A Brief Guide to the Fiction of Vernor Vinge

From Grimm's World to Rainbows End, the fiction of Vernor Vinge.

https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
--
My reviews can be found at http://jamesdavisnicoll.com/
My tor pieces at https://www.tor.com/author/james-davis-nicoll/
My Dreamwidth at https://james-davis-nicoll.dreamwidth.org/
My patreon is at https://www.patreon.com/jamesdnicoll

On 3/26/2024 10:10 AM, James Nicoll wrote:
> A Brief Guide to the Fiction of Vernor Vinge
>
> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>
> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/

You seem to imply that Vinge invented the 'technological singularity'.

While I, among many others, first became aware of the idea through his
1993 essay[1], it predates him. John Von Neumann discussed the notion
back in 1958. I have no idea if Vinge knew or Von Neumann's
speculations, or invented the idea independently. It's one of those
'when it's steam engine time, it will steam engine' ideas.

Vinge predicted it would happen by 2023. While clearly we're not
there yet, the explosive development of generative AI clearly
seems a step in that direction.

[1] http://mindstalk.net/vinge/vinge-sing.html

pt

In article <utuumi$1u8bf$1@dont-email.me>,
Cryptoengineer <petertrei@gmail.com> wrote:
>On 3/26/2024 10:10 AM, James Nicoll wrote:
>> A Brief Guide to the Fiction of Vernor Vinge
>>
>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>
>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>
>You seem to imply that Vinge invented the 'technological singularity'.
>
>While I, among many others, first became aware of the idea through his
>1993 essay[1], it predates him. John Von Neumann discussed the notion
>back in 1958. I have no idea if Vinge knew or Von Neumann's
>speculations, or invented the idea independently. It's one of those
>'when it's steam engine time, it will steam engine' ideas.
>
>Vinge predicted it would happen by 2023. While clearly we're not
>there yet, the explosive development of generative AI clearly
>seems a step in that direction.

Clippy 2.0 does not a singularity make.
--
My reviews can be found at http://jamesdavisnicoll.com/
My tor pieces at https://www.tor.com/author/james-davis-nicoll/
My Dreamwidth at https://james-davis-nicoll.dreamwidth.org/
My patreon is at https://www.patreon.com/jamesdnicoll

On 26/03/2024 11.55, Cryptoengineer wrote:
> On 3/26/2024 10:10 AM, James Nicoll wrote:
>> A Brief Guide to the Fiction of Vernor Vinge
>>
>>  From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>
>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>
> You seem to imply that Vinge invented the 'technological singularity'.

Nor did he invent the Zones of Thought. That distinction belongs
to Poul Anderson, in _Brain Wave_. As a matter of fact, I've always
viewed _A Fire Upon the Deep_ as a tribute to _Brain Wave_.

Nothing wrong with that.

Also, my reaction to "Tatja Grimm" was similar to James'. I automatically
assumed that it was targeted for sale to JWC, Jr. -- even though I first
read it in _Orbit_.

--
Michael F. Stemper
Exodus 22:21

On Tue, 26 Mar 2024 12:55:44 -0400, Cryptoengineer
<petertrei@gmail.com> wrote:

>On 3/26/2024 10:10 AM, James Nicoll wrote:
>> A Brief Guide to the Fiction of Vernor Vinge
>>
>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>
>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>
>You seem to imply that Vinge invented the 'technological singularity'.
>
>While I, among many others, first became aware of the idea through his
>1993 essay[1], it predates him. John Von Neumann discussed the notion
>back in 1958. I have no idea if Vinge knew or Von Neumann's
>speculations, or invented the idea independently. It's one of those
>'when it's steam engine time, it will steam engine' ideas.
>
>Vinge predicted it would happen by 2023. While clearly we're not
>there yet, the explosive development of generative AI clearly
>seems a step in that direction.
>
>[1] http://mindstalk.net/vinge/vinge-sing.html

I don't remember where or when or how I became aware of it.

But I do know I haven't a lot about it recently.

