Re: OT: Vaccine safety

On Sun, 5/25/2025 11:54 PM, rbowman wrote:

On Sun, 25 May 2025 14:42:43 -0400, Paul wrote:
 

The AI can't write code, in computer languages for which insufficient
"history" exists to train them. The AI hung on me (not even the safety
timer went off), when I tried that.

 
https://www.bleepingcomputer.com/news/security/ai-hallucinated-code-
dependencies-become-new-supply-chain-risk/
 
Synopsis: AIs hallucinate package names and helpful blackhats create
packages to match, filled with malware.
 

Well, you've been studying your AIs now for a while,
so you know what some of their limitations are.

1) You can't refactor Firefox (tarball or Mercurial source).
   It's too big. The machines to do this, would require astronomical
   amounts of RAM. Maybe a training machine could do it, but not
   an inference machine. Typical limits are 256,000 tokens ("words").
   Once any machine needs even tiny comm links to hook nodes together,
   all the performance of the machine is lost. Either the memory subsystem
   has terabytes per second of bandwidth, or your answer isn't coming back
   any time soon.

2) This means you will be working at the code module level.
   You still have to do your own architecting and partitioning,
   although the AI could read a magazine article about what you
   were trying to do, and pretend to understand.

That suggests to me at the moment, the leash on your AI dog is a
rather short one. You still have to review "everything" in the source
file, after it is finished. The machine comments the code, oh yes,
but the machine has no "insight". An experienced programmer would
say "if you do it this way, bad things will happen to you". Whereas
the AI says "heres your code", and it is up to you to discover
the program limitations (not enough stack, not enough heap,
bug in system library routine will limit max size the program
can handle).

I successfully managed to get it to write a 40 line program.
On another occasion, when asking for a shorter program than
that in C: , I got a programming pattern I'd never seen before.
It almost smacked of "intelligence", but I know that's
not possible, right ? :-/

But for the people using Codex, they appear to spend longer
sessions in it, implying somewhat larger scale. You still
can't point it at the Firefox tarball and say "rewrite that for
me".

The very last time I asked it to write me a program, I got
three dancing balls as my "answer", and it just sat there.
In the year 2025, every response from the computer must
be a fucking puzzle. We can't just print on the screen
"error 9" and be done with it. Three dancing balls. Hmmm.
When I asked the machine later "what happened to the program
you were writing", there was no record of the request.
Presumably there is a shovel full of dirt, underneath
the AI's hall carpeting with my query in it. I was asking
for a C program.

   Paul