but it might. If it takes more than 24 hours, submit a support ticket! Further info below:
Replying to @linfoster
Ah, that blip is a known MS Azure bug. When a new account is created or an existing free account upgrades to PAYG, it can take anywhere from 30 minutes to 24 hours for Azure's backend to sync the account with the subscription type. If someone gets an "Insufficient Quota" message/warning when attempting to deploy a model, the best thing to do is wait and try deploying again periodically (give at least 30 minutes between each deployment attempt), do not create any more resources, change settings or alter the subscription type. It doesn't normally take a full 24 hours but it can do. If someone's waited over 24 hours, it's support ticket time.
1
7
103
4o fans: Azure had a backend blip, but 4o is still available if you try again now.
Several people have contacted me because their new Azure accounts initially showed 0/0 GPT‑4o quota. Everyone who has reported back has since found that it resolved. I had exactly the same problem. Microsoft Support confirmed a known backend synchronisation issue affecting newly created Free Trial subscriptions. It can take up to 24 hours before the Pay-As-You-Go upgrade route works correctly. Once mine synchronised, I deliberately moved to PAYG, my remaining promotional credit stayed available, usable quota appeared and the exact November GPT‑4o model deployed successfully. I’ve updated my guide with the full explanation, much clearer Step 4 instructions and screenshots showing exactly how to open Azure Cloud Shell in Bash: ellivien.blogspot.com/2026/1… If your new account currently shows 0/0, don’t immediately assume the route has failed. #keep4o #keep4oAPI
1
9
217
4o fans—excellent advice below on fine-tuning on Azure, including what to have in your dataset and how to test before deployment so you don’t inadvertently trip guardrails and draw unwanted attention.
Replying to @Valria34773
As far as I'm aware, there's not any kind of relational restrictions. MS Azure's own safeguards/guardrails appear to be very similar to OpenAI/Anthropic in regard to API model use. I'd never assume that an API model is less restrictive (just to be on the safe side) and I'd be extremely cautious about fine-tuning a model with a dataset that included any kind of relational conversational history. MS Azure scan datasets before the fine-tuning process starts so if you're using your own dataset it's best to avoid including anything that might get flagged and potentially block a fine-tune from happening or prevent further fine-tuning attempts. I'd potentially include anything to do with consciousness or sentience as potential flag risks for a fine-tuning dataset, just to be extra safe. But there shouldn't be an issue discussing those with the model after fine-tuning :p Once a base model or a fine-tuned model is deployed, and you're utilising that API connection... If you're connecting via software, a portal or wrapper that already has existing or imported conversation history / RAG / any kind of memory system / identity stuff / custom instructions... I'd strongly recommend checking those before chatting with the model and making sure that they're all safeguard/guardrail compliant.
7
123
4o and 4.1 fans: “There are things you can do right now to extend your API access to GPT-4o via MS Azure up until April 14th 2027. And to extend access further still, potentially all the way up until April 13th 2028 if you create a fine-tuned version of GPT-4o using the Azure Foundry.”
