The baked-in assumptions observation is basically the opposite of the impression I get after watching Gemini 3's CoT. With the maximum reasoning effort it's able to break out of the wrong route by rethinking the strategy. For example I gave it an onion address without the .onion part, and told it to figure out what this string means. All reasoning models including Gemini 2.5 and 3 assume it's a puzzle or a cipher (because they're trained on those) and start endlessly applying different algorithms to no avail. Gemini 3 Pro is the only model that can break the initial assumption after running out of ideas ("Wait, the user said it's just a string, what if it's NOT obfuscated"), and correctly identify the string as an onion address. My guess is they trained it on simulations to enforce the anti-jailbreaking commands injected by the Model Armor, as its CoT is incredibly paranoid at times. I could be wrong, of course.
I've had some weird "thinking outside the box" behavior like this. I once asked 3 Pro what Ozzy Osbourne is up to. The CoT was a journey, I can tell you! It's not in its training data that he actually passed away. It did know he was planning a tour though. It had a real struggle trying to consolidate "suspicious search results" and even questioned whether it was fake news, or running against a simulation!, determining it wasn't going to fall for my "test".
It did ultimately decide Ozzy was alive. I pushed back on that, and it instantly corrected itself and partially blamed my query "what is he up to" for being formulated as if he was alive.
“Gemini 3 Pro was often overloaded, which produced long spans of downtime that 2.5 Pro experienced much less often”
I was unclear if this meant that the API was overloaded or if he was on a subscription plan and had hit his limit for the moment. Although I think that the Gemini plans just use weekly limits, so I guess it must be API.
Geminii CLI has a specific "model is overloaded" error message which is distinct from "you're out of quota" so I suspect whatever tools they're using for this probably have something similar, and they're referring to that.
This is exactly why I upgrade to the Pixel 10 Pro. On Black Friday, you could get a Pixel 10 Pro for about $450 on the U.S. Google Fi store (which sells unlocked phones)... which is also about how much a Pixel 9 Pro goes for on eBay; minus eBay fees and accounting for shipping, that's an upgrade for < $100. But, it's even better, the Pixel 10 Pro comes with a year of their "AI Pro" plan (which I believe costs around $240/year.) There is really, really no point in upgrading to a Pixel 10 Pro from a Pixel 9 Pro, and environmentally it pains me to be the person upgrading my phone on an annual basis (this is the fastest I've ever upgraded a phone, ever) but it's hard to turn down when Google is selling $800~ish for $400~ish.
And yeah, it's not the insanely priced AI Ultra plan, but if there are any hard limits on Gemini Pro usage I haven't found them. I have played a lot with really long Antigravity sessions to try to figure out what this thing is good for, and it seems like it will pretty much sit there and run all day. (And I can't really blame anyone for still remaining mad about AI to be completely honest, but the technology is too neat by this point to just completely ignore it.)
Seeing as Google is still giving away a bunch of free access, I'm guessing they're still in the ultra-cash-burning phase of things. My hope (hopium, realistically) is that by the time all of the cash burning is over, there will be open-weight local models that are striking near where Gemini 3 Pro strikes today. It doesn't have to be as good, getting nearby on hardware consumers can afford would be awesome.
But I'm not holding my breath, so let's hope the cash burning continues for a few years.
(There is, of course, the other way to look at it, which is that looking at the pricing per token may not tell the whole story. Given that Google is running their own data centers, it's possible the economic proposition isn't as bad as it looks. OTOH, it's also possible it is worse than it looks, if they happen to be selling tokens at a loss... but I quite doubt it, given they are currently SOTA and can charge a premium.)
That's an extrapolation to finish the entire game.
If limit your token count to a fraction of 2 billion tokens, you can try it on your own game, and of course have it complete a shorter fraction of the game.
If you go to the X page linked on the blog, the page owner mentions a “collaboration” with Google Deepmind on this project. It wouldn’t shock me if this just an elaborate advertisement for Gemini.
"Crucially, it tells the agent not to rely on its internal training data (which might be hallucinated or refer to a different version of the game) but to ground its knowledge in what it observes. "
Yes, at least to some extent. The author mentions that the base model knows the answer to the switch puzzle but does not execute it properly here.
"It is worth noting that the instruction to "ignore internal knowledge" played a role here. In cases like the shutters puzzle, the model did seem to suppress its training data. I verified this by chatting with the model separately on AI Studio; when asked directly multiple times, it gave the correct solution significantly more often than not. This suggests that the system prompt can indeed mask pre-trained knowledge to facilitate genuine discovery."
I'm wondering about this too. Would be nice to see an ablation here, or at least see some analysis on the reasoning traces.
It definitely doesn't wipe its internal knowledge of Crystal clean (that's not how LLMs work). My guess is that it slightly encourages the model to explore more and second-guess it's likely very-strong Crystal game knowledge but that's about it.
It will definitely have some effect. Why won't it? Even adding noise into prompts (like saying you will be rewarded $1000 for each correct answer) has some effect.
Whether the 'effect' something implied by the prompt, or even something we can understand, is a totally different question.
It's hard to say for sure because Gemini 3 was only tested with this prompt. But for Gemini 2.5, which is who the prompt was originally written for, yes this does cut down on bad assumptions (a specific example: the puzzle with Farfetch'd in Ilex Forest is completely different in the DS remake of the game, and models love to hallucinate elements from the remake's puzzle if you don't emphasize the need to distinguish hypothesis from things it actually observes).
I would imagine that prompting anything like this will have an excessively ironic effect like convincing it to suppress patterns which it would consider to be pre-knowledge.