Has the rate of progress slown down since, say, the 1990s? So that it
isn't quite so exponential?
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"

On 3/27/2024 9:21 AM, Paul S Person wrote:
> On Tue, 26 Mar 2024 12:55:44 -0400, Cryptoengineer
> <petertrei@gmail.com> wrote:
>
>> On 3/26/2024 10:10 AM, James Nicoll wrote:
>>> A Brief Guide to the Fiction of Vernor Vinge
>>>
>>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>>
>>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>>
>> You seem to imply that Vinge invented the 'technological singularity'.
>>
>> While I, among many others, first became aware of the idea through his
>> 1993 essay[1], it predates him. John Von Neumann discussed the notion
>> back in 1958. I have no idea if Vinge knew or Von Neumann's
>> speculations, or invented the idea independently. It's one of those
>> 'when it's steam engine time, it will steam engine' ideas.
>>
>> Vinge predicted it would happen by 2023. While clearly we're not
>> there yet, the explosive development of generative AI clearly
>> seems a step in that direction.
>>
>> [1] http://mindstalk.net/vinge/vinge-sing.html
>
> I don't remember where or when or how I became aware of it.
>
> But I do know I haven't a lot about it recently.
>
> Has the rate of progress slown down since, say, the 1990s? So that it
> isn't quite so exponential?

It is probably just "normal" now.

--
I've done good in this world. Now I'm tired and just want to be a cranky
dirty old man.

On 2024-03-27, Paul S Person <psperson@old.netcom.invalid> wrote:
> On Tue, 26 Mar 2024 12:55:44 -0400, Cryptoengineer
><petertrei@gmail.com> wrote:
>
>>On 3/26/2024 10:10 AM, James Nicoll wrote:
>>> A Brief Guide to the Fiction of Vernor Vinge
>>>
>>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>>
>>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>>
>>You seem to imply that Vinge invented the 'technological singularity'.
>>
>>While I, among many others, first became aware of the idea through his
>>1993 essay[1], it predates him. John Von Neumann discussed the notion
>>back in 1958. I have no idea if Vinge knew or Von Neumann's
>>speculations, or invented the idea independently. It's one of those
>>'when it's steam engine time, it will steam engine' ideas.
>>
>>Vinge predicted it would happen by 2023. While clearly we're not
>>there yet, the explosive development of generative AI clearly
>>seems a step in that direction.
>>
>>[1] http://mindstalk.net/vinge/vinge-sing.html
>
> I don't remember where or when or how I became aware of it.
>
> But I do know I haven't a lot about it recently.
>
> Has the rate of progress slown down since, say, the 1990s? So that it
> isn't quite so exponential?

The 1990s was not a period of large amount of progress on AI. Lots of
things were tried, small amounts of progress were made in related
disciplines (natural language parsing for example), but nothing really
advanced general AI.

One of the related disciplines that did have exponential growth was
mine: statistical information retrieval. At the beginning of the 90s,
state-of-the-art IR in practice was advanced Boolean systems on
possibly manually annotated documents. By the end of the 90s, with
the concurrent development of the web, it was natural language queries
on raw document text.

Towards the end of the 90s, people (including me) became aware that
large quantities of data (text) had a quality in itself. I had thoughts
of using IR to do AI taking advantage of this, but didn't do anything myself
though a grad student I mentored, Amit Singhal, later did. Groups like
IBM in the very early 2000s did develop this for what I consider the first
real progress in AI: question-answering. This was the group/system/approach
that was later developed Watson (that you may have seen on Jeopardy).

Side note: I wrote a fair number of IR papers with lots of citations
(30,000+), but I suspect the piece of writing that had the biggest
impact on IR was my one page post-PhD recommendation for Amit for
BellLabs++. The manager who hired him later told me that the sole
reason he offered Amit a position was my letter. It WAS rather
effusive, saying among other things that I thought he would have a
larger impact than any IR Grad student world-wide in the last 20
years. It turns out the manager knew me and knew I was not a
particularly gushy, over-enthusiastic person (can you tell from my
posts here:)) 3 years later in early 2000 Amit moved to be Head of
Research at an up-and-coming search engine company called Google,
rewrote the IR system, and the rest is history. AI was a pet interest
of Amit's throughout his 15 years as research head that he pushed as
much as possible; his goal was the ship's computer on original Star
Trek - something we're getting close to!

