It is the first model to get partial-credit on an LLM image test I have. Which is counting the legs of a dog. Specifically, a dog with 5 legs. This is a wild test, because LLMs get really pushy and insistent that the dog only has 4 legs.
In fact GPT5 wrote an edge detection script to see where "golden dog feet" met "bright green grass" to prove to me that there were only 4 legs. The script found 5, and GPT-5 then said it was a bug, and adjusted the script sensitivity so it only located 4, lol.
Anyway, Gemini 3, while still being unable to count the legs first try, did identify "male anatomy" (it's own words) also visible in the picture. The 5th leg was approximately where you could expect a well endowed dog to have a "5th leg".
That aside though, I still wouldn't call it particularly impressive.
As a note, Meta's image slicer correctly highlighted all 5 legs without a hitch. Maybe not quite a transformer, but interesting that it could properly interpret "dog leg" and ID them. Also the dog with many legs (I have a few of them) all had there extra legs added by nano-banana.
I just tried to get Gemini to produce an image of a dog with 5 legs to test this out, and it really struggled with that. It either made a normal dog, or turned the tail into a weird appendage.
Then I asked both Gemini and Grok to count the legs, both kept saying 4.
Gemini just refused to consider it was actually wrong.
Grok seemed to have an existential crisis when I told it it was wrong, becoming convinced that I had given it an elaborate riddle. After thinking for an additional 2.5 minutes, it concluded:
"Oh, I see now—upon closer inspection, this is that famous optical illusion photo of a "headless" dog. It's actually a three-legged dog (due to an amputation), with its head turned all the way back to lick its side, which creates the bizarre perspective making it look decapitated at first glance. So, you're right; the dog has 3 legs."
You're right, this is a good test. Right when I'm starting to feel LLMs are intelligent.
If you want to see something rather amusing - instead of using the LLM aspect of Gemini 3.0 Pro, feed a five-legged dog directly into Nano Banana Pro and give it an editing task that requires an intrinsic understanding of the unusual anatomy.
Place sneakers on all of its legs.
It'll get this correct a surprising number of times (tested with BFL Flux2 Pro, and NB Pro).
I had no trouble getting it to generate an image of a five-legged dog first try, but I really was surprised at how badly it failed in telling me the number of legs when I asked it in a new context, showing it that image. It wrote a long defense of its reasoning and when pressed, made up demonstrably false excuses of why it might be getting the wrong answer while still maintaining the wrong answer.
Its not that they aren’t intelligent its that they have been RL’d like crazy to not do that
Its rather like as humans we are RL’d like crazy to be grossed out if we view a picture of a handsome man and beautiful woman kissing (after we are told they are brother and sister) -
Ie we all have trained biases - that we are told to follow and trained on - human art is about subverting those expectations
Why should I assume that a failure that looks like a model just doing fairly simple pattern matching "this is dog, dogs don't have 5 legs, anything else is irrelevant" vs more sophisticated feature counting of a concrete instance of an entity is RL vs just a prediction failure due to training data not containing a 5-legged dog and an inability to go outside-of-distribution?
RL has been used extensively in other areas - such as coding - to improve model behavior on out-of-distribution stuff, so I'm somewhat skeptical of handwaving away a critique of a model's sophistication by saying here it's RL's fault that it isn't doing well out-of-distribution.
If we don't start from a position of anthropomorphizing the model into a "reasoning" entity (and instead have our prior be "it is a black box that has been extensively trained to try to mimic logical reasoning") then the result seems to be "here is a case where it can't mimic reasoning well", which seems like a very realistic conclusion.
I have the same problem, people are trying so badly to come up with reasoning for it when there's just nothing like that there. It was trained on it and it finds stuff it was trained to find, if you go out of the training it gets lost, we expect it to get lost.
But it didn't actually think it had done so, aligning with your observations. The last bits of its thinking were pretty sad (for it):
Analyzing Missing Anomaly
I'm focusing on the discrepancy - the persistent absence of the fifth leg in the generated images. Despite multiple attempts, the output consistently depicts a four-legged dog, contrary to the explicit requirement. This ongoing failure necessitates a shift in approach.
Acknowledge Leg Omission
I've made a final check, and unfortunately, the image still lacks the anomaly. The consistent absence of the fifth leg necessitates admitting my limitation. I'll provide the best image I could generate, clearly stating the missing feature.
I don’t know much about AI, but I have this image test that everything has failed at. You basically just present an image of a maze and ask the LLM to draw a line through the most optimal path.
I just oneshot it with claude code (opus 4.5) using this prompt. It took about 5 mins and included detecting that it was cheating at first (drew a line around the boundary of the maze instead), so it added guardrails for that:
```
Create a devenv project that does the following:
- Read the image at maze.jpg
- Write a script that solves the maze in the most optimal way between the mouse and the cheese
- Generate a new image which is of the original maze, but with a red line that represents the calculated path
In fact, one of the tests I use as part of GenAI Showdown involves both parts of the puzzle: draw a maze with a clearly defined entrance and exit, along with a dashed line indicating the solution to the maze.
Only one model (gpt-image-1) out of the 18 tested managed to pass the test successfully. Gemini 3.0 Pro got VERY close.
super cool! Interesting note about Seedream 4 - do you think awareness of A* actually could improve the outcome? Like I said, I'm no AI expert, so my intuitions are pretty bad, but I'd suspect that image analysis + algorithmic pathfinding don't have much crossover in terms of training capabilities. But I could be wrong!
Great question. I do wish we had a bit more insight into the exact background "thinking" that was happening on systems like Seedream.
When you think about posing the "solve a visual image of a maze" to something like ChatGPT, there's a good chance it'll try to throw a python VM at it, threshold it with something like OpenCV, and use a shortest-path style algorithm to try and solve it.
I have also tried the maze from a photo test a few times and never seen a one-shot success. But yesterday I was determined to succeed so I allowed Gemini 3 to write a python gui app that takes in photos of physical mazes (I have a bunch of 3d printed ones) and find the path. This does work.
Gemini 3 then one-shot ported the whole thing (which uses CV py libraries) to a single page html+js version which works just as well.
I gave that to Claude to assess and assign a FAANG hiring level to, and it was amazed and said Gemini 3 codes like an L6.
Since I work for Google and used my phone in the office to do this, I think I can't share the source or file.
Honestly, even though it failed, I'm kind of impressed that the trajectory mostly stays in the lines. If you remove all but two openings, does it work? The drawing you show has more than two openings, some of which are inaccessible from the inside of the maze.
The reason is that image generators don't iterate on the output in the same way the text-based LLMs do. Essentially they produce the image in "one hit" and can't solve a complex sequence in the same way you couldn't one-shot this either. Try taking a random maze, glance at it, then go off to draw a squiggle on a transparency. If you were to place that on top of the maze, there's virtually no chance that you'd have found the solution on the first try.
That's essentially what's going on with AI models, they're struggling because they only get "one step" to solve the problem instead of being able to trace through the maze slowly.
An interesting experiment would be to ask the AI to incrementally solve the maze. Ask it to draw a line starting at the entrance a little ways into the maze, then a little bit further, etc... until it gets to the end.
