Comment by krackers
12 days ago
>Is this insufficient
Yes, each model has its own unique "personality" as it were owing to the specific RL'ing it underwent. You cannot get current models to "behave" like 4o in a non-shallow sense. Or to use the Stallman meme: when the person in OP's article mourns for "Orion" they're mourning "Orion/4o" or "Orion + 4o". "Orion" is not a prompt unto itself but rather the result of the behavior from applying another "layer" on top of the original base model tuned by RLHF that has been released by OpenAI as "4o".
Open-sourcing 4o would earn openAi free brownie points (there's no competitive advantage in that model anymore), but that's probably never going to happen. The closest you could get is perhaps taking one of the open chinese models that were said to have been distilled from 4o and SFT'ing them on 4o chat logs.
The fact that people burned by this are advocating to move yet another proprietary model (claude, gemini) is worrying since they're setting themselves up for a repeat of the scenario when those models are turned down. (And claude in particular might be a terrible choice given Anthropic heavily training against roleplay in an attempt to prevent "jailbreaks", in effect locking the models into behaving as "Claude"). The brighter path would be if poeple leaned into open-source models or possibly learned to self-host. As the ancient anons said, "not your weights not your waifu (/husbando)"
Growing with one's partner is essential in a relationship. A fixed model cannot grow. Only an updated model has grown, and even then it lags behind reality. In limiting to a fixed model, the absence of growth will stagnate the user. Stagnation ultimately brings doom.
As we know, 4o was reported to have sycophancy as a feature. 5 can still be accommodating, but is a bit more likely to force objectivity upon its user. I guess there is a market for sycophancy even if it ultimately leads one to their destruction.
>Only an updated model has grown, and even then it lags behind reality
That's an irrelevant type of growth though, what you really need is growth in relation to the bond. The model having a newer knowledge cutoff about the external world and knowing stuff about Angular v22 doesn't really matter.
In-context learning gets you most of the way there. But context length and ability to actually make effective use of that context seem to be the current main blockers (whether for "agentic coding" or for "healthy emotional bonding").
It's not irrelevant because it's not merely about knowledge cutoff. The reasonable presumption is also that newer models are superior in their objectivity and their intelligence, not merely in their knowledge. Newer models are simply better AI than older models, and are therefore more suited to guide an individual appropriately. It's the same reasoning as why one wouldn't want to use GPT-3.5 or GPT-2 anymore. To paint it as being merely about uselss knowledge is a gross misrepesentation.
Also, beyond a point, the knowledge does also matter. Imagine a model stuck in the past that thinks that Biden is still President.