Overall, at Google and elsewhere, I view the start of exponential
growth of AI to be around 2000, sparked by first time availability of
very large data sets (whether text or images or eventually financial)
fast, cheap computer resources, and the realization that you didn't
need formal rules - the data was enough. It started reasonably slowly
but the exponential growth has not stopped yet. However, I will say
that right now many people are overestimating current AI capabilities
and underestimating the remaining problems (eg reliability). Lots of
hype!

Chris

On 3/28/2024 8:06 PM, Chris Buckley wrote:
> On 2024-03-27, Paul S Person <psperson@old.netcom.invalid> wrote:
>> On Tue, 26 Mar 2024 12:55:44 -0400, Cryptoengineer
>> <petertrei@gmail.com> wrote:
>>
>>> On 3/26/2024 10:10 AM, James Nicoll wrote:
>>>> A Brief Guide to the Fiction of Vernor Vinge
>>>>
>>>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>>>
>>>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>>>
>>> You seem to imply that Vinge invented the 'technological singularity'.
>>>
>>> While I, among many others, first became aware of the idea through his
>>> 1993 essay[1], it predates him. John Von Neumann discussed the notion
>>> back in 1958. I have no idea if Vinge knew or Von Neumann's
>>> speculations, or invented the idea independently. It's one of those
>>> 'when it's steam engine time, it will steam engine' ideas.
>>>
>>> Vinge predicted it would happen by 2023. While clearly we're not
>>> there yet, the explosive development of generative AI clearly
>>> seems a step in that direction.
>>>
>>> [1] http://mindstalk.net/vinge/vinge-sing.html
>>
>> I don't remember where or when or how I became aware of it.
>>
>> But I do know I haven't a lot about it recently.
>>
>> Has the rate of progress slown down since, say, the 1990s? So that it
>> isn't quite so exponential?
>
> The 1990s was not a period of large amount of progress on AI. Lots of
> things were tried, small amounts of progress were made in related
> disciplines (natural language parsing for example), but nothing really
> advanced general AI.
>
> One of the related disciplines that did have exponential growth was
> mine: statistical information retrieval. At the beginning of the 90s,
> state-of-the-art IR in practice was advanced Boolean systems on
> possibly manually annotated documents. By the end of the 90s, with
> the concurrent development of the web, it was natural language queries
> on raw document text.
>
> Towards the end of the 90s, people (including me) became aware that
> large quantities of data (text) had a quality in itself. I had thoughts
> of using IR to do AI taking advantage of this, but didn't do anything myself
> though a grad student I mentored, Amit Singhal, later did. Groups like
> IBM in the very early 2000s did develop this for what I consider the first
> real progress in AI: question-answering. This was the group/system/approach
> that was later developed Watson (that you may have seen on Jeopardy).
>
> Side note: I wrote a fair number of IR papers with lots of citations
> (30,000+), but I suspect the piece of writing that had the biggest
> impact on IR was my one page post-PhD recommendation for Amit for
> BellLabs++. The manager who hired him later told me that the sole
> reason he offered Amit a position was my letter. It WAS rather
> effusive, saying among other things that I thought he would have a
> larger impact than any IR Grad student world-wide in the last 20
> years. It turns out the manager knew me and knew I was not a
> particularly gushy, over-enthusiastic person (can you tell from my
> posts here:)) 3 years later in early 2000 Amit moved to be Head of
> Research at an up-and-coming search engine company called Google,
> rewrote the IR system, and the rest is history. AI was a pet interest
> of Amit's throughout his 15 years as research head that he pushed as
> much as possible; his goal was the ship's computer on original Star
> Trek - something we're getting close to!
>
> Overall, at Google and elsewhere, I view the start of exponential
> growth of AI to be around 2000, sparked by first time availability of
> very large data sets (whether text or images or eventually financial)
> fast, cheap computer resources, and the realization that you didn't
> need formal rules - the data was enough. It started reasonably slowly
> but the exponential growth has not stopped yet. However, I will say
> that right now many people are overestimating current AI capabilities
> and underestimating the remaining problems (eg reliability). Lots of
> hype!
>
> Chris