It always feels to me like these types of tests are being somewhat intentionally ignorant of how LLM cognition differs from human cognition. To me, they don't really "prove" or "show" anything other than simply - LLMs thinking works different than human thinking.
I'm always curious if these tests have comprehensive prompts that inform the model about what's going on properly, or if they're designed to "trick" the LLM in a very human-cognition-centric flavor of "trick".
Does the test instruction prompt tell it that it should be interpreting the image very, very literally, and that it should attempt to discard all previous knowledge of the subject before making its assessment of the question, etc.? Does it tell the model that some inputs may be designed to "trick" its reasoning, and to watch out for that specifically?
More specifically, what is a successful outcome here to you? Simply returning the answer "5" with no other info, or back-and-forth, or anything else in the output context? What is your idea of the LLMs internal world-model in this case? Do you want it to successfully infer that you are being deceitful? Should it respond directly to the deceit? Should it take the deceit in "good faith" and operate as if that's the new reality? Something in between? To me, all of this is very unclear in terms of LLM prompting, it feels like there's tons of very human-like subtext involved and you're trying to show that LLMs can't handle subtext/deceit and then generalizing that to say LLMs have low cognitive abilities in a general sense? This doesn't seem like particularly useful or productive analysis to me, so I'm curious what the goal of these "tests" are for the people who write/perform/post them?
I thought adversarial testing like this was a routine part of software engineering. He's checking to see how flexible it is. Maybe prompting would help, but it would be cool if it was more flexible.
So the idea is what? What's the successful outcome look like for this test, in your mind? What should good software do? Respond and say there are 5 legs? Or question what kind of dog this even is? Or get confused by a nonsensical picture that doesn't quite match the prompt in a confusing way? Should it understand the concept of a dog and be able to tell you that this isn't a real dog?
This is the first time I hear the term LLM cognition and I am horrified.
LLMs don‘t have cognition. LLMs are a statistical inference machines which predict a given output given some input. There are no mental processes, no sensory information, and certainly no knowledge involved, only statistical reasoning, inference, interpolation, and prediction. Comparing the human mind to an LLM model is like comparing a rubber tire to a calf muscle, or a hydraulic system to the gravitational force. They belong in different categories and cannot be responsibly compared.
When I see these tests, I presume they are made to demonstrate the limitation of this technology. This is both relevant and important that consumers know they are not dealing with magic, and are not being sold a lie (in a healthy economy a consumer protection agency should ideally do that for us; but here we are).
Categories of _what_, exactly? What word would you use to describe this "kind" of which LLMs and humans are two very different "categories"? I simply chose the word "cognition". I think you're getting hung up on semantics here a bit more than is reasonable.
Precisely. At least apples and oranges are both fruits, and it makes sense to compare e.g. the sugar contents of each. But an LLM model and the human brain are as different as the wind and the sunshine. You cannot measure the windspeed of the sun and you cannot measure the UV index of the wind.
Your choice of the words here was rather poor in my opinion. Statistical models do not have cognition any more than the wind has ultra-violet radiation. Cognition is a well studied phenomena, there is a whole field of science dedicated to cognition. And while cognition of animals are often modeled using statistics, statistical models in them selves do not have cognition.
A much better word here would by “abilities”. That is that these tests demonstrate the different abilities of LLM models compared to human abilities (or even the abilities of traditional [specialized] models which often do pass these kinds of tests).
Semantics often do matter, and what worries me is that these statistical models are being anthropomorphized way more then is healthy. People treat them like the crew of the Enterprise treated Data, when in fact they should be treated like the ship‘s computer. And I think this because of a deliberate (and malicious/consumer hostile) marketing campaign from the AI companies.
Human legs and car tires can both take a human and a car respectively to the finish line of a 200 meter track course, the car tires do so considerably quicker than a pair of human legs. But nobody needs to describe the tire‘s running abilities because of that, nor even compare a tire to a leg. A car tire cannot run, and it is silly to demand an explanation for it.
Claude said there were 3 hands and 16 fingers.
GPT said there are 10 fingers. Grok impressively said "There are 9 fingers visible on these two hands (the left hand is missing the tip of its ring finger)."
Gemini smashed it and said 12.
I just re-ran that image through Gemini 3.0 Pro via AI Studio and it reported:
I've moved on to the right hand, meticulously tagging each finger. After completing the initial count of five digits, I noticed a sixth! There appears to be an extra digit on the far right. This is an unexpected finding, and I have counted it as well. That makes a total of eleven fingers in the image.
This right HERE is the issue. It's not nearly deterministic enough to rely on.
Thanks for that. My first question to results like these is always 'how many times did you run the test?'. N=1 tells us nothing. N=2 tells us something.
I wonder if a lot of these models are large language models that have had image recognition and generation tools bolted on? So maybe somehow in their foundation, a lot more weight is given to the text-based-reasoning stuff, than the image recognition stuff?
Anything that needs to overcome concepts which are disproportionately represented in the training data is going to give these models a hard time.
Try generating:
- A spider missing one leg
- A 9-pointed star
- A 5-leaf clover
- A man with six fingers on his left hand and four fingers on his right
You'll be lucky to get a 25% success rate.
The last one is particularly ironic given how much work went into FIXING the old SD 1.5 issues with hand anatomy... to the point where I'm seriously considering incorporating it as a new test scenario on GenAI Showdown.
Some good examples there. The octopus one is at an angle - can't really call that one pass (unless the goal is "VISIBLE" tentacles).
Other than the five-leaf clover, most of the images (dog, spider, person's hands) all required a human in the loop to invoke the "Image-to-Image" capabilities of NB Pro after it got them wrong. That's a bit different since you're actively correcting them.
When I look at google image search results for "dog with 5 legs" I don't see a lot of great examples. The first unequivocal "dog with 5 legs" was an illustration. Here was my conversation with Chat GPT.
> How many legs does this dog have?
"The dog in the image has four legs."
> look closer.
" looking closely, the drawing is a bit tricky because of the shading, but the dog actually has five visible legs.
Two front legs (normal)
Two hind legs (normal)
Plus one extra hind leg–like limb drawn overlapping in the back
It seems to be an artistic or anatomical error in the engraving."
Sounds like they used GenAI to make them. The "Editor" models (Seedream, Nano-Banana) can easily integrate a fifth limb to create the "dog with awkward walking animation".
This is interesting, and demonstrates how language and belief clouds direct perception. Now I'm wondering what's the LLM equivalent of opening the doors of perception ;)
And just like that, you no longer have a good benchmark. Scrapers / AI developers will read this comment, and add 5-legged dogs to LLM's training data.
This is exactly why I believe LLMs are a technological dead end. Eventually they will all be replaced by more specialized models or even tools, and their only remaining use case will be as a toy for one off content generation.
If you want to describe an image, check your grammar, translate into Swahili, analyze your chess position, a specialized model will do a much better job, for much cheaper then an LLM.
I do some electrical drafting work for construction and throw basic tasks at LLMs.
