9 Apr
2025
9 Apr
'25
8:21 p.m.
At 12:39 PM 4/9/2025, Bill Gunshannon via cctalk wrote:
I'll continue to play with it but my initial
reactions still stand.
No sign of intelligence and not ready for prime-time.
What did you ask it to do? Exactly, I mean? When it didn't give you
the answer you wanted, what did you say next?
It would not surprise me if comp-sci departments now offer a major
in writing GPT queries.
Today's AI wranglers might be writing queries that range from a few
words to a many pages of text. They're also having a conversation
with it to refine their query. It's more than google.com circa 2005.
Many of us spent years learning how to get better results from Google
more often, after all.
I don't have a problem with skepticism about it. Call it a parlor trick
all day long, tell me it's not *truly* intelligent, but you should also
examine it enough to get to the point where you'll say "that's truly an
amazing and perhaps sometimes very useful parlor trick."
- John
9 Apr
9 Apr
9:36 p.m.
New subject: Why I am not worried about AI taking over the world!
On 09/04/2025 20:21, John Foust via cctalk wrote:
I don't have a problem with skepticism about it.
Call it a parlor trick
all day long, tell me it's not *truly* intelligent, but you should also
examine it enough to get to the point where you'll say "that's truly an
amazing and perhaps sometimes very useful parlor trick."
It is a glorified markov's chain. Great at reproducing common patterns
that don't require any real understanding.
Julf
10:02 p.m.
New subject: Why I am not worried about AI taking over the world!
On Wednesday, April 9th, 2025 at 12:36, Johan Helsingius via cctalk
<cctalk(a)classiccmp.org> wrote:
It is a glorified markov's chain. Great at
reproducing common patterns
that don't require any real understanding.
I've been tinkering a little bit with a small model (I have a copy of Nemotron running
inside of
Ollama on my laptop) and it seems like a fun tool for developing non-player characters for
tabletop
games. It's helped me flesh them out in terms of personality and motivations.
The Doctor [412/724/301/703/415/510]
WWW: https://drwho.virtadpt.net/
Get thee down. Be thou funky.
9:59 p.m.
New subject: Why I am not worried about AI taking over the world!
On Wednesday, April 9th, 2025 at 11:21, John Foust via cctalk
<cctalk(a)classiccmp.org> wrote:
It would not surprise me if comp-sci departments now
offer a major
in writing GPT queries.
Perhaps it'll mean having some experience with creative writing will gain respect
again in
the eyes of interviewers. Talking about having any kind of liberal arts education in
your
background when you're interviewing in tech these days is a bad idea (speaking from
experience after the 2023 round of layoffs throughout the industry, and earlier as an
interviewer).
The Doctor [412/724/301/703/415/510]
WWW: https://drwho.virtadpt.net/
Get thee down. Be thou funky.
11:19 p.m.
New subject: Why I am not worried about AI taking over the world!
> It would not surprise me if comp-sci departments
now offer a major
> in writing GPT queries.
On Wed, 9 Apr 2025, Carlos E Murillo-Sanchez via cctalk wrote:
Nowadays, "prompt engineering" in the
context of LLMs is a thing.
35 years ago, it was known as "Query Engineering"
Not a major, nor even a full course, but discussed quite a bit in
"Information Retrieval" classes
--
Grumpy Ol' Fred cisin(a)xenosoft.com
MLIS (Master of Library and INFORMATION Studies), 1991
11:54 p.m.
New subject: Why I am not worried about AI taking over the world!
In pre-computer engineering it was called 'asking the right questions' and
speaks of understanding the problem domain and appreciating the impact of
the consequences of the design. It's not only useful for the engineer's
ability but also to prompt the rest of the team into further expanding the
analysis.
How similar is 'prompt engineering' ?
Sometimes it seems to be pre-charging the LLM with what amounts to the real
(or desired) answers. Other times it's directing the LLM to search in areas
for which it doesn't easily find the connection but does actually have
extensive details.
On Wed, Apr 9, 2025 at 10:19 PM Fred Cisin via cctalk <cctalk(a)classiccmp.org>
wrote:
> It would
not surprise me if comp-sci departments now offer a major
> in writing GPT queries.
On Wed, 9 Apr 2025, Carlos E Murillo-Sanchez via cctalk wrote:
Nowadays, "prompt engineering" in the
context of LLMs is a thing.