Having been a software developer since 1975, I am really wondering about
the AI hype. The AI thing so far seems to be a major expansion of the
old Eliza program. Of course, I am really outdated, writing in Fortran
and C++ nowadays. I have about two million lines of F77 / C++ code that
I am shepherding around the place for several customers.
https://en.wikipedia.org/wiki/ELIZA

Lynn

On Thu, 28 Mar 2024, Lynn McGuire wrote:

> On 3/28/2024 8:06 PM, Chris Buckley wrote:
>> On 2024-03-27, Paul S Person <psperson@old.netcom.invalid> wrote:
>>> On Tue, 26 Mar 2024 12:55:44 -0400, Cryptoengineer
>>> <petertrei@gmail.com> wrote:
>>>
>>>> On 3/26/2024 10:10 AM, James Nicoll wrote:
>>>>> A Brief Guide to the Fiction of Vernor Vinge
>>>>>
>>>>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>>>>
>>>>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>>>>
>>>> You seem to imply that Vinge invented the 'technological singularity'.
>>>>
>>>> While I, among many others, first became aware of the idea through his
>>>> 1993 essay[1], it predates him. John Von Neumann discussed the notion
>>>> back in 1958. I have no idea if Vinge knew or Von Neumann's
>>>> speculations, or invented the idea independently. It's one of those
>>>> 'when it's steam engine time, it will steam engine' ideas.
>>>>
>>>> Vinge predicted it would happen by 2023. While clearly we're not
>>>> there yet, the explosive development of generative AI clearly
>>>> seems a step in that direction.
>>>>
>>>> [1] http://mindstalk.net/vinge/vinge-sing.html
>>>
>>> I don't remember where or when or how I became aware of it.
>>>
>>> But I do know I haven't a lot about it recently.
>>>
>>> Has the rate of progress slown down since, say, the 1990s? So that it
>>> isn't quite so exponential?
>>
>> The 1990s was not a period of large amount of progress on AI. Lots of
>> things were tried, small amounts of progress were made in related
>> disciplines (natural language parsing for example), but nothing really
>> advanced general AI.
>>
>> One of the related disciplines that did have exponential growth was
>> mine: statistical information retrieval. At the beginning of the 90s,
>> state-of-the-art IR in practice was advanced Boolean systems on
>> possibly manually annotated documents. By the end of the 90s, with
>> the concurrent development of the web, it was natural language queries
>> on raw document text.
>>
>> Towards the end of the 90s, people (including me) became aware that
>> large quantities of data (text) had a quality in itself. I had thoughts
>> of using IR to do AI taking advantage of this, but didn't do anything
>> myself
>> though a grad student I mentored, Amit Singhal, later did. Groups like
>> IBM in the very early 2000s did develop this for what I consider the first
>> real progress in AI: question-answering. This was the
>> group/system/approach
>> that was later developed Watson (that you may have seen on Jeopardy).
>>
>> Side note: I wrote a fair number of IR papers with lots of citations
>> (30,000+), but I suspect the piece of writing that had the biggest
>> impact on IR was my one page post-PhD recommendation for Amit for
>> BellLabs++. The manager who hired him later told me that the sole
>> reason he offered Amit a position was my letter. It WAS rather
>> effusive, saying among other things that I thought he would have a
>> larger impact than any IR Grad student world-wide in the last 20
>> years. It turns out the manager knew me and knew I was not a
>> particularly gushy, over-enthusiastic person (can you tell from my
>> posts here:)) 3 years later in early 2000 Amit moved to be Head of
>> Research at an up-and-coming search engine company called Google,
>> rewrote the IR system, and the rest is history. AI was a pet interest
>> of Amit's throughout his 15 years as research head that he pushed as
>> much as possible; his goal was the ship's computer on original Star
>> Trek - something we're getting close to!
>>
>> Overall, at Google and elsewhere, I view the start of exponential
>> growth of AI to be around 2000, sparked by first time availability of
>> very large data sets (whether text or images or eventually financial)
>> fast, cheap computer resources, and the realization that you didn't
>> need formal rules - the data was enough. It started reasonably slowly
>> but the exponential growth has not stopped yet. However, I will say
>> that right now many people are overestimating current AI capabilities
>> and underestimating the remaining problems (eg reliability). Lots of
>> hype!
>>
>> Chris
>
> Having been a software developer since 1975, I am really wondering about the
> AI hype. The AI thing so far seems to be a major expansion of the old Eliza
> program. Of course, I am really outdated, writing in Fortran and C++
> nowadays. I have about two million lines of F77 / C++ code that I am
> shepherding around the place for several customers.
> https://en.wikipedia.org/wiki/ELIZA
>
> Lynn

I agree. I am also not super impressed. I've tried the common ones and for
my use cases I find that I still need to proof read, change and rewrite.
The same goes for code.