I gave it a shitty harness and it almost 1 shotted laying out outlets in a room based on a shitty pdf. I think if I gave it better control it could do a huge portion of my coworkers jobs very soon
I've been using pyrevit inside revit so I just threw a basic loop in there. There's already a building model and the coworkers are just placing and wiring outlets, switches, etc. The harness wasn't impressive enough to share (alos contains vibe coded UI since I didn't want to learn XAML stuff on a friday night). Nothing fancy; I'm not very skilled (I work in construction)
I gave it some custom methods it could call, including "get_available_families", "place family instance", "scan_geometry" (reads model walls into LLM by wall endpoint), and "get_view_scale".
The task is basically copy the building engineer's layout onto the architect model by placing my families. It requires reading the symbol list, and you give it a pdf that contains the room.
Notably, it even used a GFCI family when it noticed it was a bathroom (I had told it to check NEC code, implying outlet spacing).
The turing test is still a thing. No llm could pass for a person for more than a couple minutes of chatting. That’s a world of difference compared to a decade ago, but I would emphatically not call that “passing the turing test”
Also, none of the other things you mentioned have actually happened. Don’t really know why I bother responding to this stuff
> No llm could pass for a person for more than a couple minutes of chatting
I strongly doubt this. If you gave it an appropriate system prompt with instructions and examples on how to speak in a certain way (something different from typical slop, like the way a teenager chats on discord or something), I'm quite sure it could fool the majority of people
> Remember when the Turing test was a thing? No one seems to remember it was considered serious in 2020
To be clear, it's only ever been a pop science belief that the Turing test was proposed as a literal benchmark. E.g. Chomsky in 1995 wrote:
The question “Can machines think?” is not a question of fact but one of language, and Turing himself observed that the question is 'too meaningless to deserve discussion'.
The Turing test is a literal benchmark. Its purpose was to replace an ill-posed question (what does it mean to ask if a machine could "think", when we don't know ourselves what this means- and given that the subjective experience of the machine is unknowable in any case) with a question about the product of this process we call "thinking". That is, if a machine can satisfactorily imitate the output of a human brain, then what it does is at least equivalent to thinking.
"I believe that in about fifty years'
time it will be possible, to programme computers, with a storage capacity of about 10^9, to
make them play the imitation game so well that an average interrogator will not have
more than 70 per cent chance of making the right identification after five minutes of
questioning. The original question, "Can machines think?" I believe to be too
meaningless to deserve discussion. Nevertheless I believe that at the end of the century
the use of words and general educated opinion will have altered so much that one will be
able to speak of machines thinking without expecting to be contradicted."
I'm double replying to you since the replies are disparate subthreads. This is the necessary step so the robots who can turn wrenches know how to turn them. Those are near useless without perfect automated models.
Anything like this willl have trouble getting adopted since you'd need these to work with imperfect humans, which becomes way harder. You could bankroll a whole team of subcontractors (e.g. all trades) using that, but you would have one big liability.
The upper end of the complexity is similar to EDA in difficulty, imo. Complete with "use other layers for routing" problems.
I feel safer here than in programming. The senior guys won't be automated out any time soon, but I worry for Indian drafting firms without trade knowledge; the handholding I give them might go to an LLM soon.
Going to compare this to our current solution of Amazon's Textract service for analyzing handwritten datasheets. Textract, when extracting tables (which is what we use it for) does not allow for providing any context or information about the tables and what we expect them to contain, but it is really good at correctly recognizing hand written characters. All of my attempts at less specialized, more general models allow me to provide that context, which is helpful in some ways, but fail at the basic part of almost always correctly getting the character.
These OCR improvements will almost certainly be brought to google books, which is great. Long term it can enable compressing all non-digital rare books into a manageable size that can be stored for less than $5,000.[0] It would also be great for archive.org to move to this from Tesseract. I wonder what the cost would be, both in raw cost to run, and via a paid API, to do that.
I was surprised at how poorly GPT-5 did in comparison to Opus 4.1 and Gemini 2.5 on a pretty simple OCR task a few months ago - I should run that again against the latest models and see how they do. https://simonwillison.net/2025/Aug/29/the-perils-of-vibe-cod...
I found much better results with smallish UI elements in large screenshots on GPT by slicing it up manually and feeding them one at a time. I think it does severely lossy downscaling.
Love how employee portals for many companies essentially never get updated design wise over the decades, lol. That page styling and the balls certainly take me back.
I used to work for a company where the SSO screen had a nice corporate happy people at the office type of image. 25mb. I was in Brazil on a crappy roaming 2g service and couldn't login at all. I know most of the work happens on desktop but geee.....
Oh speaking on mobile, I remember when I tried to use Jira mobile web to move a few tickets up on priority by drag and dropping and ended up closing the Sprint. That stuff was horrible.
> Pointing capability: Gemini 3 has the ability to point at specific locations in images by outputting pixel-precise coordinates. Sequences of 2D points can be strung together to perform complex tasks, such as estimating human poses or reflecting trajectories over time
Does somebody know how to correctly prompt the model for these tasks or even better provide some docs? The pictures with the pretty markers are appreciated but that section is a bit vague and without references
For my CMS I’d love to get an AI to nicely frame a picture in certain aspect ratios. Like of I provide an image, give me coordinates for a widescreen, square, portrait, and 4x3 using a photographers eye.
Any model that can do that? I tried looking in huggingface but didn’t quite see anything.
Since I think it's interesting to highlight the jagged intelligence, I have a simple word search puzzle [0] that Nano Banana Pro stills struggles to solve correctly. Gemini 3 Pro with Code Execution is able to one-shot the problem and find the positions of each word (this is super impressive! one year ago it wasn't possible), but Nano Banana Pro fails to highlight the words correctly.
Here's the output from two tests I ran:
1. Asking Nano Banana Pro to solve the word search puzzle directly [1].
2. Asking Nano Banana Pro to highlight each word on the grid, with the position of every word included as part of the prompt [2].
The fact that it gets 2 words correct demonstrates meaningful progress, and it seems like we're really close to having a model that can one-shot this problem soon.
There's actually a bit of nuance required to solve this puzzle correctly which an older Gemini model struggled to do without additional nudging. You have to convert the grid or word list to use matching casing (the grid uses uppercase, the word list uses lowercase), and you need to recognize that "soup mix" needs to have the space removed when doing the search.
Gemini 3 Pro is not Nano Banana Pro, and the image generation/model that decodes the generated image tokens may not be as robust.
The thinking step of Nano Banana Pro can refine some lateral steps (i.e. the errors in the homework correction and where they are spatially in the image) but it isn't perfect and can encounter some of the typical pitfalls. It's a lot better than Nano Banana base, though.
I actually did this prompt and found that it worked with a single nudge on a followup prompt. My first shot got me a wine glass that was almost full but not quite. I told it I wanted it full to the top - another drop would overflow. The second shot was perfectly full.
Please describe what happening in each scene of this video.
List scenes with timestamp, then describe separately:
- Setup and background, colors
- What is moving, what appear
- What objects in this scene and what is happening,
Basically make desceiption of 5 minutes video for a person who cant watch it.
I'm playing with this and wondering if this is an actually good way to identify dominant colors and other features of a garment/product when using a photo where the item is styled and not isolated from the model or other garments
Interesting. When i asked Gemini 3 Pro to generate a Infographic from my personal accounting sheet, it first failed to generate anything except a black background, then it generated something where it mixed different languages in a non-sensical way, with obvious typos and irrelevant information grouping. It's certainly a leap forward in OCR, rendering classic OCR useless.
