If “aim” is already highly probable after the first line, the example is much less interesting. A more cool question is whether the model is carrying a latent

If “aim” is already highly probable after the first line, the example is much less interesting. A more cool question is whether the model is carrying a latent representation of the destination that isn’t yet reflected in the immediate token probabilities? I don’t know. Anthropic is investigating: https://www.anthropic.com/research/tracing-thoughts-language-model?s=prompt+injection&utm_source=chatgpt.com https://www.anthropic.com/research/tracing-thoughts-language...