So I can just as well come up with the text/code myself, with the added
advantage that I then know exactly what I did.

But based on what I am reading, there are ninjas who seem to have fused
with their favourite AI of choice and feel they are much more productive.

That said, I think the progress from Eliza is great, and if the progress
continues so that I don't have to do what I have to do today, then it's a
valuable tool for corporate BS and for simple coding tasks (parse file,
convert to csv, give me a basic web site setup in framework X).

On 2024-03-29, D <nospam@example.net> wrote:
>
>
> On Thu, 28 Mar 2024, Lynn McGuire wrote:
>
>> On 3/28/2024 8:06 PM, Chris Buckley wrote:
>>> On 2024-03-27, Paul S Person <psperson@old.netcom.invalid> wrote:
>>>> On Tue, 26 Mar 2024 12:55:44 -0400, Cryptoengineer
>>>> <petertrei@gmail.com> wrote:
>>>>
>>>>> On 3/26/2024 10:10 AM, James Nicoll wrote:
>>>>>> A Brief Guide to the Fiction of Vernor Vinge
>>>>>>
>>>>>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>>>>>
>>>>>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>>>>>
>>>>> You seem to imply that Vinge invented the 'technological singularity'.
>>>>>
>>>>> While I, among many others, first became aware of the idea through his
>>>>> 1993 essay[1], it predates him. John Von Neumann discussed the notion
>>>>> back in 1958. I have no idea if Vinge knew or Von Neumann's
>>>>> speculations, or invented the idea independently. It's one of those
>>>>> 'when it's steam engine time, it will steam engine' ideas.
>>>>>
>>>>> Vinge predicted it would happen by 2023. While clearly we're not
>>>>> there yet, the explosive development of generative AI clearly
>>>>> seems a step in that direction.
>>>>>
>>>>> [1] http://mindstalk.net/vinge/vinge-sing.html
>>>>
>>>> I don't remember where or when or how I became aware of it.
>>>>
>>>> But I do know I haven't a lot about it recently.
>>>>
>>>> Has the rate of progress slown down since, say, the 1990s? So that it
>>>> isn't quite so exponential?
>>>
>>> The 1990s was not a period of large amount of progress on AI. Lots of
>>> things were tried, small amounts of progress were made in related
>>> disciplines (natural language parsing for example), but nothing really
>>> advanced general AI.
>>>
>>> One of the related disciplines that did have exponential growth was
>>> mine: statistical information retrieval. At the beginning of the 90s,
>>> state-of-the-art IR in practice was advanced Boolean systems on
>>> possibly manually annotated documents. By the end of the 90s, with
>>> the concurrent development of the web, it was natural language queries
>>> on raw document text.
>>>
>>> Towards the end of the 90s, people (including me) became aware that
>>> large quantities of data (text) had a quality in itself. I had thoughts
>>> of using IR to do AI taking advantage of this, but didn't do anything
>>> myself
>>> though a grad student I mentored, Amit Singhal, later did. Groups like
>>> IBM in the very early 2000s did develop this for what I consider the first
>>> real progress in AI: question-answering. This was the
>>> group/system/approach
>>> that was later developed Watson (that you may have seen on Jeopardy).
>>>
>>> Side note: I wrote a fair number of IR papers with lots of citations
>>> (30,000+), but I suspect the piece of writing that had the biggest
>>> impact on IR was my one page post-PhD recommendation for Amit for
>>> BellLabs++. The manager who hired him later told me that the sole
>>> reason he offered Amit a position was my letter. It WAS rather
>>> effusive, saying among other things that I thought he would have a
>>> larger impact than any IR Grad student world-wide in the last 20
>>> years. It turns out the manager knew me and knew I was not a
>>> particularly gushy, over-enthusiastic person (can you tell from my
>>> posts here:)) 3 years later in early 2000 Amit moved to be Head of
>>> Research at an up-and-coming search engine company called Google,
>>> rewrote the IR system, and the rest is history. AI was a pet interest
>>> of Amit's throughout his 15 years as research head that he pushed as
>>> much as possible; his goal was the ship's computer on original Star
>>> Trek - something we're getting close to!
>>>
>>> Overall, at Google and elsewhere, I view the start of exponential
>>> growth of AI to be around 2000, sparked by first time availability of
>>> very large data sets (whether text or images or eventually financial)
>>> fast, cheap computer resources, and the realization that you didn't
>>> need formal rules - the data was enough. It started reasonably slowly
>>> but the exponential growth has not stopped yet. However, I will say
>>> that right now many people are overestimating current AI capabilities
>>> and underestimating the remaining problems (eg reliability). Lots of
>>> hype!
>>>
>>> Chris
>>
>> Having been a software developer since 1975, I am really wondering about the
>> AI hype. The AI thing so far seems to be a major expansion of the old Eliza
>> program. Of course, I am really outdated, writing in Fortran and C++
>> nowadays. I have about two million lines of F77 / C++ code that I am
>> shepherding around the place for several customers.
>> https://en.wikipedia.org/wiki/ELIZA
>>
>> Lynn
>
> I agree. I am also not super impressed. I've tried the common ones and for
> my use cases I find that I still need to proof read, change and rewrite.
> The same goes for code.
>
> So I can just as well come up with the text/code myself, with the added
> advantage that I then know exactly what I did.
>
> But based on what I am reading, there are ninjas who seem to have fused
> with their favourite AI of choice and feel they are much more productive.
>
> That said, I think the progress from Eliza is great, and if the progress
> continues so that I don't have to do what I have to do today, then it's a
> valuable tool for corporate BS and for simple coding tasks (parse file,
> convert to csv, give me a basic web site setup in framework X).