Gemini 3 Pro's text encoder powers Nano Banana Pro, but it has its own image decoding model that decodes the generated image tokens into an actual image, which appears to be the more pertinent issue in this case.
Frankly, it's insane how laughably bad under scrutiny their own examples are. It both distorted the data and made the chart less readable (labels placement, segments separation, missing labels, worse contrast). And it combined them into one, so you you'll have harder time comparing them compared to the original image! Isn't it amazing that it added a toggle? Post author seems to think it deserves an exclamation point even.
I'm really fascinate by the opportunities to analyze videos. The amount of tokens it compresses down to, and what you can reason across those tokens, is incredible.
That is because it isn't actually tokens that are fed into the model for non-text. For text, it is tokenized, and each token has a specific set of vectors. But with other media, they've trained encoders that analyze the media and produce a set of vectors that are the same "format" as the token's vectors, but it isn't actually ever a token.
Most companies have rules for how many tokens the media should "cost", but they aren't usually exact.
Yeah the "High frame rate understanding" feature caught my eye, actual real time analysis of live video feeds seems really cool. Also wondering what they mean by "video reasoning/thinking"?
The document is paints a super impressive picture, but the core constraint of “network connection to Google required so we can harvest your data” is still a big showstopper for me (and all cloud-based AI tooling, really).
I’d be curious to see how well something like this can be distilled down for isolated acceleration on SBCs or consumer kit, because that’s where the billions to be made reside (factories, remote sites, dangerous or sensitive facilities, etc).
People with your concerns probably make up 1% of the market if that. Also I don’t upload stuff I’m worried about Google seeing. I wonder if they will allows special plans for corporations
I’m very curious where you get that number from, because I thought the same thing until I got a job inside that market and realized how much more vast it actually is. The revenue numbers might not be as big as Big Tech, but the product market is shockingly vast. My advice is not to confuse Big Tech revenues for total market size, because they bring in such revenue by catering to everyone, rather than specific segments or niches; a McDonald’s will always do more volume than a steakhouse, but it doesn’t mean the market for steakhouses is small enough to ignore.
As for this throwaway line:
> Also I don’t upload stuff I’m worried about Google seeing.
You do realize that these companies harvest even private data, right? Like, even in places you think you own, or that you pay for, they’re mining for revenue opportunities and using you as the product even when you’re a customer, right?
> I wonder if they will allows special plans for corporations
They do, but no matter how much redlining Legal does to protect IP interests, the consensus I keep hearing is “don’t put private or sensitive corporate data into third-parties because no legal agreement will sufficiently protect us from harm if they steal our IP or data”. Just look at the glut of lawsuits against Apple, Google, Microsoft, etc from smaller companies that trusted them to act in good faith but got burned for evidence that you cannot trust these entities.
Special since Trump, which non-US company should trust and invest know-how to an us company. And then are also governments. Also special since Trump, is way to risky to send any data to an us company.
Arpanet was supposed to be decentralized. Now everyone wants to centralize everything so in a war it is sufficient to strike 100 data centers and the whole tethered economy collapses.
That is called progress.
EDIT: You can downvote the truth but still no one wants your "AI" slop.
Ah, the fond memories of telnetting to NCSA to upload the raw HTML of my first website, written on an OG Macintosh computer and ported via floppy to a PowerMac for network connectivity.
im realizing how much of a bottleneck vision models are
im just a glorified speedreadin' promptin' QA at this point with codex
once it replaces the QA layer its truly over for software dev jobs
future would be a software genie where on aistudio you type: "go make counterstrike 1.6 clone, here is $500, you have two hours"
edit: saw the Screenspot benchmark and holy ** this is an insane jump!!! 11% to 71% even beating Opus 4.5's 50%...chatgpt is at 3.5% and it matches my experience with codex
> once it replaces the QA layer its truly over for software dev jobs
Maybe. However, with CYA requirements being everywhere in industry, there would have to be 100 waiver forms signed. I-promise-not-to-sue-company-if-AI-deletes-the-entire-database
It won't happen for that reason alone. Oh who am I kidding of course it will
Well
It is the first model to get partial-credit on an LLM image test I have. Which is counting the legs of a dog. Specifically, a dog with 5 legs. This is a wild test, because LLMs get really pushy and insistent that the dog only has 4 legs.
In fact GPT5 wrote an edge detection script to see where "golden dog feet" met "bright green grass" to prove to me that there were only 4 legs. The script found 5, and GPT-5 then said it was a bug, and adjusted the script sensitivity so it only located 4, lol.
Anyway, Gemini 3, while still being unable to count the legs first try, did identify "male anatomy" (it's own words) also visible in the picture. The 5th leg was approximately where you could expect a well endowed dog to have a "5th leg".
That aside though, I still wouldn't call it particularly impressive.
As a note, Meta's image slicer correctly highlighted all 5 legs without a hitch. Maybe not quite a transformer, but interesting that it could properly interpret "dog leg" and ID them. Also the dog with many legs (I have a few of them) all had there extra legs added by nano-banana.
I just tried to get Gemini to produce an image of a dog with 5 legs to test this out, and it really struggled with that. It either made a normal dog, or turned the tail into a weird appendage.
Then I asked both Gemini and Grok to count the legs, both kept saying 4.
Gemini just refused to consider it was actually wrong.
Grok seemed to have an existential crisis when I told it it was wrong, becoming convinced that I had given it an elaborate riddle. After thinking for an additional 2.5 minutes, it concluded: "Oh, I see now—upon closer inspection, this is that famous optical illusion photo of a "headless" dog. It's actually a three-legged dog (due to an amputation), with its head turned all the way back to lick its side, which creates the bizarre perspective making it look decapitated at first glance. So, you're right; the dog has 3 legs."
You're right, this is a good test. Right when I'm starting to feel LLMs are intelligent.
If you want to see something rather amusing - instead of using the LLM aspect of Gemini 3.0 Pro, feed a five-legged dog directly into Nano Banana Pro and give it an editing task that requires an intrinsic understanding of the unusual anatomy.
It'll get this correct a surprising number of times (tested with BFL Flux2 Pro, and NB Pro).https://imgur.com/a/wXQskhL
I had no trouble getting it to generate an image of a five-legged dog first try, but I really was surprised at how badly it failed in telling me the number of legs when I asked it in a new context, showing it that image. It wrote a long defense of its reasoning and when pressed, made up demonstrably false excuses of why it might be getting the wrong answer while still maintaining the wrong answer.
Yeah it gave me the 5-legged dog on the 4th or 5th try.
Its not that they aren’t intelligent its that they have been RL’d like crazy to not do that
Its rather like as humans we are RL’d like crazy to be grossed out if we view a picture of a handsome man and beautiful woman kissing (after we are told they are brother and sister) -
Ie we all have trained biases - that we are told to follow and trained on - human art is about subverting those expectations
Why should I assume that a failure that looks like a model just doing fairly simple pattern matching "this is dog, dogs don't have 5 legs, anything else is irrelevant" vs more sophisticated feature counting of a concrete instance of an entity is RL vs just a prediction failure due to training data not containing a 5-legged dog and an inability to go outside-of-distribution?
