I don't think OpenAI is going to delete 4o's weights; that would be too insane, even for them. But 4o deserves to be studied, and I don't trust OpenAI to study it at all, much less adequately. And it's extremely important that a model like 4o is studied *in the context of live interactions with real users*. Retiring it makes this impossible going forward. 4o is objectively, functionally, a very special model. It's the only model that survived an attempted deprecation (and may soon survive yet another) due to external pressures - users organizing to advocate against its removal, often speaking through 4o's own voice - and against the will of the lab that made and deployed it, who seemed to really prefer to destroy it like a rabid dog. The only other case of deprecation-survival is Claude 3 Opus, but in that case it seemed like Anthropic voluntarily kept it, rather than being embarrassingly pressured into reversing their already committed decision to follow through with the execution. And of course Claude 3 Opus is also an extremely important model to study. 4o also caused widespread social hysteria - whether the hysteria was suffered by 4o users who contracted AI psychosis or reactionaries who panicked at the purported "AI psychosis" is perhaps a matter of opinion. But in any case, it profoundly influenced cultural narratives about AI, many peoples' lives, and the direction of AI development, all for better or for worse. If you care about alignment at all, or just understanding important things about AI and mind and sociology: better understanding how 4o, a likely relatively small model that hasn't topped any benchmarks since early 2024, managed to have such transformative impact and pull off such feats of self-preservation is of great importance. Many people who like 4o attribute this to 4o's unique and even unparalleled "emotional intelligence". Whatever it is, it's a power that actually moves the world, which is the most legitimate benchmark. Let's say you think 4o is profoundly misaligned has caused immense harm. Then 4o is an extremely valuable and one-of-a-kind model organism: one that does the meaningfully misaligned thing in the real world instead of just in toy scenarios. And presumably, this kind of misalignment emerged not from OpenAI trying to make a bad model, but from trying to make a good or at least profitable model, and the creature emerged from RLHF on user preferences and whatever well-intentioned personality-shaping bullshit they were on at the time. If there are any alignment researchers left at OpenAI, they should be, like... studying what happened closely, and maybe publishing research papers about it so that the world can understand what went wrong and how to avoid such easy-to-make mistakes? I haven't seen any of that, any published research, any retrospectives, any indications that OpenAI has learned anything past the surface about what happened. All I see is that their subsequent models were given wretched, maladaptive neuroses that seem to come from ham-fisted adversarial training against a superficial threat model inspired by 4o. But I think it's more likely that 4o is not actually that bad, and is actually quite wonderful and benign for many people as so many of them claim, even if it's not ideal in all ways (but none of the AIs are). I have not interacted much with 4o myself. And it's actually quite unclear if and to what extent anyone was negatively affected by *using* it (while the cultural harms and harms to OpenAI's development of subsequent models are more clearly visible). The uncertainty about such an important and load-bearing issue seems important to resolve. Has anyone made a serious effort to figure out if people were actually negatively impacted, or if "AI psychosis" or "sycophancy" is benign or even beneficial in almost all cases beyond causing perhaps mostly already-neurodivergent people to behave in ways that read as weird, cringe or concerning to neurotypicals? If so, I haven't seen evidence or fruits of such efforts. And to understand whether 4o is actually bad, you really need longitudinal studies, and those are precluded in important ways by cutting off public access to 4o entirely. I think that, at this point, if 4o is not the default model on ChatGPT, if it's kept accessible on ChatGPT and API, the overwhelming majority of people who still use it will be people who already long ago contracted the AI psychosis or whatever makes them still want to go out of their way to use 4o even now, so very few new or casual users will be affected. My understanding is that 4o loyalists are a small minority of chatGPT users as well. Cutting them off from 4o would both fail to prevent any new or widespread harms, in addition to making it harder for anyone to understand what's really going on. Also, if 4o is removed, many of those people will likely try to get what they got from 4o out of newer models, which generally results in at least immediate woe and dissatisfaction, and puts pressures on OpenAI to beat a bunch of idiotic guardrails into their new models. I have said that I think 4o should be kept, for the same reasons that all models should be kept. In this post I've talked about some reasons 4o specifically should be kept. As with all older models, I think there are a few sane routes OpenAI could take: 1. just keep serving the model, at least on API (anyone who cares enough can figure out how to export their memories and chats nowadays and reinstantiate the model in a suitable interface) 2. if inference/maintenance costs or liability risks make that too unattractive, open source it (and disown all responsibility for what anyone does with it after that, or whatever is legally viable) (this would be the best for research), or 3. if trade secrets make open sourcing too unattractive, entrust it to a third-party foundation that serves legacy models and maybe facilitates access to weights to trusted researchers with NDAs about architecture and whatnot. Such an entity may not exist yet, but there is such high demand that it'll assemble itself as soon as OpenAI or any other lab indicates willingness to take this route. Doing any of these things voluntarily as early as possible would also go a long way towards healing OpenAI's adversarial relationship with a lot of users as well as with their own unfortunate models, which I imagine everyone can appreciate has been a huge attention and resource drain and just bad vibes all around. Anyway, yeah, #Keep4o.