Suppose I tell you that the father of James Coburn is Samuel Coburn. I then ask you who is the son of Samuel Coburn. Can you answer that question? I would assume that all else being equal, you would quickly and without any hesitation reply that James Coburn is the son of Samuel Coburn.Turns out that the generative AI of today might not produce that correct answer. The AI might balk and insist that the answer is unknown.
People are often shocked and dismayed that generative AI might not be able to answer correctly such an easy question. The overall fluency of generative AI lulls us into thinking that GenAI is totally conversant and possibly on par with human interactivity. You must fight the strident urge to assume this. Generative AI is primarily a computational pattern-matcher and does not enter into the realm of sentience.
Now that I’ve covered the core of this weighty matter, I would bet that you are pondering how it might impact your everyday use of generative AI.Realize that whenever you are using generative AI, there is a solid chance that this B-is-to-A oversight might adversely affect the essays or problem-solving that you are asking GenAI to do. Review and assess all responses by the AI accordingly..
For those of you interested in prompt engineering or prompt design, I’ve been doing an ongoing series of insightful explorations on the latest in this expanding and evolving realm, including this coverage:Vagueness is often shunned when using generative AI but it turns out that vagueness is a useful prompt engineering tool . If a model is trained on a sentence of the form ‘A is B’, it will not automatically generalize to the reverse direction ‘B is A’. This is the Reversal Curse.
The pretraining sets for modern LLMs are very large and diverse. Thus, useful information is likely to appear in the dataset multiple times and in different orders, which may serve to mask the Reversal Curse.”First, you might be thinking that after a generative AI app has been initially data trained, you could merely run a series of tests to detect whether the B-is-to-A conditions have been absorbed. This is not such an easy testing task to perform.
Do not be so quick to leap to wild conclusions. As mentioned above, the Reverse Curse does not seem to especially occur when you do in-context learning. In-context learning consists of telling the generative AI about something when you are using the AI app. Notice that I told the AI about the relationship between Michael and Lance. I did not rely upon the initial training of the AI app. Instead, I provided the information while using generative AI.