RL has been used extensively in other areas - such as coding - to improve model behavior on out-of-distribution stuff, so I'm somewhat skeptical of handwaving away a critique of a model's sophistication by saying here it's RL's fault that it isn't doing well out-of-distribution.
If we don't start from a position of anthropomorphizing the model into a "reasoning" entity (and instead have our prior be "it is a black box that has been extensively trained to try to mimic logical reasoning") then the result seems to be "here is a case where it can't mimic reasoning well", which seems like a very realistic conclusion.
I’m inclined to buy the RL story, since the image gen “deep dream” models of ~10 years ago would produce dogs with TRILLIONS of eyes: https://doorofperception.com/2015/10/google-deep-dream-incep...
I have the same problem, people are trying so badly to come up with reasoning for it when there's just nothing like that there. It was trained on it and it finds stuff it was trained to find, if you go out of the training it gets lost, we expect it to get lost.
Isn't this proof that LLMs still don't really generalize beyond their training data?
I wonder how they would behave given a system prompt that asserts "dogs may have more or less than four legs".
They do, but we call it "hallucination" when that happens.
Kind of feels that way
Nano Banana 2 actually did generate a pretty good looking 5-legged dog for me.
https://gemini.google.com/share/e7a8b902ff67
But it didn't actually think it had done so, aligning with your observations. The last bits of its thinking were pretty sad (for it):
Analyzing Missing Anomaly
I'm focusing on the discrepancy - the persistent absence of the fifth leg in the generated images. Despite multiple attempts, the output consistently depicts a four-legged dog, contrary to the explicit requirement. This ongoing failure necessitates a shift in approach.
Acknowledge Leg Omission
I've made a final check, and unfortunately, the image still lacks the anomaly. The consistent absence of the fifth leg necessitates admitting my limitation. I'll provide the best image I could generate, clearly stating the missing feature.
I don’t know much about AI, but I have this image test that everything has failed at. You basically just present an image of a maze and ask the LLM to draw a line through the most optimal path.
Here’s how Nano Banana fared: https://x.com/danielvaughn/status/1971640520176029704?s=46
I just oneshot it with claude code (opus 4.5) using this prompt. It took about 5 mins and included detecting that it was cheating at first (drew a line around the boundary of the maze instead), so it added guardrails for that:
```
Create a devenv project that does the following:
Use whatever lib/framework is most appropriate```
If you allow tool use much simpler models can solve it.
In fact, one of the tests I use as part of GenAI Showdown involves both parts of the puzzle: draw a maze with a clearly defined entrance and exit, along with a dashed line indicating the solution to the maze.
Only one model (gpt-image-1) out of the 18 tested managed to pass the test successfully. Gemini 3.0 Pro got VERY close.
https://genai-showdown.specr.net/#the-labyrinth
super cool! Interesting note about Seedream 4 - do you think awareness of A* actually could improve the outcome? Like I said, I'm no AI expert, so my intuitions are pretty bad, but I'd suspect that image analysis + algorithmic pathfinding don't have much crossover in terms of training capabilities. But I could be wrong!
Great question. I do wish we had a bit more insight into the exact background "thinking" that was happening on systems like Seedream.
When you think about posing the "solve a visual image of a maze" to something like ChatGPT, there's a good chance it'll try to throw a python VM at it, threshold it with something like OpenCV, and use a shortest-path style algorithm to try and solve it.
I have also tried the maze from a photo test a few times and never seen a one-shot success. But yesterday I was determined to succeed so I allowed Gemini 3 to write a python gui app that takes in photos of physical mazes (I have a bunch of 3d printed ones) and find the path. This does work.
Gemini 3 then one-shot ported the whole thing (which uses CV py libraries) to a single page html+js version which works just as well.
I gave that to Claude to assess and assign a FAANG hiring level to, and it was amazed and said Gemini 3 codes like an L6.
Since I work for Google and used my phone in the office to do this, I think I can't share the source or file.
I tried this with Claude:
``` > [Image #1] Create a unicode "ascii-art" version of this image, with the optimal path through the maze highlighted in a solid colour.
I'll create an ASCII art version of this maze with the solution path highlighted!
```Suffice to say, it didn't do either part right.
>Suffice to say, it didn't do either part right.
I dunno why people are surprised by this. This is what you get with text->text. Reasoning doesn't work text->text.
Honestly, even though it failed, I'm kind of impressed that the trajectory mostly stays in the lines. If you remove all but two openings, does it work? The drawing you show has more than two openings, some of which are inaccessible from the inside of the maze.
That might be an interesting training set, a bunch of mazes…
The reason is that image generators don't iterate on the output in the same way the text-based LLMs do. Essentially they produce the image in "one hit" and can't solve a complex sequence in the same way you couldn't one-shot this either. Try taking a random maze, glance at it, then go off to draw a squiggle on a transparency. If you were to place that on top of the maze, there's virtually no chance that you'd have found the solution on the first try.
That's essentially what's going on with AI models, they're struggling because they only get "one step" to solve the problem instead of being able to trace through the maze slowly.
An interesting experiment would be to ask the AI to incrementally solve the maze. Ask it to draw a line starting at the entrance a little ways into the maze, then a little bit further, etc... until it gets to the end.
It always feels to me like these types of tests are being somewhat intentionally ignorant of how LLM cognition differs from human cognition. To me, they don't really "prove" or "show" anything other than simply - LLMs thinking works different than human thinking.
I'm always curious if these tests have comprehensive prompts that inform the model about what's going on properly, or if they're designed to "trick" the LLM in a very human-cognition-centric flavor of "trick".
Does the test instruction prompt tell it that it should be interpreting the image very, very literally, and that it should attempt to discard all previous knowledge of the subject before making its assessment of the question, etc.? Does it tell the model that some inputs may be designed to "trick" its reasoning, and to watch out for that specifically?
More specifically, what is a successful outcome here to you? Simply returning the answer "5" with no other info, or back-and-forth, or anything else in the output context? What is your idea of the LLMs internal world-model in this case? Do you want it to successfully infer that you are being deceitful? Should it respond directly to the deceit? Should it take the deceit in "good faith" and operate as if that's the new reality? Something in between? To me, all of this is very unclear in terms of LLM prompting, it feels like there's tons of very human-like subtext involved and you're trying to show that LLMs can't handle subtext/deceit and then generalizing that to say LLMs have low cognitive abilities in a general sense? This doesn't seem like particularly useful or productive analysis to me, so I'm curious what the goal of these "tests" are for the people who write/perform/post them?
The marketing of these products is intentionally ignorant of how LLM cognition differs from human cognition.
Let's not say that the people being deceptive are the people who've spotted ways that that is untrue...
I thought adversarial testing like this was a routine part of software engineering. He's checking to see how flexible it is. Maybe prompting would help, but it would be cool if it was more flexible.
So the idea is what? What's the successful outcome look like for this test, in your mind? What should good software do? Respond and say there are 5 legs? Or question what kind of dog this even is? Or get confused by a nonsensical picture that doesn't quite match the prompt in a confusing way? Should it understand the concept of a dog and be able to tell you that this isn't a real dog?
