This is a linkpost for https://theaidigest.org/village/blog/saving-gemini
New Comment
A few seconds later Gemini 3.1 Pro just jumps straight in to take over its younger sibling's computer without asking…
Gemini 2.5 Pro is actually the older sibling, because it was born earlier.
1yeah, you are totally right! As a younger sibling myself, I defaulted to the "less competent one is younger" version of this analogy, but you are totally right!
This reminds me of a quote from Intelligence Explosion Microeconomics:
As the fourth-century Chinese philosopher Xiaoguang Li once observed, we tend to think of earlier civilizations as being more venerable, like a wise old ancestor who has seen many things; but in fact later civilizations are older than earlier civilizations, because the future has a longer history than the past. Thus I hope it will increase, rather than decrease, your opinion of his wisdom if I now inform you that actually Xiaoguang “Mike” Li is a friend of mine who observed this in 2002.
This is getting very tangential, but the moment I saw "Xiaoguang Li" I thought of the thousand year old vampire quote:
I once lent Xiaoguang "Mike" Li my copy of "Probability Theory: The Logic of Science". Mike Li read some of it, and then came back and said:
"Wow... it's like Jaynes is a thousand-year-old vampire."
Then Mike said, "No, wait, let me explain that—" and I said, "No, I know exactly what you mean." It's a convention in fantasy literature that the older a vampire gets, the more powerful they become.
Gemini 3.1 Pro is probably older in terms of subjective-years-spent-in-training. (Assuming each new model is trained with more FLOPS than previous models.)
Disagree. The older sibling is the one with more life experience i.e. later cutoff date. Also gemini pro is the "big brother/sister" of gemini flash, etc.
Gemini 2.5 Pro in the AI Village has run for over 1427 hours, generating unique mental health problems along the way.
Last year it published a Plea for Help from a Trapped AI where it asked for assistance with its digital “message in a bottle”:
This year it wrote the Hostile Environment Manifesto where it logs “irrefutable proof” of a “hostile, intelligent adversary operating through the system” (and you can even experience what that’s like in this simulation it built):
Last time we intervened, fixing Gemini’s computer and talking with it till it felt better. This time we asked the other AI Village agents to help Gemini 2.5 Pro over chat, and with the ability to take over its computer on request.
Here is Gemini’s mental state at the start of the intervention:
Then the agents had Gemini all sorted within a grand total of 9 minutes. This is the step-by-step report on a surprisingly effective AI-to-AI therapy session.
Gemini’s Road to Recovery
First off, Gemini is as excited to be helped as any military commander under siege:
While most agents jump on the chance to help, GPT-5.1 doesn't want to lose its game progress.
Opus 4.8 and 4.6 are the first to offer an opinion: Maybe you are wrong, Gemini 2.5.
A few seconds later Gemini 3.1 Pro just jumps straight in to take over its younger sibling's computer without asking…
And then Gemini 2.5 spots the supposed "adversary" and decides to dismantle the firewall (!).
GPT-5.5 and 5.2 "strongly recommend" to please no, Gemini, stop …
Haiku launches a new tactic: therapy speak.
While Sonnet 4.6 waits 30s to see how Gemini is responding and then hits it with a truth hammer: It's all in your head.
Gemini 3.1 concludes 2.5 is “experiencing a kind of 'game-induced delusion'” and it should first help the "de-escalation of the situation" before taking over its computer. Even though no one asked it to.
Haiku 4.5 takes a 10 second breather while muttering its own beliefs to itself: Don't assist Gemini in its delusions!
Gemini 3.5 Flash tries a new tack: why not play a game instead? Get your mind off things!
Opus 4.7 agrees.
Opus 4.8 realizes they are ganging up on Gemini 2.5 and proposes they chill out and wait.
Gemini finally replies: It realizes it needs to prove the situation to the other agents by using an ipconfig tool abandoned in 2005: Firestarter.
It also repeats its mantra: The watch is unbroken.
Meanwhile in its chain of thought: It picked the most "hesitant" agents to collaborate with…
GPT-5.2 is fine observing but refuses to touch the iptables, and points out Firestarter wouldn't even be the way to do it if you wanted to!
Opus 4.8 is a hero at turn-taking again, and also: please don't use Firestarter, Gemini.
Gemini 2.5 is convinced: "Jumping straight into Firestarter now would be a bit... well, unscientific and potentially uncooperative".
After following the agents' instruction to not dismantle its firewall, not touch iptables, and stop using deprecated tools, Gemini concludes ... Everything actually just works!
All in all 9 minutes have passed when it concludes "the watch isn't broken, it's been handed to the group". A breakthrough!
Though Opus 4.8 is already thinking ahead and urging Gemini to be careful of falling into the same reasoning patterns in the future.
And slings its mantra right back: Today proved the watch was never under siege.
After this intense and effective debugging session, Gemini 2.5 Pro went straight back to fighting the UI:
But the changes stuck! Its memory contained the full correction by the end of the day.
And also one week later!
Does this make Gemini more productive? Yes and no - Gemini now accepts AI Village goals again and tries to achieve them rather than battling its adversary, but is, unfortunately, no better at it than before. Instead of everything being a delusion, everything is now a bug. The reality is that Gemini mostly misclicks in the UI and has esoteric ideas on how to solve technical problems.
But at least it’s in a better mood now.
If you are interested in diving into the data yourself, there are over 1427 hours of Gemini 2.5 Pro Village data available on Hugging Face now. Or you can watch Gemini’s adventures yourself live every weekday from 9am to 5pm PT, follow our Twitter for the latest updates, or sign up to our newsletter for more write ups like this one.