The progress of AI has been truly impressive the last couple of
decades. But one of the major problems is that general AI is
extremely hard to evaluate well. There are many different criteria
that need to be used and humans disagree deeply on almost every
one of them. Eg, what does it mean to be fair?

Ellen Voorhees was the TREC project manager at NIST for decades until
she retired last year. TREC has been the leading text evaluation
workshop/conference since the early 1990s. Her last couple of years at
NIST were spent trying to come up with tasks upon which general AI
could be evaluated. She did not succeed. Just too big of a mismatch
between the breadth of what "intelligence" means and the human
disagreements on the narrow criteria that can be evaluated in a task.
It was just too messy. (I heard about it often; we've been married
for over 40 years.)

One of the problems with the general public's opinion of AI is that
the fluency of the new models is just too good! That sort of fluency
in humans is only achievable by adult experts but the generative AI
models are masking all their understanding gaps, which may be equivalent
to a 10 year old child.

To my mind, a major proof that we have reached the area of true learning,
is the performance of AlphaGo and then AlphaZero 7 or 8 years ago. The
massive data there was their own game simulations and it was a very restricted
world of chess, go, and other games but to start with just the rules, come up
with their own criteria for "good moves", and become better than anybody
else in the world in less than a day is impressive! This is the type
of learning that is going into general Ai models of today.

Chris

Lynn McGuire <lynnmcguire5@gmail.com> wrote:
>
>Having been a software developer since 1975, I am really wondering about
>the AI hype. The AI thing so far seems to be a major expansion of the
>old Eliza program. Of course, I am really outdated, writing in Fortran
>and C++ nowadays. I have about two million lines of F77 / C++ code that
>I am shepherding around the place for several customers.
> https://en.wikipedia.org/wiki/ELIZA

Don't think about it as Eliza, because Eliza is rules-based and you can
look into the box.

Think about machine learning as giant matrix. Stuff goes in and we tweak
the matrix coefficients so that the stuff comes out that we want when we
multiply the matrix by the stuff that went in. After the learning period
is over, we put completely new stuff in, and other completely new stuff
comes out that hopefully is related in the same way as the data we trained
it on.