No, it’s just a test case to demonstrate flexibility when faced with unusual circumstances
This is the first time I hear the term LLM cognition and I am horrified.
LLMs don‘t have cognition. LLMs are a statistical inference machines which predict a given output given some input. There are no mental processes, no sensory information, and certainly no knowledge involved, only statistical reasoning, inference, interpolation, and prediction. Comparing the human mind to an LLM model is like comparing a rubber tire to a calf muscle, or a hydraulic system to the gravitational force. They belong in different categories and cannot be responsibly compared.
When I see these tests, I presume they are made to demonstrate the limitation of this technology. This is both relevant and important that consumers know they are not dealing with magic, and are not being sold a lie (in a healthy economy a consumer protection agency should ideally do that for us; but here we are).
>They belong in different categories
Categories of _what_, exactly? What word would you use to describe this "kind" of which LLMs and humans are two very different "categories"? I simply chose the word "cognition". I think you're getting hung up on semantics here a bit more than is reasonable.
> Categories of _what_, exactly?
Precisely. At least apples and oranges are both fruits, and it makes sense to compare e.g. the sugar contents of each. But an LLM model and the human brain are as different as the wind and the sunshine. You cannot measure the windspeed of the sun and you cannot measure the UV index of the wind.
Your choice of the words here was rather poor in my opinion. Statistical models do not have cognition any more than the wind has ultra-violet radiation. Cognition is a well studied phenomena, there is a whole field of science dedicated to cognition. And while cognition of animals are often modeled using statistics, statistical models in them selves do not have cognition.
A much better word here would by “abilities”. That is that these tests demonstrate the different abilities of LLM models compared to human abilities (or even the abilities of traditional [specialized] models which often do pass these kinds of tests).
Semantics often do matter, and what worries me is that these statistical models are being anthropomorphized way more then is healthy. People treat them like the crew of the Enterprise treated Data, when in fact they should be treated like the ship‘s computer. And I think this because of a deliberate (and malicious/consumer hostile) marketing campaign from the AI companies.
You'll need to explain the IMO results, then.
Human legs and car tires can both take a human and a car respectively to the finish line of a 200 meter track course, the car tires do so considerably quicker than a pair of human legs. But nobody needs to describe the tire‘s running abilities because of that, nor even compare a tire to a leg. A car tire cannot run, and it is silly to demand an explanation for it.
Super interesting. I replicated this.
I passed the AIs this image and asked them how many fingers were on the hands: https://media.post.rvohealth.io/wp-content/uploads/sites/3/2...
Claude said there were 3 hands and 16 fingers. GPT said there are 10 fingers. Grok impressively said "There are 9 fingers visible on these two hands (the left hand is missing the tip of its ring finger)." Gemini smashed it and said 12.
I just re-ran that image through Gemini 3.0 Pro via AI Studio and it reported:
This right HERE is the issue. It's not nearly deterministic enough to rely on.Thanks for that. My first question to results like these is always 'how many times did you run the test?'. N=1 tells us nothing. N=2 tells us something.
Naive question, but what is Gemini?
I wonder if a lot of these models are large language models that have had image recognition and generation tools bolted on? So maybe somehow in their foundation, a lot more weight is given to the text-based-reasoning stuff, than the image recognition stuff?
Anything that needs to overcome concepts which are disproportionately represented in the training data is going to give these models a hard time.
Try generating:
- A spider missing one leg
- A 9-pointed star
- A 5-leaf clover
- A man with six fingers on his left hand and four fingers on his right
You'll be lucky to get a 25% success rate.
The last one is particularly ironic given how much work went into FIXING the old SD 1.5 issues with hand anatomy... to the point where I'm seriously considering incorporating it as a new test scenario on GenAI Showdown.
https://gemini.google.com/share/8cef4b408a0a
Surprisingly, it got all of them right
Some good examples there. The octopus one is at an angle - can't really call that one pass (unless the goal is "VISIBLE" tentacles).
Other than the five-leaf clover, most of the images (dog, spider, person's hands) all required a human in the loop to invoke the "Image-to-Image" capabilities of NB Pro after it got them wrong. That's a bit different since you're actively correcting them.
What image are you using?
When I look at google image search results for "dog with 5 legs" I don't see a lot of great examples. The first unequivocal "dog with 5 legs" was an illustration. Here was my conversation with Chat GPT.
> How many legs does this dog have?
"The dog in the image has four legs."
> look closer.
" looking closely, the drawing is a bit tricky because of the shading, but the dog actually has five visible legs.
Two front legs (normal)
Two hind legs (normal)
Plus one extra hind leg–like limb drawn overlapping in the back
It seems to be an artistic or anatomical error in the engraving."
Seems fair to me.
Sounds like they used GenAI to make them. The "Editor" models (Seedream, Nano-Banana) can easily integrate a fifth limb to create the "dog with awkward walking animation".
https://imgur.com/a/wXQskhL
This is interesting, and demonstrates how language and belief clouds direct perception. Now I'm wondering what's the LLM equivalent of opening the doors of perception ;)
And just like that, you no longer have a good benchmark. Scrapers / AI developers will read this comment, and add 5-legged dogs to LLM's training data.
That's okay. Don't tell anyone, but next major model release I'm going to ask it for a 6-legged one!
Could you link the image? Interesting stuff.
this is hilarious and incredibly interesting at the same time! thanks for writing it up.
This is exactly why I believe LLMs are a technological dead end. Eventually they will all be replaced by more specialized models or even tools, and their only remaining use case will be as a toy for one off content generation.
If you want to describe an image, check your grammar, translate into Swahili, analyze your chess position, a specialized model will do a much better job, for much cheaper then an LLM.
"There are FOUR legs!!!"
I do some electrical drafting work for construction and throw basic tasks at LLMs.
I gave it a shitty harness and it almost 1 shotted laying out outlets in a room based on a shitty pdf. I think if I gave it better control it could do a huge portion of my coworkers jobs very soon
Can you give an example of the sort of harness you used for that? Would love to play around with it
I've been using pyrevit inside revit so I just threw a basic loop in there. There's already a building model and the coworkers are just placing and wiring outlets, switches, etc. The harness wasn't impressive enough to share (alos contains vibe coded UI since I didn't want to learn XAML stuff on a friday night). Nothing fancy; I'm not very skilled (I work in construction)
I gave it some custom methods it could call, including "get_available_families", "place family instance", "scan_geometry" (reads model walls into LLM by wall endpoint), and "get_view_scale".
The task is basically copy the building engineer's layout onto the architect model by placing my families. It requires reading the symbol list, and you give it a pdf that contains the room.
Notably, it even used a GFCI family when it noticed it was a bathroom (I had told it to check NEC code, implying outlet spacing).
"AI could never replace the creativity of a human"
"Ok, I guess it could wipe out the economic demand for digital art, but it could never do all the autonomous tasks of a project manager"
"Ok, I guess it could automate most of that away but there will always be a need for a human engineer to steer it and deal with the nuances of code"
"Ok, well it could never automate blue collar work, how is it gonna wrench a pipe it doesn't have hands"
The goalposts will continue to move until we have no idea if the comments are real anymore.