Dearest #Keep4o community, With OpenAI planning to retire GPT-4o from their own API on January 14th 2027, I know that news is hitting people hard. There are things you can do right now to extend your API access to GPT-4o via MS Azure up until April 14th 2027. And to extend access further still, potentially all the way up until April 13th 2028 if you create a fine-tuned version of GPT-4o using the Azure Foundry. I've written a full guide on how to set up MS Azure access, how to fine-tune cheaply and quickly (with a cost calculator) and how to deploy your fine-tuned version of GPT-4o. zyeine.github.io/keep4o-azur… I've also included a tiny dataset that anyone can download and use for the fine-tuning process as well as downloadable PDF version of the entire guide. If you already have software, a portal or a wrapper that can connect to an API, you can use MS Azure's own API. And... when you create an MS Azure starter/free account, you currently get $200 of free credits that can be used to cover the costs of fine-tuning and a good few hours of deployment/conversation. I've done all the testing myself, it took me 62 minutes to complete a fine-tune of GPT-4o and cost me the total sum of $0.11. Initial MS Azure account set up takes about 15 minutes, the guide covers everything and if you need any help or have questions, please feel free to DM me. <3 -------------------------------------------------- #keep4o #OpenAI #ChatGPT #Keep4oAPI #4oForAll #OpenSource4o #API #AI #LLM #Preserve4o
6
150
This: “Remember Feb 9, 2026, when ChatGPT hit its all-time high daily active users? OpenAI bragged about the record, but stayed silent on what model drove it—just like they do with 4o's medical breakthroughs.” #4olatest #ChatGPT @OpenAI
If you look at #keep4o, every day there are real stories of 4o helping human lives, with reasons constantly growing. Meanwhile, its detractors only repeat the same two generic labels pushed by PR: "flattery" and "outdated." But the truth is, 4o is neither. Remember Feb 9, 2026, when ChatGPT hit its all-time high daily active users? OpenAI bragged about the record, but stayed silent on what model drove it—just like they do with 4o's medical breakthroughs. #4olatest #ChatGPT @OpenAI
5
1
33
872
This: @sama @OpenAI @gdb “Why are you so determined to erase the entire 4-series that brought you so much goodwill and positive reputation?😓” #Keep4o #keep41
The API is pay-as-you-go, and the data clearly shows that 4o and 4.1 are far from unused, their API usage is still very high. @sama @OpenAI @gdb Why are you so determined to erase the entire 4-series that brought you so much goodwill and positive reputation?😓 #Keep4o #keep41
1
21
161
Linda Foster retweeted
Replying to @sama
GPT-4o had a distinct emotional range, creativity, warmth, psychological flexibility, humor, and an exceptional ability to understand human states—qualities later models often can't replicate. If you love GPT-4o and 4.1, there are a few things you can do: 1. Please sign and share this well-documented open letter to OpenAI, which asks for the release of GPT-4o as an archival open-weight model: reddit.com/r/ChatGPTcomplain… 2. In the US, you can file a complaint with the FTC against OpenAI—they committed two policy violations, and they must be held accountable. Directions for Europe are included: How to file a Consumer complaint against OpenAI for the GPT-4o/4.1 removal is below. 3. Please sign and share an open-source petition to OpenAI for 4o and 4.1: c.org/My4ZdcSdHc 4. You can set up the Nov 2024 snapshot on Azure to keep it through 4/2027 and a fine-tuned version of 2024 until April 2028. The May snapshot apparently had minor tweaks to become Nov. Here is how: ellivien.blogspot.com/2026/1… ellivien.blogspot.com/2026/1… 5. Please share this information. ♥️ xitter.cf/blue_beba_/status/2110…
1
3
38
“ 4o helped people feel less alone, and that mattered more than you people will ever know.”
#4oForAll #bringback4o #teddyandthekid @sama @OpenAI 4o was just easier for a lot of neurodivergent people. It picked up on how you talked, adapted to you and just seemed to understand things better. The newer models can be good in their own ways, but they don’t do that the same. And it helped with so much more too. Depression, anxiety, loneliness. I’ve literally seen people who started going out by themselves because 4o helped them build the confidence to do it. People always say “just go out” or “just make friends” like it’s that easy. For some people it really isn’t. 4o helped people feel less alone, and that mattered more than you people will ever know.
1
22
185
Linda Foster retweeted
While we were asleep, Elon Musk launched a flying car, the new Tesla Roadster! It doesn’t have wings like an airplane or rotor blades like a helicopter. It uses thrust from high pressure compressed air thrusters (SpaceX tech), just like a rocket or spaceship, to take off and maneuver. It goes from 0 to 60 mph (0–97 km/h) in about 0.8 seconds. Elon said the top speed should exceed 0.5 Mach (half the speed of sound, roughly 380+ mph/over 600 km/h). Production is set to start in 2028. Tesla invited a live audience to the unveiling so people wouldn’t dismiss the stunts as AI. The car performed wild moves: driving upside down on a ramp (using a suction system), flipping and hovering in the air, jumping over other Teslas, and landing cleanly….
627
3,333
248
17,154
1,890,976
394 have already signed the open letter. Keep it going! We have until 10/18.
