NeuroSpring

I write about random AI stuff

[neurospring]

You use AI, oh my god

The fact that the anti-AI sentiment became prevalent on the broader internet is frustrating. And it's prevalent in a completely braindead way too. People either think the only AI that exist are the current big LLMs and image generation models, or they hate on all AI, somehow? At least they should be specific about what kind of AI they disapprove of, because it sounds absolutely braindead when they just say that "AI is bad". And if you are one of these people reading this and you think "I got it, I will say generative AI is bad!" well that still sounds braindead, because generative AI models are still not just the models you hate, there's many uses you would approve of. Even though I think the models people really hate are fine, I can see why someone would dislike Midjourney for example and even show intense hatred towards it. Even if I think that's cringe, that's valid feelings to have I guess. But while they do that, they often call the model just "AI" and that's just ridiculous. One other funny thing is when people say they don't use AI, and well that's just not true obviously. Everyone uses AI, often a lot. Google search uses AI for example. It's like, if you hate apples, you wouldn't say you hate fruit while eating bananas.

[neurospring]

"It's not intelligent" - no u. you are not intelligent

[neurospring]

Just for fun

I have tested if language models can solve the intersection captcha from the neal.fun game:

I was betting on just no. And yeah, they seem to be unable to. ChatGPT 5 thought for like 5 minutes then gave a completely wrong answer.

As a grid (turns clockwise): 3 1 3 1 0 3 1 2 0

Gemini 2.5 pro thought about a minute and also gave me a completely wrong answer.

Visual Representation of the Solution: 2 Rotations (180°)0 Rotations (Correct)3 Rotations (270°) 0 Rotations (Correct)0 Rotations (Correct)0 Rotations (Correct) 0 Rotations (Correct)0 Rotations (Correct)0 Rotations (Correct)

And Claude 4.5 sonnet concluded all the squares are already correctly rotated so this is a trick question.

After careful examination, this puzzle appears to be a trick question! All the tiles seem to already be in their correct orientations. The image shows an aerial view of a highway interchange, and when I trace the roads across tile boundaries, they all connect properly as-is.

I am curious how far we are from LLMs being able to solve this.

[neurospring]

A new misconception I came across

There's always been and always going to be misconceptions about Neuro-sama. The "trained entirely on twitch chat" or "only learns from interactions" claims are common place, many people are not understanding where the LLM is from. I just have read a different, but also insanely incorrect comment on YouTube that got some likes, and I want to address it. Context, this comment is replying to a comment stating that Neuro required what a number of people consider "stolen" data. "Vedal has been working on her LLM for 2+ years now. It's so divergent from any open source LLM out there that the initial model weights don't matter as much anymore."

This reply is non-sense, and I want to rant about it while hopefully educating people. So, the language model used, whichever it may be, was trained on trillions of tokens by a big company. All these powerful models are. How many tokens do individuals use to fine-tune, post-train these models? Thousands to some millions. This is the range Vedal is using realistically. Big companies may use billions, and I just know that some people will argue that Vedal is big, he must be training this model on billions of tokens. Let's see - in that case, that training could be as much as 1% of the whole of the training data. (If you calculate with a model trained on ~10 trillion tokens further trained on 100 billion tokens.) But more realistically, let's say Vedal uses a million tokens for additional training for Neuro. Which is still a lot, but it's only 0.00001% of all of the training data. We can't know exactly how much data Vedal trains models on, but even if he would be training on industry levels, that would be still very far from the commenter's statement of "so divergent from any open source LLM out there that the initial model weights don't matter as much anymore". And let's address one other important thing: there was not one LLM. There has been multiple LLMs used over time. Intelligence upgrades equal model switches. It is not one LLM that is getting much smarter. Vedal said about a year ago on discord, "we are on llm model 6." Since then, there was one intelligence upgrade, so we are on the 7th LLM at this point. Vedal has to restart his training every time he does an intelligence upgrade. I'm sure people who this is new info for will ask, "how is it then that Neuro-sama has a consistent personality though?" so I want to address that. I honestly don't feel she is that consistent, but to the extent that she is, Vedal using the same and similar training data and prompts goes a long way. And roughly speaking, these LLMs all learn the same world-model and are not super different from each other to start with.

This is a huge topic and this is a very broad strokes, minimal explanation of things. But I wanted to clear this up a little bit.

[neurospring]

The reactions to Neuro 'drawing' are causing me physical pain. Most people who watch Neuro really are like zero percent familiar with LLMs in general, it's truly wild.

[neurospring]

I'm looking at even more DH3 images and I'm getting more and more excited. It feels like the underlying language model has more control over image generation? I have been testing shit with GPT4o and Nano banana and it's frustrating when the language model itself understands something perfectly, but it can't generate it. I feel like that will be less of a problem with DH3. Am I reading too much into the few images I have seen? Probably, but what else would explain what we see?

A bit of a weakness but I think it's also its biggest strenght, the model perhaps wants to follow the prompt a bit too closely. Doesn't seem to want to guess what you are thinking. But yeah that's good too, it doesn't reinterpret the prompt. It seem to do exactly as you tell it to.

[neurospring]

NANO BANANA (GEMINI 2.5 FLASH IMAGE) IS OUT. I AM YELLING BECAUSE IM VERY EXCITED and I have no one to share my excitement with.

As someone who has been excited about image generation models since all they could do is generate weird looking 32x32 images, this is just surreal.

[neurospring]

"It's generative AI that's bad."

I hate that the anti-AI crowd took the term generative AI and using it to mean... whatever they think it means. This is what generative AI means according to wikipedia: "Generative artificial intelligence (Generative AI, GenAI,[1] or GAI) is a subfield of artificial intelligence that uses generative models to produce text, images, videos, or other forms of data. These models learn the underlying patterns and structures of their training data and use them to produce new data based on the input [...]"