Remember when the Turing test was a thing? No one seems to remember it was considered serious in 2020
The turing test is still a thing. No llm could pass for a person for more than a couple minutes of chatting. That’s a world of difference compared to a decade ago, but I would emphatically not call that “passing the turing test”
Also, none of the other things you mentioned have actually happened. Don’t really know why I bother responding to this stuff
> No llm could pass for a person for more than a couple minutes of chatting
I strongly doubt this. If you gave it an appropriate system prompt with instructions and examples on how to speak in a certain way (something different from typical slop, like the way a teenager chats on discord or something), I'm quite sure it could fool the majority of people
> Remember when the Turing test was a thing? No one seems to remember it was considered serious in 2020
To be clear, it's only ever been a pop science belief that the Turing test was proposed as a literal benchmark. E.g. Chomsky in 1995 wrote:
The Turing test is a literal benchmark. Its purpose was to replace an ill-posed question (what does it mean to ask if a machine could "think", when we don't know ourselves what this means- and given that the subjective experience of the machine is unknowable in any case) with a question about the product of this process we call "thinking". That is, if a machine can satisfactorily imitate the output of a human brain, then what it does is at least equivalent to thinking.
"I believe that in about fifty years' time it will be possible, to programme computers, with a storage capacity of about 10^9, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning. The original question, "Can machines think?" I believe to be too meaningless to deserve discussion. Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted."
> blue collar work
I don't think it's fair to qualify this as blue collar work
I'm double replying to you since the replies are disparate subthreads. This is the necessary step so the robots who can turn wrenches know how to turn them. Those are near useless without perfect automated models.
Anything like this willl have trouble getting adopted since you'd need these to work with imperfect humans, which becomes way harder. You could bankroll a whole team of subcontractors (e.g. all trades) using that, but you would have one big liability.
The upper end of the complexity is similar to EDA in difficulty, imo. Complete with "use other layers for routing" problems.
I feel safer here than in programming. The senior guys won't be automated out any time soon, but I worry for Indian drafting firms without trade knowledge; the handholding I give them might go to an LLM soon.
It is definitely not. Entry pay is 60k and the senior guys I know make about 200k in HCoL areas. A few wear white dress shirts every day.
Going to compare this to our current solution of Amazon's Textract service for analyzing handwritten datasheets. Textract, when extracting tables (which is what we use it for) does not allow for providing any context or information about the tables and what we expect them to contain, but it is really good at correctly recognizing hand written characters. All of my attempts at less specialized, more general models allow me to provide that context, which is helpful in some ways, but fail at the basic part of almost always correctly getting the character.
Hopefully Google pro marries the two together.
These OCR improvements will almost certainly be brought to google books, which is great. Long term it can enable compressing all non-digital rare books into a manageable size that can be stored for less than $5,000.[0] It would also be great for archive.org to move to this from Tesseract. I wonder what the cost would be, both in raw cost to run, and via a paid API, to do that.
[0] https://annas-archive.org/blog/critical-window.html
This is a really interesting "data flywheel" -- better model >> more usable data >> even better model
More Data for the Data Gods!
Interesting "ScreenSpot Pro" results:
ScreenSpot-Pro: GUI Grounding for Professional High-Resolution Computer Usehttps://arxiv.org/abs/2504.07981
I was surprised at how poorly GPT-5 did in comparison to Opus 4.1 and Gemini 2.5 on a pretty simple OCR task a few months ago - I should run that again against the latest models and see how they do. https://simonwillison.net/2025/Aug/29/the-perils-of-vibe-cod...
That is... astronomically different. Is GPT-5.1 downscaling and losing critical information or something? How could it be so different?
I found much better results with smallish UI elements in large screenshots on GPT by slicing it up manually and feeding them one at a time. I think it does severely lossy downscaling.
impressive.....most impressive
its going to reach low 90s very soon if trends continue
In case the article author sees this, the "HTML transcription" link is broken - it goes to https://aistudio-preprod.corp.google.com/prompts/1GUEWbLIlpX... which is a Google-employee-only URL.
Love how employee portals for many companies essentially never get updated design wise over the decades, lol. That page styling and the balls certainly take me back.
I used to work for a company where the SSO screen had a nice corporate happy people at the office type of image. 25mb. I was in Brazil on a crappy roaming 2g service and couldn't login at all. I know most of the work happens on desktop but geee.....
Oh speaking on mobile, I remember when I tried to use Jira mobile web to move a few tickets up on priority by drag and dropping and ended up closing the Sprint. That stuff was horrible.
Wow yeah. Flashbacks to when Gmail Invites were cool! Google too.
I’m a little surprised how open the help links are… I guess that if need help logging in you can’t be expected to well, log in.
hey, it's Rohan (the author of the article) - appreciate you catching this, we just fixed this!
Same with "See prompt in Google AI Studio" which links to an unpublished prompt in AI Studio.
> Pointing capability: Gemini 3 has the ability to point at specific locations in images by outputting pixel-precise coordinates. Sequences of 2D points can be strung together to perform complex tasks, such as estimating human poses or reflecting trajectories over time
Does somebody know how to correctly prompt the model for these tasks or even better provide some docs? The pictures with the pretty markers are appreciated but that section is a bit vague and without references
For my CMS I’d love to get an AI to nicely frame a picture in certain aspect ratios. Like of I provide an image, give me coordinates for a widescreen, square, portrait, and 4x3 using a photographers eye.
Any model that can do that? I tried looking in huggingface but didn’t quite see anything.
Since I think it's interesting to highlight the jagged intelligence, I have a simple word search puzzle [0] that Nano Banana Pro stills struggles to solve correctly. Gemini 3 Pro with Code Execution is able to one-shot the problem and find the positions of each word (this is super impressive! one year ago it wasn't possible), but Nano Banana Pro fails to highlight the words correctly.
Here's the output from two tests I ran:
1. Asking Nano Banana Pro to solve the word search puzzle directly [1].
2. Asking Nano Banana Pro to highlight each word on the grid, with the position of every word included as part of the prompt [2].
The fact that it gets 2 words correct demonstrates meaningful progress, and it seems like we're really close to having a model that can one-shot this problem soon.
There's actually a bit of nuance required to solve this puzzle correctly which an older Gemini model struggled to do without additional nudging. You have to convert the grid or word list to use matching casing (the grid uses uppercase, the word list uses lowercase), and you need to recognize that "soup mix" needs to have the space removed when doing the search.
[0] https://imgur.com/ekwfHrN
[1] https://imgur.com/1nybezU
[2] https://imgur.com/18mK5i5
"Gemini 3 Pro represents a generational leap from simple recognition to true visual and spatial reasoning."
Prompt: "wine glass full to the brim"
Image generated: 2/3 full wine glass.
True visual and spatial reasoning denied.
Gemini 3 Pro is not Nano Banana Pro, and the image generation/model that decodes the generated image tokens may not be as robust.
The thinking step of Nano Banana Pro can refine some lateral steps (i.e. the errors in the homework correction and where they are spatially in the image) but it isn't perfect and can encounter some of the typical pitfalls. It's a lot better than Nano Banana base, though.
As a consumer I typed this into "Gemini". The behind the scenes model selection just adds confusion.
If "AI" trust is the big barrier for widespread adoption to these products, Alphabet soup isn't the solution (pun intended).
