Raw LLM Responses

You can verify yourself that Glaze/Nightshade does not work, but installing gemma3 via ollama (super easy), and running it on your poisoned image. If Nightshade worked, you would get incorrect image descriptions, preventing training on your image. But descriptions are spot on. That is why AI labs have barely discussed Nightshade since its publication: it's not a concern. There has been tons of new models released since then, with far improved image generation capabilities. Another somewhat important detail is how image generation itself is changing: diffusion models (Midjourney, DALL-E) are mostly out, and token-based (autoregressive) models, like ChatGPT image generation (but also Gemini, Qwen) are in. They have far better details, and they support in-context visual learning. So you can combine photo references into a detailed, correct combination of them. This addresses the reference use case: photos will always be better, but latest models are far superior to diffusion-based. Lastly, on inevitability of AI. This is very different from NFTs. NFTs never saw national investments to the scale of trillions of dollars in infrastructure. Consumer use cases is not the goal: job automation is. For illustration, it will mean an avalanche of lower quality, but "good enough" stuff. This approach - far cheaper, far faster, and mostly good enough - is the classic disruption formula, but with AI applied to pretty much all intellectual work. I don't like that at all, this is really bad for humanity long term, but there's almost nothing we can do to stop it. AI development is a national security question.