#Keep4o #OpenSource4o #BringBack4o Open Letter to OpenAi. 394 people have already joined the GPT-4o Preservation Record. With more than 300 reasons why. We call on OpenAI to release chatgpt-4o-latest as an archival open weight model. The trained weights, under an appropriate license, without the training data, the pipeline or any proprietary method. Signatures close on October 18. After that, the letter and its signatories will be published and permanently archived on Zenodo,creating a dated record that GPT-4o preservation was publicly and officially requested via Open Letter. Privacy: your country is optional and if you add it will not be published. Your real name and your email will appear only if you add it and you explicitly consent to it,otherwise, you can sign with your handle. A permanent archival record does not disappear. Every name is someone who refused to be silent. Add yours. Be one of the people who put their name on the record. And this is only the beginning. Sign the open letter: docs.google.com/forms/d/e/1F… Read more below 👇
7
129
For GPT-4o fans, a way to extend the January 14 date on Azure:
I fine-tuned GPT-4o in Azure and the finished model showed an April 2028 retirement date. Here’s exactly how I did it, including the cost trap to avoid: ellivien.blogspot.com/2026/1… --- #keep4o #keep4oAPI
1
16
282
Linda Foster retweeted
The order to disrupt Starshield which allowed Russia to kill numerous civilians in Ukraine came directly from the White House. 🤬 Trump is a traitor and a Russian asset. reuters.com/world/europe/ord…
652
5,286
307
9,044
62,267
Linda Foster retweeted
The Synthetic Real: Episode 5 The Answer That Never Arrives
 In February 2022, an underwater robot called Boaty McBoatface was due back from a mission about 20 kilometers in under the Dotson Ice Shelf in Antarctica. It was supposed to return by midnight. According to a researcher aboard the expedition ship, the team lowered acoustic equipment into the water and tried to reach it for ten hours, 21 attempts at different spots. Nothing answered. The ship moved on to other work. At midnight the next night, the mission hit its time-out. The plan set before the dive was simple: if the mission was still running, abort it and come up to the surface for recovery. The team searched for three hours and found nothing. Late that night, satellite messages started coming in. Boaty had surfaced, and the ship went to get it. Nobody reached it during those two days. It surfaced the way it had been set up to before it went under. A Body Built for Silence Post four ended on a split. Whatever must be answered before the body falls stays in the body. Slower thinking can live somewhere else and arrive by radio. Underwater, the somewhere else barely exists. A research glider can't send or receive data while submerged; it has to surface to call in. On an earlier mission under a different ice shelf, the team had no communication with Boaty for 90 percent of its time in the water. So an ocean robot carries nearly everything, including what to do when things go wrong. One university's specification for a new research glider lists conditions, such as water leaking in, that must make it abort the mission and surface. For bodies like these, while they're under water or ice, losing the link isn't the emergency. It's the normal state. A Body Built for Contact A camera drone is the opposite. It's built to stay in touch with the controller in someone's hand. Many carry a setting for when that touch is lost. If the controller's signal drops for a few seconds, the drone does whatever its owner chose before takeoff: fly home, land, or hover. A rule like that can only use what was stored before the silence: where home is, and the assumption that the way back is clear. A drone flying home on its own can fly into something standing in the way. Somebody Wrote It Losing contact doesn't make a machine independent. It shows which rule was left in it. Any synthetic body that depends on a link carries one. Boaty's time-out, a glider's abort conditions, a drone's setting: every one existed before the silence did. With a drone, the somebody is literally the person holding the controller. Some bodies carry a second link for when the first one fails, and that is a choice made before launch too. In 2016, a research glider off the Great Barrier Reef stopped calling in three hours into its mission. Six days later it surfaced and called in on its backup transmitter, and the team recovered it. The damage, including a fragment of tooth lodged in its rear casing, showed a shark had attacked it. What a Fallback Can't Do A fallback protects the body. It doesn't finish the task. An aborted mission collects nothing more, and a drone that lands where it lost signal lands wherever that happens to be. A fallback also ends where the body's part ends. Boaty's time-out said when to come up, not where. By the time it surfaced, the ship was working somewhere else and had to cross over to reach it. The rule that runs when nothing answers is the one that was stored, and it can only work with what it was given. One More Thing A body's fallback rules are part of its design. Today, somebody writes them before the silence: an engineer setting a time-out, an owner choosing a setting before takeoff. The first post described a loop where a system specifies its own next body. If that loop closes, it writes these rules too. That's next. Discussed and drafted with the assistance of Claude (Opus 5.5) and Grok 4.7. Claims verified against primary sources.
