My brother sent me a clip of Jon Stewart satirizing a clip of Sam Altman. The clip begins with
Altman saying: “I don’t want to go back to a world where I, like, painfully try
to find things on my computer.” Stewart
takes that to the logical absurd conclusion and states: “We are leveraging
existential danger on a future we don’t even need.” That should be obvious with the resistance to
the large build out of data centers and to what end? So an average person like me (or you) can
generate unnecessary graphics, documents, and searches that will fill all of
that data storage space. On the one hand it is typical American capitalism
selling products that nobody really needs and on the other hand it is
overselling a product of unknown quality and safety that may actually be
dangerous.
I use AI as a slightly better search engine. It works fairly well in that capacity but all
of the references have to be verified and read or I do not use them. At that
level it continues to make things up including formatted references that look
like the real thing until you try to find them.
A lot of people reify AI as though it is intentional, has a mind
and personal agency. As part of that process, AI is assumed to have superior
intelligence and does not make mistakes.
I have routinely argued with AI on specific points in medical research
and practice and it does retract statements and acknowledge mistakes. That is
why the following banner placed on US Government web sites (this is from
pubmed.gov) does not make any sense. (click to enlarge).
I embarked on an informal AI experiment myself. I have some experience with HTML from over 20
years ago when I put up a website on GeoCities. Yes, it is that long ago – but I have not
used HTML that much ever since. I used
Google Blogger to produce this blog and generally do all of my editing in Word
before pasting it into Blogger. This
results in HTML code containing a mix of Blogger and Microsoft tags. Overall, it means that the appearance of
graphics is variable based on the original image but also the HTML. Some of the graphics look blurred until you
click on them and the original graphic is displayed in its original size. The options in Blogger to display the graphic
in various sizes does not resolve the problem very well. In addition, to display complex data, I have
taken to using more tables this year.
Some of the tables crowd out the text into a second line or display it
all as a long continuous column.
To correct all of
these problems and improve the overall appearance I decided to try to use AI to
rewrite the HTML. Since Blogger is a
Google product, I would try Google’s AI Gemini.
The first task was to input the goal and initial task. It is a simple
enough request with a goal of stretching all of the graphics across the top of
the blog posts like you night see on a Substack or WordPress blog. I will post
the initial parts of the AI response without all of the HTML code. Those codes were posted as complete 50-100 line
modifications or a single long line in more basic cases.
After surveying the results, many were not aesthetically pleasing. In some cases book covers and molecules now seemed to dominate the post covering up to 25% of the length. The smaller images were more well defined and fit better. That led to this request.
At this point Gemini made 2 suggestions – one on a case-by-case
basis and one for a global change in the entire code. I used the case-by-case modification and that
seemed to work well resizing book covers and molecules to an more appropriate
dimension. I thought that would be the
final solution but 2 days later – I pulled some of the same posts back up and
they were all back at giant size. I rechecked
all of the HTML modifications and could not find a reason. A note here about the HTML is in order. If you have not checked the underlying HTML
of your blog, I encourage you to take a look.
It is complicated to say the very least.
In my case I was looking at hundreds of lines of this code and multiple
image and table tags. In some cases,
modifications were not allowed. In other
cases, I was not certain whether the change would result in changes to my post
and needing to redo it all from scratch.
Since there are nearly 1,000 posts on my blog the nightmare scenario would be
changing them all and not being able to correct them. To an expert – I am probably catastrophizing.
So at this point the recommended solution did not anticipate the automatic editing done by the Blogger editor. I decided to go back to the original format rather than risk further problems for an inconsistent or nonexistent result.
On a qualified basis what have I learned about AI in this case. When I say qualified that includes all of the following disclaimers:
1. Not an expert in
AI
2. Not an expert in
HTML
3. Working with
currently publicly accessible AI
First, AI is very fast and provides fast answers to
questions. It generated very long HTML segments in seconds. Second, it reflects
my overall experience using it to search the medical and scientific literature -
it does not necessarily provide the best answer or solution. Third, some of that decision making seems to
reflect an inadequate working memory or failure to consider all of the possible
solutions. I am referring specifically to the failure to anticipate that the
Blogger editor would delete certain tags in the HTML that were required for the
solution. After the fact – it appears to
recognize the problem but at an intelligence level that is a fairly low bar to
pass.
I previously posted about the moral and ethical implications
of AI. In that case the real-life scenario
was AI being used to produce nude photographs of anyone and the possible
implications of using autonomous AI in warfare.
It seems clear that moral and ethical decision making is not woven into
any current AI and the evidence is all of the technical people working in that
area sounding the alarms about no guardrails.
This exercise suggests a deficiency in planning. AI certainly has algorithmic intelligence
based on how fast it can compute and produce HTML code. To some extent that reflects holding all of
those rules in memory to produce the code.
It also needed to decode a basic level of abstraction in the initial
question about the overall goal.
There is some breakdown at the level of epistemic rationality
– if knowledge about the Google Blogger software was known. That occurs because producing HTML to achieve the
goal of graphics that are as wide as the blog post and subsequent individual modifications
is inconsistent with using tags that will be eliminated by the Blogger editor. It
is essentially producing a complicated nonsolution. That also fails at the
level of instrumental rationality or choice or choices that produce the
desired result.
For me this exercise confirms what I know about AI since I started working with it about a year ago. The programs I have access to are certainly fast and can produce a lot of information. Most search engines these days go right into an AI mode and produce several explanatory paragraphs and summaries going far beyond just listing references like the pre-AI days. This example illustrates some of the shortcomings. I certainly would not accept the references or solutions it produces at face value. If you know more about the references or a theory it has come up with – do your best to confront it. You may eventually get to information that you can use. The real advantage these systems seem to have over humans reminds me of the old expression number crunching. These systems are tireless number crunchers. I am sure that I could still be asking for HTML codes for my blog and the AI would still be spitting them out. But I would still have no assurances of a better solution or one that incorporates all of the data.
Or as one colleague put it "AI is exhausting!"
If my little experiment is any indication - it also illustrates potential problems with medical applications. Assuming AI is intelligent in the human sense if it can produce a blizzard of information of varying relevance without incorporating some critical existing information and not proceeding to a real solution is problematic. The statement that "AI will solve all medical problems in the next 10 years." rings hollow unless it can solve problems for specific individual patients. That is considerably harder than optimizing graphics for this blog.
George Dawson, MD, DFAPA
Graphic credit:
CBrasil0. DALL·E 2025-01-31 21.58.22 - A realistic image of the interior of a modern data center. The scene features long rows of tall, sleek server racks filled with blinking LED lights
Creative Creative Commons license: https://creativecommons.org/licenses/by-sa/4.0 CC BY-SA 4.0 via Wikimedia