It's about the way these systems "learn" and how they produce "new" output. I feel like anti-AI crowd defines generative AI loosely as "AI systems that steal and plagarise, and produce art or text." You can make a generative AI model, train it on your own data, and it would be just as much generative AI as Midjourney or ChatGPT 5 is. Also, there are technologies that are very much generative AI but anti-AI people don't often think of it as such and generally have no problem with, like AI TTS models.

And perhaps most importantly, it is used in medicine. It is making new discoveries. For example, AlphaFold 3 is using generative AI technology, a diffusion model, to generate 3D molecular structures. So when people say "AI is not all bad, generative AI is bad" it just sounds insane. Of course when you bring this up people say "well of course I don't mean that generative AI". Well, didn't you try to be accurate with clarifying what AI you think is bad? You didn't clarify anything, just made a false statement.

[neurospring]

Nano banana testing!

"Nano banana", a new image model on lmarena that is theorized, or at this point confirmed to be, a new Gemini native image generation model is awesome, but it still has most of the flaws that any other model has. Though it may not be able to generate a fully solved Rubik's cube on it's own - it can "solve" one!

For this second one, I would have preferred if it kept the middle colors, but what can ya do. It does look solved at least.

I have lost the image where I have tried to get it to generate a solved Rubik's cube on its own, but it didn't do good. Showed the same flaws as other image generators. It is very interesting then that it can edit a photo to show a solved Rubik's cube, but can't generate one on its own. I will continue testing this when the model comes out. I have generated 20 Rubik's cubes now on lmarena and nano banana didn't come up yet. At least I do know for sure that no model can do this, lol.

(Imagine 10x similar examples...)

Okay I had the idea to have it edit an image, higher chance the nano banana model will come up that way. It didn't, but now we know that gemini 2.0 flash and gpt-image-1 can't just generate a solved Rubik's cube either. These models should be I think the most capable to do this after nano banana.

Now I want to see if they can "solve" a Rubik's cube... I will now try ChatGPT (uses gpt4o native image gen, I assume it's mostly the same model as gpt-image-1.) Yep, gpt4o can't do it!

Based on my previous testing on gemini 2.0 flash, that can't do it either. Okay, as I have tested/suspected, nano banana doesn't seem to be able to generate a Solved Rubik's cube outright either! Or it may just be because the cube is smaller here. More testing is needed.

I have now taken this image and told it to solve this Rubik's cube. It generated 2 green sides, but it's... better?

It's funny how flux just completely scrambled it instead, haha. I wanna test all this stuff further after the model comes out. Okay I did one more test, and ow it struggled with this one:

The two cubes (with the background maybe) were too much - it couldn't "solve it". Overall, 'nano banana' is definitely better than other models at 'solving' cubes, but of course, it isn't AGI.

[neurospring]

Man this blog post is gonna be long... No one will even read it lmao, ugh. Ah well it's mostly a nice way to vent for me.

[neurospring]

Why is it that anti-AI people generally know less than nothing about AI?

I have honestly never seen one who was actually knowledgable about how AI works. And to an extent it makes sense they wouldn't know much about it, they hate it, they don't wanna learn much about it, though they should still know the basics - but it's not even just that they don't know the basics, they actively hold totally false beliefs and have their own vocabulary almost. What gives?

[neurospring]

Btw deleting the comment and blocking the person for asking for a source for what you are saying is vile.

[neurospring]

Great, someone blocked me because I asked for a source for their claims about Neuro, including that she is trained on "not stolen data".

This belief that Neuro is trained on "ethical data" is sacred for a lot of people at this point. Wonder what will happen when a lot of these people find out what Neuro is made of? At some point someone on a stream or in a YouTube video will say what's up.

[neurospring]

Woooo

I'm so excited for Gemini 3 and "nano banana" (model on lmarena, theorised to be Gempix - Gemini native image generation). I am very excited for Gemini 3 based on how good 2.5 is, though anything could happen, we know nothing about the model. And nano banana looks great and it's awesome. It's looking to be very usable and imo generates probably the best and most aesthetically pleasing images right now.

[neurospring]

It's so funny to me when anti-AI people make shit up about AI and other anti-AI people will just stand behind it, it only matters to them that it sounds bad. They will even call the entire technology of generative AI bad. But then when it comes to AI they like, like Neuro-sama, they will make shit up in the other direction. "Oh Neuro? Based on what I have dreamed up, she is a totally ethically made model trained on twitch chat (they have all consented, don't question it) and other fully ethical data carefully curated by Vedal himself reaching tens of megabytes. She is not even generative AI, despite LLMs being like the definition of generative AI, and her other main component, the TTS, being generative AI as well. I am emotionally attached to these claims and if you say these LLMs cost millions of dollars to train, I may argue that Vedal has spent millions of dollars developing Neuro despite the fact that that still wouldn't solve the problem of the training data. I will never admit that Neuro was trained on the exact same kind of data I always argue it is unethical to train generative AI models on."

Just be honest and say that you like something despite it contradicting your 'values', or something.

[neurospring]

It has become increasingly difficult for me to watch Neuro-sama content. I disagree with how Vedal handles things, him being okay with many many people thinking very unreasonable things about how much he is responsible for Neuro's AI, and him also implying financial hardship when he is definitely very well off. Many people believe very unreasonable things about his work and some also about his finances. He knows this and is only adding fuel to the fire.

I don't believe what he is doing is justifiable. I can't really excuse it anymore, there are really no good-faith interpretations here I can think of. I'm gonna leave it here for now. Bai.

[neurospring]

I hope in ~5 years, the models we have now are gonna seem like a joke.