The Osaka Expo opened on April 13.  In search of news of the event I
consulted Google search (in Japanese).  Google replied in the ordinary
way, with the URL of the official Expo site on top followed by news
items, images and such.  Among them, on the first page was a rather
new field which listed "related questions".  The first of them was a
great surprise.  It read:

  "For what reason(s) was the 2025 Osaka Expo canceled?"

Other items of the list discussed ticket prices, access to the
grounds.  Obviously the above text was generated by AI, for it is
unlikely for a human being to err in this manner, but Google does not
clarify.  (Confirmed April 28 Japan time)  So, one part of Google
(probably AI) treats the exposition as canceled while another part
understands that it is in progress.  There is a failure to notice
an obvious contradiction.

Can AI distinguish between "free" as in "free drinks" and "free" as in
"freedom"?  We should watch out for this.  The above makes me think
that we should not expect much.  Even if we observe AI getting the
distinction right in a number of cases we should not expect it to be
always correct.

For one thing many people are confused and much has been written
based on the mistake.  it is unreasonable to expect that none of the
erroneous material has been fed to the large language model (LLM)
neural networks for training.  In addition the above example
indicates that AI is not good at coping with lack of integrity.
Perhaps a contradiction which is obvious to a human being is not
so for AI.

---

Related article: Researchers have found methods to investigate what is
going on inside the neural networks that power AI.  They have
discovered that the process differs greatly from human reasoning.

We Now Know How AI 'Thinks'--and It's Barely Thinking at All
https://www.yahoo.com/finance/news/now-know-ai-thinks-barely-010000603.html

Though I don't follow developments in this field closely, what is
written in this article confirms my suspicions.

Another way to see this is that the AI developers now have the
equivalent of a debugger which allows them to probe the internal
process.  This is likely to affect development.

_______________________________________________
libreplanet-discuss mailing list
[email protected]
https://lists.libreplanet.org/mailman/listinfo/libreplanet-discuss

Reply via email to