35 years ago, it was known as "Query Engineering"
Not a major, nor even a full course, but discussed quite a bit in
"Information Retrieval" classes
--
Grumpy Ol' Fred cisin(a)xenosoft.com
MLIS (Master of Library and INFORMATION Studies), 1991
10 Apr
10 Apr
3:26 a.m.
New subject: Why I am not worried about AI taking over the world!
On Wed, Apr 09, 2025 at 02:19:11PM -0700, Fred Cisin via cctalk wrote:
Am I the only one that remembers how everyone was going to
be a "knowledge engineer" when expert systems were the be
all end all thing in the '80s?
Can anyone say 5th generation computing?
BLS
> It would
not surprise me if comp-sci departments now offer a major
> in writing GPT queries.
On Wed, 9 Apr 2025, Carlos E Murillo-Sanchez via
cctalk wrote:
Nowadays, "prompt engineering"? in the
context of LLMs is a thing.
35 years ago, it was known as "Query Engineering"
Not a major, nor even a full course, but discussed quite a bit in
"Information Retrieval" classes
10:21 p.m.
New subject: Why I am not worried about AI taking over the world!
On Wednesday, April 9th, 2025 at 18:26, Brian L. Stuart via cctalk
<cctalk(a)classiccmp.org> wrote:
Am I the only one that remembers how everyone was
going to
be a "knowledge engineer" when expert systems were the be
all end all thing in the '80s?
Nope, you're not the only one. I call myself a "knowledge worker" from time
to time.
The Doctor [412/724/301/703/415/510]
WWW: https://drwho.virtadpt.net/
Get thee down. Be thou funky.
12 Apr
12 Apr
6:31 p.m.
New subject: Why I am not worried about AI taking over the world!
On Wed, Apr 09, 2025 at 01:21:06PM -0500, John Foust via cctalk wrote:
At 12:39 PM 4/9/2025, Bill Gunshannon via cctalk
wrote:
Because LLMs are - grossly simplified - regexes & Markov chains on
steroids and meth. They are massively scaled up pattern matching
systems and while there are plenty of use cases where this can, in
fact be quite useful (e.g. when the job is, indeed, pattern matching
such as image recognition, feature extraction from still/moving images
etc[0]), they DO NOT THINK.
I'll continue to play with it but my initial
reactions still stand.
No sign of intelligence and not ready for prime-time.
What did you ask it to do? Exactly, I mean? When it didn't give you
the answer you wanted, what did you say next?
It would not surprise me if comp-sci departments now offer a major
in writing GPT queries.
Thus making the "You don't need to understand anything, just be able
to bullshit your way along is good enough" approach official. GPT
and friends will cheerfully emit bullshit[1]. Among the more newsworthy
examples are lawyers having LLMs write the papers for cases where
the LLM (aka bullshit generator) happily cited non-existing precedent
cases to support their argument - with the lawyers being scolded by the
very irritated judge and the lawyers (yes, plural, of course this
happened more than once) IIRC facing disbarment for this.
Today's AI wranglers might be writing queries that
range from a few
words to a many pages of text. They're also having a conversation
with it to refine their query. It's more than google.com circa 2005.
Many of us spent years learning how to get better results from Google
more often, after all.
I don't have a problem with skepticism about it. Call it a parlor trick
all day long, tell me it's not *truly* intelligent, but you should also
examine it enough to get to the point where you'll say "that's truly an
amazing and perhaps sometimes very useful parlor trick."
Yes, right until it tells you to put glue into the spaghetti sauce.
Kind regards,
Alex.
[0] Swiss Railways (SBB) uses an "AI" (read: oversold pattern matcher)
to detect and identify railway carriage wheel damage based on the
noises this damage causes when a train is running over the rails.
They trained it with carefully curated data sets and it apparently
works _very_ well. The alternative would be manual inspection of
these wheels, which due to scale, is just not doable with a useful
level of coverage.
[1] See Harry Frankfurt: On Bullshit
--
"Opportunity is missed by most people because it is dressed in overalls and
looks like work." -- Thomas A. Edison
13
days inactive
16
days old
9 comments
8 participants
participants (8)
-
Adrian Godwin -
Alexander Schreiber -
Brian L. Stuart -
Carlos E Murillo-Sanchez -
Fred Cisin -
Johan Helsingius -
John Foust -
The Doctor