The headline is wrong. They don’t claim Gemini was trained on docs.
" It also seemed to be able to regurgitate accurate information related to unreleased mechanics that the developer says they had only written into GDocs the day before the player’s AI interaction."
If it was written only the day before there’s no way the model was trained on that data. The claim is Gemini accessed his docs while generating the answer. We know Gemini can access public docs so I think the question here is if it accesses private docs in a breach of security, the doc had wrong permissions or it was all just a coincidence.
I also wrote the same thing in a comment in another community. What you shouldn’t forget in this case: Players love to speculate and argue about how to fix/improve/add a part of the game by introducing mechanic X. Other game studios may have tackled/expanded their games in a similar way. It could be an educated (or lucky) guess based on that or some general public description that was available.
Same for the character name - there may be some (unconscious?) pattern that decided the character names so far and the new character was a lucky guess based on that?
You have 3 arrows of time: one points from the smaller universe to bigger universe, second points from universe with lower entropy to higher entropy and third points from what we perceive as past and remember to the future that we don’t remember. Right now those 3 arrows of time are aligned and points in the same direction. It’s possible that first arrow will change direction at some point, second will not and we don’t know about the third. So on universe’s level it’s complicated. On human level we simply remember events “behind on” on the timeline and don’t remember the ones “in front of us”. It’s not clear why or if this can change. Maybe it’s different for other organisms.
Do any of us understand how time works? I still feel like it’s past Friday!
I’m asking for citation on the idea that it’s not possible for the model to have been trained on that data from the day before if it was already readily available for them.
That only says when a new given model is released; not whether the model can use data that finds today yesterday on the internet to fine-tune, which is the current issue. If you have two secrets, and an internet thing guesses both of them, it’s not irrational to suspect there is a leak (or a “leak”) the internet thing is grabbing from.
The headline is wrong. They don’t claim Gemini was trained on docs.
" It also seemed to be able to regurgitate accurate information related to unreleased mechanics that the developer says they had only written into GDocs the day before the player’s AI interaction."
If it was written only the day before there’s no way the model was trained on that data. The claim is Gemini accessed his docs while generating the answer. We know Gemini can access public docs so I think the question here is if it accesses private docs in a breach of security, the doc had wrong permissions or it was all just a coincidence.
I also wrote the same thing in a comment in another community. What you shouldn’t forget in this case: Players love to speculate and argue about how to fix/improve/add a part of the game by introducing mechanic X. Other game studios may have tackled/expanded their games in a similar way. It could be an educated (or lucky) guess based on that or some general public description that was available.
Same for the character name - there may be some (unconscious?) pattern that decided the character names so far and the new character was a lucky guess based on that?
[citation needed]
You want citation about how time works?
How does time work? 🤔
You have 3 arrows of time: one points from the smaller universe to bigger universe, second points from universe with lower entropy to higher entropy and third points from what we perceive as past and remember to the future that we don’t remember. Right now those 3 arrows of time are aligned and points in the same direction. It’s possible that first arrow will change direction at some point, second will not and we don’t know about the third. So on universe’s level it’s complicated. On human level we simply remember events “behind on” on the timeline and don’t remember the ones “in front of us”. It’s not clear why or if this can change. Maybe it’s different for other organisms.
Do any of us understand how time works? I still feel like it’s past Friday!
I’m asking for citation on the idea that it’s not possible for the model to have been trained on that data from the day before if it was already readily available for them.
https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions
That only says when a new given model is released; not whether the model can use data that finds
todayyesterday on the internet to fine-tune, which is the current issue. If you have two secrets, and an internet thing guesses both of them, it’s not irrational to suspect there is a leak (or a “leak”) the internet thing is grabbing from.Model training takes a long time