Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I'm not saying this is a "live" update, but all usage is collected and used to inform later offline training or fine-tuning.

Whether they are used directly with the positive/negative signal given from users, or whether it's something more abstract, doesn't really matter. The important thing is that feedback is used to improve the responses over time.

As for whether a version is immutable, it seems this research may have been done on a previous version. But also I'm not sure if the model and weights are immutable, or whether it's just the model structure. It's clear the model is not stable so it's not like there's an API contract being met with fixed weights.

Edit: others are suggesting that the author used GPT-4 via ChatGPT, not by pinning the model. This would suggest that at least the ChatGPT tuned model is being frequently changed?



Being pedantic, (a) this comment is also incorrect, and (b) even if correct, wouldn't fix all these results immediately.

The simplest explanation is researcher error.


Assuming the researcher didn't lie, it seems unlikely that they got the responses wrong in some way.

The most likely alternative explanation I can think of is that this is the seemingly well know instability of results caused by the way the MoE architecture is implemented for GPT-4?

I'd love to understand what exactly is wrong in my understanding. I realise I've only got a layman's understanding of this, but it seems clear that OpenAI and others depend on these feedback loops to improve things over time. Is that not the case?


It is explicitly known the the training data cutoff for GPT4 is September 2021. While we can assume that feedback is taken into consideration for future training of new models, the training data used to train all current models is a specific bundle of data with that cutoff date.




Consider applying for YC's Fall 2025 batch! Applications are open till Aug 4

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: