Positive users hail directional prompts as a clever breakthrough letting GPT-5.6 bypass ghost fonts, while negative users worry about model collapse and fragile security from such techniques.
Based on 13 visible X reactions from 21 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@incriptionnn To me it’s less interesting whether models can do it “themselves,” my point is more that it’s easy to prompt a model to be able to read these. If Ghost Font takes off enough that LLMs learn of it (or see it in search results), they will see how it works and succeed at reading it.
@goodside big if OpenAI trains that adversarial noise pattern into the next batch of synthetic data — model collapse via font steganography is a wild failure mode.
@goodside Prompting a model how to perceive rather than just what to look for is the next level of prompt engineering. Absolute masterclass.
@goodside the fact that it understood the "sliding upward" description and decoded it from that is actually insane 😁
A simple motion-tracking prompt helped the model decode the text.
@incriptionnn To me it’s less interesting whether models can do it “themselves,” my point is more that it’s easy to prompt a model to be able to read these. If Ghost Font takes off enough that LLMs learn of it (or see it in search results), they will see how it works and succeed at reading it.
@goodside big if OpenAI trains that adversarial noise pattern into the next batch of synthetic data — model collapse via font steganography is a wild failure mode.
@goodside Prompting a model how to perceive rather than just what to look for is the next level of prompt engineering. Absolute masterclass.
@goodside the fact that it understood the "sliding upward" description and decoded it from that is actually insane 😁
@goodside This is wild—so much for "AI-proof" fonts. Really shows how far prompting alone can go.
@goodside That’s a clever way to get it to see.
@teortaxesTex I feel like it needs some perception of movement in video, otherwise the intuition to reveal the same via tool calls. If you hint the direction the noise is moving it gets it. The way it fails otherwise is just trying every steganalysis method known to man and getting lost: https://twitter.com/goodside/status/2076361122328858829
I got GPT-5.6 Sol to read “Ghost Font” (a proposed method for making text legible to humans but not AI) by just telling it in what direction it looks like the noise is moving: https://x.com/goodside/status/2076361122328858829/video/1 https://twitter.com/ericlu/status/2075876651574210643
@AndrewMayne That’s interesting; I suspect reasoning effort does a lot of heavy lifting because without hints Sol tends to just throw every steganalysis method it knows at it and either finds the motion pattern or doesn’t—and in the latter case it often hallucinates.
I got GPT-5.6 Sol to read “Ghost Font” by just telling it in what direction it looks like the noise is moving: https://x.com/goodside/status/2076346888538800207/video/1 https://twitter.com/ericlu/status/2075876651574210643
@goodside It's so good at some things, and yet so bad at the basics at the same time... https://x.com/davidmanheim/status/2076373798694052039
@goodside I don't care that it can succeed via tools that's a cognitive success, but perceptually it still doesn't see anything
Positive users hail directional prompts as a clever breakthrough letting GPT-5.6 bypass ghost fonts, while negative users worry about model collapse and fragile security from such techniques.
Based on 13 visible X reactions from 21 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@goodside That’s a clever way to get it to see.
@teortaxesTex I feel like it needs some perception of movement in video, otherwise the intuition to reveal the same via tool calls. If you hint the direction the noise is moving it gets it. The way it fails otherwise is just trying every steganalysis method known to man and getting lost: https://twitter.com/goodside/status/2076361122328858829
I got GPT-5.6 Sol to read “Ghost Font” (a proposed method for making text legible to humans but not AI) by just telling it in what direction it looks like the noise is moving: https://x.com/goodside/status/2076361122328858829/video/1 https://twitter.com/ericlu/status/2075876651574210643
@AndrewMayne That’s interesting; I suspect reasoning effort does a lot of heavy lifting because without hints Sol tends to just throw every steganalysis method it knows at it and either finds the motion pattern or doesn’t—and in the latter case it often hallucinates.
I got GPT-5.6 Sol to read “Ghost Font” by just telling it in what direction it looks like the noise is moving: https://x.com/goodside/status/2076346888538800207/video/1 https://twitter.com/ericlu/status/2075876651574210643
@goodside It's so good at some things, and yet so bad at the basics at the same time... https://x.com/davidmanheim/status/2076373798694052039