Long time readers will know that I’ve been leveraging Google NotebookLM for AI generated conversations of my posts.
For the most part, they are very good. There are always interesting ideas, observations, and sometimes fascinating twists on the writing. There are some hallucinations, but many that are knowable only to me. Listeners may not even recognize these as hallucinations.
But recently, many of my posts have become more analytic and problem solving in nature. Some very directive on “here are the equations and analysis we use.”
And in those conversations, I notice the models really struggle. They tend to invent completely false concepts, ideas, acting randomly.
For a test, I wrote a very short, succinct post. No hidden message, no hidden pattern, just 7 lines. Here it is:
1+1= 2
2+2 =4
3×3=9
9-3=4
4×4=16
5-1=4
4/2=1
That’s it, just 7 rows of math calculations and 2 are erroneous I intended them to be nothing more than what you see.
Now, I will ask you to scroll to the bottom and listen to the 5 minute discussion and insights into these 7 rows of equations. They get into the questions of the history of numbers, whether there is another way to interpret 1+1, and all sorts of other “serious discussions.”
All completely fabricated, nothing having to do with what I wrote. Any human observer looking at the same “post” might say, “Dave, you doing your homework online? Dave, did you have to use your fingers and toes to figure these out? Dave, you screwed up……”
And through all of this, they never identified the two erroneous equations! They treated every equation as perfectly correct!
In just this simple post, we see so much of the weaknesses inherent in these tools. They have no understanding of meaning. They have no understanding of intent. They have no reasoning power.
Then this causes me to reflect back on my “idea” posts. The posts where I’m talking about ideas, concepts, observations, sometimes provocation. Are the conversations being generated for these as flawed as the conversation 7 equations? Are the hallucinations as bad, is the construct as bad, but do I fail to recognize them because of the abstraction of the ideas.
We have come so far with these tools. But the more I use them, I realize how very far we have to go.
Enjoy this discussion of the meaning of 7 lines of arithmetic equations. It’s a little frightening. At the same time, the stories and discussion they invent are hilarious when you realize the total absence of meaning in the discussion.
Leave a Reply