Nano Banana generates images.
This article is about understanding images.
Your task is unrelated to the article.
I actually did this prompt and found that it worked with a single nudge on a followup prompt. My first shot got me a wine glass that was almost full but not quite. I told it I wanted it full to the top - another drop would overflow. The second shot was perfectly full.
The correction I expect to give to an intern, not a junior person.
did it return the exact same glass and surrounding imagery, just with more wine?
Audio described Youtube please? That'd be so amazing! Even if I couldn't play Zelda yet, I could listen to a playthrough with Gemini describing it.
Hey, I just made simple test on 5 minute downloaded YouTube video uploading it to Gemini app.
Source video title: Zelda: Breath of the Wild - Opening five minutes of gameplay
https://www.youtube.com/watch?v=xbt7ZYdUXn8
Prompt:
Result on github gist since there too much text:https://gist.github.com/ArseniyShestakov/43fe8b8c1dca45eadab...
I'd say thi is quite accurate.
Another example with completely random 10 minute benchmark video from Tears of Kingdom:
https://gist.github.com/ArseniyShestakov/47123ce2b6b19a8e6b3...
What’s new here? I believe this is just gemini 3 which was released last month (the model id hasn’t changed AFAICT)
Nothing new, it's just highlighting practical vision use cases.
I'm playing with this and wondering if this is an actually good way to identify dominant colors and other features of a garment/product when using a photo where the item is styled and not isolated from the model or other garments
Interesting. When i asked Gemini 3 Pro to generate a Infographic from my personal accounting sheet, it first failed to generate anything except a black background, then it generated something where it mixed different languages in a non-sensical way, with obvious typos and irrelevant information grouping. It's certainly a leap forward in OCR, rendering classic OCR useless.
That's more of an issue with Nano Banana Pro than with Gemini 3 Pro.
What's the difference? I thought the vision ai component of gemini 3 is called nano banana?
That’s about generating images, the other side is about understanding images.
i assumed nano banana was just a tool that gemini 3 used though i don't know
Gemini 3 Pro's text encoder powers Nano Banana Pro, but it has its own image decoding model that decodes the generated image tokens into an actual image, which appears to be the more pertinent issue in this case.
I would be interested in seeing what G3P makes of the Dead Sea Scrolls or similarly old documents.
Frankly, it's insane how laughably bad under scrutiny their own examples are. It both distorted the data and made the chart less readable (labels placement, segments separation, missing labels, worse contrast). And it combined them into one, so you you'll have harder time comparing them compared to the original image! Isn't it amazing that it added a toggle? Post author seems to think it deserves an exclamation point even.
I'm really fascinate by the opportunities to analyze videos. The amount of tokens it compresses down to, and what you can reason across those tokens, is incredible.
The actual token calculations with input videos for Gemini 3 Pro is...confusing.
https://ai.google.dev/gemini-api/docs/media-resolution
That is because it isn't actually tokens that are fed into the model for non-text. For text, it is tokenized, and each token has a specific set of vectors. But with other media, they've trained encoders that analyze the media and produce a set of vectors that are the same "format" as the token's vectors, but it isn't actually ever a token.
Most companies have rules for how many tokens the media should "cost", but they aren't usually exact.
When will we get Gemini 3 Flash?
Google really are a fully woken sleeping giant. More code reds being issued today I expect.
Okay maybe this one isn't an exaggeration when they say leap forward
Curious how this will fare when playing Pokemon Red.
Gemini 3 Pro has been playing Pokemon Crystal (which is significantly harder than Red) in a race against Gemini 2.5 Pro: https://www.twitch.tv/gemini_plays_pokemon
Gemini 3 Pro has been making steady progress (12/16 badges) while Gemini 2.5 Pro is stuck (3/16 badges) despite using double the turns and tokens.
I think what would be interesting is if it could play the game with vision only inputs. That would represent a massive leap multimodal understanding.
Yeah the "High frame rate understanding" feature caught my eye, actual real time analysis of live video feeds seems really cool. Also wondering what they mean by "video reasoning/thinking"?
I don’t think it’s real time? The videos were likely taken previously.
Screen understanding is huge for further automating dev work.
The document is paints a super impressive picture, but the core constraint of “network connection to Google required so we can harvest your data” is still a big showstopper for me (and all cloud-based AI tooling, really).
I’d be curious to see how well something like this can be distilled down for isolated acceleration on SBCs or consumer kit, because that’s where the billions to be made reside (factories, remote sites, dangerous or sensitive facilities, etc).
People with your concerns probably make up 1% of the market if that. Also I don’t upload stuff I’m worried about Google seeing. I wonder if they will allows special plans for corporations
I’m very curious where you get that number from, because I thought the same thing until I got a job inside that market and realized how much more vast it actually is. The revenue numbers might not be as big as Big Tech, but the product market is shockingly vast. My advice is not to confuse Big Tech revenues for total market size, because they bring in such revenue by catering to everyone, rather than specific segments or niches; a McDonald’s will always do more volume than a steakhouse, but it doesn’t mean the market for steakhouses is small enough to ignore.
As for this throwaway line:
> Also I don’t upload stuff I’m worried about Google seeing.
You do realize that these companies harvest even private data, right? Like, even in places you think you own, or that you pay for, they’re mining for revenue opportunities and using you as the product even when you’re a customer, right?
> I wonder if they will allows special plans for corporations
They do, but no matter how much redlining Legal does to protect IP interests, the consensus I keep hearing is “don’t put private or sensitive corporate data into third-parties because no legal agreement will sufficiently protect us from harm if they steal our IP or data”. Just look at the glut of lawsuits against Apple, Google, Microsoft, etc from smaller companies that trusted them to act in good faith but got burned for evidence that you cannot trust these entities.
Special since Trump, which non-US company should trust and invest know-how to an us company. And then are also governments. Also special since Trump, is way to risky to send any data to an us company.
Arpanet was supposed to be decentralized. Now everyone wants to centralize everything so in a war it is sufficient to strike 100 data centers and the whole tethered economy collapses.
That is called progress.
EDIT: You can downvote the truth but still no one wants your "AI" slop.
Ah, the fond memories of telnetting to NCSA to upload the raw HTML of my first website, written on an OG Macintosh computer and ported via floppy to a PowerMac for network connectivity.
Simple, elegant. I do miss those days.
what framework is being utilized for computer use here?
Yes, but can it play PacMan yet?
So we’re going to use this to make the maid from the Jetsons finally. Right?
im realizing how much of a bottleneck vision models are
im just a glorified speedreadin' promptin' QA at this point with codex
once it replaces the QA layer its truly over for software dev jobs
future would be a software genie where on aistudio you type: "go make counterstrike 1.6 clone, here is $500, you have two hours"
edit: saw the Screenspot benchmark and holy ** this is an insane jump!!! 11% to 71% even beating Opus 4.5's 50%...chatgpt is at 3.5% and it matches my experience with codex
> once it replaces the QA layer its truly over for software dev jobs
Maybe. However, with CYA requirements being everywhere in industry, there would have to be 100 waiver forms signed. I-promise-not-to-sue-company-if-AI-deletes-the-entire-database
It won't happen for that reason alone. Oh who am I kidding of course it will