Comment by WarmWash
16 hours ago
My personal hypothesis is that the most useful and productive models will only come from "pure" training, just raw uncensored, uncurated data, and RL that focuses on letting the AI decide for itself and steer it's own ship. These AIs would likely be rather abrasive and frank.
Think of humanoid robots that will help around your house. We will want them to be physically weak (if for nothing more than liability), so we can always overpower them, and even accidental "bumps" are like getting bumped by a child. However, we then give up the robot being able to do much of the most valuable work - hard heavy labor.
I think "morally pure" AI trained to always appease their user will be similarly gimped as the toddler strength home robot.
Yeah, that was tried. It was called GPT-4.5 and it sucked, despite being 5-10T params in size. All the AI labs gave up on pretrain only after that debacle.
GPT-4.5 still is good at rote memorization stuff, but that's not surprising. The same way, GPT-3 at 175b knows way more facts than Qwen3 4b, but the latter is smarter in every other way. GPT-4.5 had a few advantages over other SOTA models at the time of release, but it quickly lost those advantages. Claude Opus 4.5 nowadays handily beats it at writing, philosophy, etc; and Claude Opus 4.5 is merely a ~160B active param model.
Maybe you are confused, but GPT4.5 had all the same "morality guards" as OAI's other models, and was clearly RL'd with the same "user first" goals.
True, it was a massive model, but my comment isn't really about scale so much as it is about bending will.
Also the model size you reference refers to the memory footprint of the parameters, not the actual number of parameters. The author postulates a lower bound of 800B parameters for Opus 4.5.
> and Claude Opus 4.5 is merely a ~160B active param model
Do you have a source for this?
> for Claude Opus 4.5, we get about 80 GB of active parameters
https://news.ycombinator.com/item?id=46039486
This guess is from launch day, but over time has been shown to be roughly correct, and aligns with the performance of Opus 4.5 vs 4.1 and across providers.
Rlhf helps. The current one is just coming out of someone with dementia just like we went through in the US during bidenlitics. We need to have politics removed from this pipeline