If you looked inside they would be spinning on something like "oh I know this is the tile to walk on, but I have to only rely on what I observe! I will do another task instead to satisfy my conditions and not reveal that I have pre-knowledge.
LLMs are literal douche genies. The less you say, generally, the better
Nice writeup! I need to start blogging about my antics. I rigged up several cutting edge small local models to an emulator all in-browser and unsuccessfully tried to get them to play different Pokémon games. They just weren't as sharp as the frontier models.
This was a good while back but I'm sure a lot of people might find the process and code interesting even if it didn't succeed. Might resurrect that project.
In my project I rigged up an in-browser emulator and directly fed captured images of the screen to local multimodal models.
So it just looks right at what's going on, writes a description for refinement, and uses all of that to create and manage goals, write to a scratchpad and submit input. It's minimal scaffolding because I wanted to see what these raw models are capable of. Kind of a benchmark.
As a fun comparison, Gemini 3 Pro took 17 days to beat the game.
Twitch Plays Pokemon, which was frequently random, chaotic, even malicious, took 13 days to clear Crystal.
It would unfortunately also need several runs of each to be reliable. There's nothing in TFA to indicate the results shown aren't to a large degree affected by random chance!
(I do think from personal benchmarks that Gemini 3 is better for the reasons stated by the author, but a single run from each is not strong evidence.)
So after years of being gleefully told that AI will replace all jobs an omniscient state of the art model, with heavy assistance, takes more than two weeks and thousands of dollars in tokens to do what child me did in a few days? Huh.
I used to think the same until latest agents started adding perfectly fine features to a large existing react app with just basic input (in English) . Most of the jobs require levels of intelligence below that. It's just a matter of time before agents get to that.
How certain can we be that these improvements aren't just a result of Gemini 3 Pro pre-training on endless internet writeups of where 2.5 has struggled (and almost certainly what a human would have done instead)?
In other words, how much of this improvement is true generalization vs memorization?
You're too kind. Even the CEO of Google retweeted how well Gemini 2.5 did on Pokemon. There is a high chance that now it's explicitly part of the training regime. We kind of need a different kind of game to know how well it generalizes.
There were no such writeups, 99% of the discussion about difficulties in Crystal were in twitch and discord chats where Google doesn't scrape. (It hadn't yet gotten the public attention that Claude and Gemini's runs of Pokemon Red and Blue have gotten.)
That said, this writeup itself will probably be scraped and influence Gemini 4.
Readit News
It did ultimately decide Ozzy was alive. I pushed back on that, and it instantly corrected itself and partially blamed my query "what is he up to" for being formulated as if he was alive.
I was unclear if this meant that the API was overloaded or if he was on a subscription plan and had hit his limit for the moment. Although I think that the Gemini plans just use weekly limits, so I guess it must be API.
And yeah, it's not the insanely priced AI Ultra plan, but if there are any hard limits on Gemini Pro usage I haven't found them. I have played a lot with really long Antigravity sessions to try to figure out what this thing is good for, and it seems like it will pretty much sit there and run all day. (And I can't really blame anyone for still remaining mad about AI to be completely honest, but the technology is too neat by this point to just completely ignore it.)
Seeing as Google is still giving away a bunch of free access, I'm guessing they're still in the ultra-cash-burning phase of things. My hope (hopium, realistically) is that by the time all of the cash burning is over, there will be open-weight local models that are striking near where Gemini 3 Pro strikes today. It doesn't have to be as good, getting nearby on hardware consumers can afford would be awesome.
But I'm not holding my breath, so let's hope the cash burning continues for a few years.
(There is, of course, the other way to look at it, which is that looking at the pricing per token may not tell the whole story. Given that Google is running their own data centers, it's possible the economic proposition isn't as bad as it looks. OTOH, it's also possible it is worse than it looks, if they happen to be selling tokens at a loss... but I quite doubt it, given they are currently SOTA and can charge a premium.)
If limit your token count to a fraction of 2 billion tokens, you can try it on your own game, and of course have it complete a shorter fraction of the game.
Did the streamer get subsidized by Google?
(The stream isn't run by Google themselves, is it?)
Does this even have any effect?
"It is worth noting that the instruction to "ignore internal knowledge" played a role here. In cases like the shutters puzzle, the model did seem to suppress its training data. I verified this by chatting with the model separately on AI Studio; when asked directly multiple times, it gave the correct solution significantly more often than not. This suggests that the system prompt can indeed mask pre-trained knowledge to facilitate genuine discovery."
It definitely doesn't wipe its internal knowledge of Crystal clean (that's not how LLMs work). My guess is that it slightly encourages the model to explore more and second-guess it's likely very-strong Crystal game knowledge but that's about it.
Whether the 'effect' something implied by the prompt, or even something we can understand, is a totally different question.
If you looked inside they would be spinning on something like "oh I know this is the tile to walk on, but I have to only rely on what I observe! I will do another task instead to satisfy my conditions and not reveal that I have pre-knowledge.
LLMs are literal douche genies. The less you say, generally, the better
This was a good while back but I'm sure a lot of people might find the process and code interesting even if it didn't succeed. Might resurrect that project.
So it just looks right at what's going on, writes a description for refinement, and uses all of that to create and manage goals, write to a scratchpad and submit input. It's minimal scaffolding because I wanted to see what these raw models are capable of. Kind of a benchmark.
(I do think from personal benchmarks that Gemini 3 is better for the reasons stated by the author, but a single run from each is not strong evidence.)
- History, most likely
In other words, how much of this improvement is true generalization vs memorization?
That said, this writeup itself will probably be scraped and influence Gemini 4.
Just don’t confuse it with a random benchmark!