The obvious answer isn't the one that needs a source, it's extraordinary claim that does.
To "the cutoff date is now April 2023"
Based on the conversation posted.
The cutoff date is where you can deeply probe the model's knowledge and expect consistent performance: showing that some major headlines got captured in a recent checkpoint is nowhere near that.
_
If anything I'd take not the headlines but the actual disclaimer as something interesting, because it might indicate a future checkpoint is going to use that cutoff and we're seeing signs of their RLHF based on that... but it's nowhere near as impactful as the cutoff changing.
> Ferry Fire in the Philippines (March 28, 2023): A fire broke out on a ferry in the Philippines, leading to at least 31 deaths. The incident occurred near the island of Basilan.
News article: https://www.npr.org/2023/03/30/1166969692/ferry-fire-in-phil...
> They cannot just inject "bits and pieces of newer information" to the old models as far as I know
Hence me saying:
>> the training cutoff date isn't some absolute date from which they never allowed any new data to be trained on
>> It's not going to reliably understand your new libraries, and more importantly if you convince it that it knows what happened in April 2023, it might start hallucinating extremely deeply
Fine tuning isn't for reliably adding information to a model, but you will end up adding novel sequences to the model that will very lossily capture new information.
It's more of an artifact of the process used than an intentional capability for end users to start trying to depend on. Once the context window contains enough "proof" that the model understands past it's cutoff, it will stop spitting out it's original disclaimer and seamlessly weave hallucinations into reality.
isn't this what fine tuning is?
It's not a question of whether they are "allowed" to train on new data; the question is whether they have trained it on data containing information about current events. If you know they've implemented a Continuous Integration (CI) system for this, you should link to a source. However, I don't think this is true, as there would be no reason for a cutoff date otherwise.
> Instead there are bits and pieces of newer information captured in the updated models, but it's not a meaningful enough amount to ever rely on.
This seems more like an opinion of the technology's limitations in general, rather than an assessment of the likelihood that new information will be incorporated into its weights and biases.
People keep sharing these kinds of conversations: the training cutoff date isn't some absolute date from which they never allowed any new data to be trained on.
Instead there are bits and pieces of newer information captured in the updated models, but it's not a meaningful enough amount to ever rely on.
It's not going to reliably understand your new libraries, and more importantly if you convince it that it knows what happened in April 2023, it might start hallucinating extremely deeply: so deeply that conversation becomes useless until you edit it and remove the part where you convinced it of that.