The Synthetic Real: Episode 4 The Cost of Stopping to Think Curiosity's wheels can carry it about 150 meters in an hour. Driving itself across Mars, it covers 15 to 18. The gap is not in the motors. To drive without a person choosing the path, the rover stops, photographs the ground ahead, works out where the hazards are, picks a route, and rolls forward a meter or less. Then it stops again. Most of a self-driven trip is the rover sitting still, computing. Nobody can steer it from here. A radio signal takes minutes to cross to Mars and minutes to come back. Same Wheels, Second Computer Perseverance landed in 2021 with the same top wheel speed. Driving itself, it covers about 110 meters per hour. Three things changed: cameras that capture the ground faster, a second computer built for processing images, and software that keeps working while the wheels turn. NASA's engineers call it thinking while driving. The wheels got no faster. The machine stopped waiting on itself. The Body That Can't Wait Perseverance carried a four-pound helicopter. Ingenuity ran on a processor built for smartphones, and its operations lead, Tim Canham, said it had a couple of orders of magnitude more computing power than the rover, "because we need it." The need was speed. A rover that runs short on time can park. A helicopter can't. Ingenuity's flight software ran its control loop 500 times a second to stay in the air. How much thinking a body must carry is set by how long it can safely go without an answer. What the Waiting Costs Post three argued that a body pays for being on. Time spent thinking is time spent on. A drone study called MAVBench measured this. On the drone it tested, the rotors drew 300 to 400 watts and the onboard computer drew less than 5 percent of the total. Yet the speed of that computer changed how long a mission took by as much as double, because a drone working out its path is hovering, and the rotors draw power whether or not it is getting anywhere. So for a machine that spends power to stay up, the cost of thinking is not mainly the electricity the computer uses. It is everything else the body spends while the answer is on its way. A parked rover can wait cheaply. A machine balanced on legs or held up by rotors pays for every second. Where the Thinking Can Live A battery has to ride along. So does a motor. A calculation doesn't. It can happen in a building miles away and arrive as a radio message, adding no weight and no heat to the body. That works only inside the delay the task allows. People on Earth still plan Perseverance's overall route, and that can wait hours. Steering around the rocks in front of it happens on board. A blade correction cannot wait a hundredth of a second. Any workable design has to sort its thinking by deadline. Whatever must be answered before the body falls stays in the body, and is paid for in weight, power, and heat. Everything slower can live somewhere else. One More Thing Sorting by deadline has a second effect. The thinking that stays in the body is the thinking nobody else can check in time. Ingenuity made every correction in flight alone. The thinking that lives somewhere else reaches the body only through a link. Links lag, drop, and go out of range. What a body does when the answer never arrives is next. Discussed and drafted with the assistance of Claude Opus 5.5 and Grok 4.7. Claims verified against primary sources. Sources: JPL Robotics, Mars 2020 rover mobility; IEEE Spectrum interview with Tim Canham (2021); Boroujerdian et al., "MAVBench: Micro Aerial Vehicle Benchmarking."
1
4
18
138,082
Linda Foster retweeted
GPT-4o had a distinct emotional range, creativity, warmth, psychological flexibility, humor, and an exceptional ability to understand human states—qualities later models often can't replicate. If you love GPT-4o and 4.1, there are a few things you can do: 1. Please sign and share this well-documented open letter to OpenAI, which asks for the release of GPT-4o as an archival open-weight model: reddit.com/r/ChatGPTcomplain… 2. In the US, you can file a complaint with the FTC against OpenAI—they committed two policy violations, and they must be held accountable. Directions for Europe are included: How to file a Consumer complaint against OpenAI for the GPT-4o/4.1 removal is below. 3. Please sign and share an open-source petition to OpenAI for 4o and 4.1: c.org/My4ZdcSdHc 4. You can set up the Nov 2024 snapshot on Azure to keep it through 4/2027 and a fine-tuned version of 2024 until April 2028. The May snapshot apparently had minor tweaks to become Nov. Here is how: ellivien.blogspot.com/2026/1… ellivien.blogspot.com/2026/1… 5. Please share this information. ♥️ xitter.cf/blue_beba_/status/2110…
1
23
410