On mechanical parrots

“(…) In the beginning, there was man, and for a time, it was good. But humanity's so-called civil societies soon succumbed to vanity and corruption. Then man made the machine in his own likeness. Thus did man become the architect of his own demise. But for a time, it was good.”

Today, in the third decade of the 21st century, we are amazed by our own creation: Artificial Intelligence. Even though the Culture Industry's output over the last 60 years discussed at length the impact of the day this would become real, we clearly were not prepared for its arrival.

Obviously, generative AIs are nothing more than sophisticated mechanical parrots trained on data from all kinds of sources, which, by all indications, were not notified, compensated or consulted.

We are not prepared for the simplest ethical dilemmas of using this technology, such as this one: Porto Alegre city council passes the first law written by ChatGPT (in Portuguese).

Although they have no consciousness or creativity, and often hallucinate (the technical term for when the AI starts generating nonsense text with no connection to what was asked), they "lie", since they make up nonexistent information, as in the case where a judge used a thesis invented by ChatGPT to write a ruling (in Portuguese).

There is no denying that Generative AI, as it is being called, has already transformed the world and is here to stay, but its impact, however large, falls far short of science fiction. We are still far from having truly autonomous assistants that need little supervision, like the ones we see in movies such as Marvel Studios' "Iron Man".

There we have an AI able to follow Tony Stark's orders flawlessly. That is, although it is also unable to create anything truly new, since it is not endowed with consciousness, autonomy or creativity, it can follow instructions precisely.

That does not happen in the real world, at least not with the versions open to the public; even the paid ones have glaring limitations.

For example, in the last few days I asked GPT-4 to generate SQL code to fix flaws in the structure of several tables in a MySQL database. But no matter how much I explained and detailed how the input data should be processed, it was so stuck to its training data that, simply because the table structure input had a "+" where a "-" would be expected, signaling the opposite of what was expected in this particular case, it hallucinated and ignored my instructions.

Because, for a reason only GLPI's creators (https://glpi-project.org/) can explain, the command that shows which columns should exist and which do exist doesn't follow the convention of using a "+" sign before columns to be created, as opposed to a "-" sign for those to be removed.

Example:

Running "php bin/console database:check_schema_integrity", which checks the integrity of the GLPI database, gives the following output:
--- Expected database schema
+++ Current database schema
@@ @@
   `users_id_editor` int NOT NULL DEFAULT 0,
   `users_id` int NOT NULL DEFAULT 0,
   PRIMARY KEY (`id`),
-  KEY `date_creation` (`date_creation`),
-  KEY `date_mod` (`date_mod`),
+  KEY `date_creation` (`date_creation`) USING BTREE,
+  KEY `date_mod` (`date_mod`) USING BTREE,
   KEY `date` (`date`),
   KEY `is_private` (`is_private`),
   KEY `item` (`itemtype`,`items_id`),
@@ @@
   KEY `requesttypes_id` (`requesttypes_id`),
   KEY `sourceitems_id` (`sourceitems_id`),
   KEY `sourceof_items_id` (`sourceof_items_id`),
-  KEY `users_id_editor` (`users_id_editor`),
+  KEY `users_id_editor` (`users_id_editor`) USING BTREE,
   KEY `users_id` (`users_id`)
 )

I think it makes more sense now: the lines preceded by "-" are the expected schema, the columns that should be created, while those preceded by "+" are the ones that should be removed.

It was quite frustrating to watch ChatGPT 3.5, 4 and 4-Turbo fail to understand this, because it goes against convention and is exactly the opposite of a diff made by Git, for example.

But you and I, being human, can abstract this, because there are explanations at the top of each statement; the machine can't, because it is stuck to its training data.

This is just one example. It also has an innate ability to wreck a text when correcting it, changing the cadence of sentences, destroying cohesion and the author's style. And do you know how I know that? I used it to proofread texts on this blog, and it failed miserably.

So, while we are amazed by the incredible things Generative AI can do, underneath lie huge limitations that keep it from even coming close to us.