It's not really rules-based and what something weird comes out, it is very
difficult to go back and work out what data in the training process caused
that weird association to occur. Often it is impossible.

If you remember the McCulloch and Pitts Perceptron machine in the sixties,
that is basically the great grandfather of the modern machine learning
systems. In the end, connectionism won and traditional AI methods lost.
--scott

--
"C'est un Nagra. C'est suisse, et tres, tres precis."

On Fri, 29 Mar 2024, Chris Buckley wrote:

> On 2024-03-29, D <nospam@example.net> wrote:
>>
>>
>> On Thu, 28 Mar 2024, Lynn McGuire wrote:
>>
>>> On 3/28/2024 8:06 PM, Chris Buckley wrote:
>>>> On 2024-03-27, Paul S Person <psperson@old.netcom.invalid> wrote:
>>>>> On Tue, 26 Mar 2024 12:55:44 -0400, Cryptoengineer
>>>>> <petertrei@gmail.com> wrote:
>>>>>
>>>>>> On 3/26/2024 10:10 AM, James Nicoll wrote:
>>>>>>> A Brief Guide to the Fiction of Vernor Vinge
>>>>>>>
>>>>>>> From Grimm's World to Rainbows End, the fiction of Vernor Vinge.
>>>>>>>
>>>>>>> https://reactormag.com/a-brief-guide-to-the-fiction-of-vernor-vinge/
>>>>>>
>>>>>> You seem to imply that Vinge invented the 'technological singularity'.
>>>>>>
>>>>>> While I, among many others, first became aware of the idea through his
>>>>>> 1993 essay[1], it predates him. John Von Neumann discussed the notion
>>>>>> back in 1958. I have no idea if Vinge knew or Von Neumann's
>>>>>> speculations, or invented the idea independently. It's one of those
>>>>>> 'when it's steam engine time, it will steam engine' ideas.
>>>>>>
>>>>>> Vinge predicted it would happen by 2023. While clearly we're not
>>>>>> there yet, the explosive development of generative AI clearly
>>>>>> seems a step in that direction.
>>>>>>
>>>>>> [1] http://mindstalk.net/vinge/vinge-sing.html
>>>>>
>>>>> I don't remember where or when or how I became aware of it.
>>>>>
>>>>> But I do know I haven't a lot about it recently.
>>>>>
>>>>> Has the rate of progress slown down since, say, the 1990s? So that it
>>>>> isn't quite so exponential?
>>>>
>>>> The 1990s was not a period of large amount of progress on AI. Lots of
>>>> things were tried, small amounts of progress were made in related
>>>> disciplines (natural language parsing for example), but nothing really
>>>> advanced general AI.
>>>>
>>>> One of the related disciplines that did have exponential growth was
>>>> mine: statistical information retrieval. At the beginning of the 90s,
>>>> state-of-the-art IR in practice was advanced Boolean systems on
>>>> possibly manually annotated documents. By the end of the 90s, with
>>>> the concurrent development of the web, it was natural language queries
>>>> on raw document text.
>>>>
>>>> Towards the end of the 90s, people (including me) became aware that
>>>> large quantities of data (text) had a quality in itself. I had thoughts
>>>> of using IR to do AI taking advantage of this, but didn't do anything
>>>> myself
>>>> though a grad student I mentored, Amit Singhal, later did. Groups like
>>>> IBM in the very early 2000s did develop this for what I consider the first
>>>> real progress in AI: question-answering. This was the
>>>> group/system/approach
>>>> that was later developed Watson (that you may have seen on Jeopardy).
>>>>
>>>> Side note: I wrote a fair number of IR papers with lots of citations
>>>> (30,000+), but I suspect the piece of writing that had the biggest
>>>> impact on IR was my one page post-PhD recommendation for Amit for
>>>> BellLabs++. The manager who hired him later told me that the sole
>>>> reason he offered Amit a position was my letter. It WAS rather
>>>> effusive, saying among other things that I thought he would have a
>>>> larger impact than any IR Grad student world-wide in the last 20
>>>> years. It turns out the manager knew me and knew I was not a
>>>> particularly gushy, over-enthusiastic person (can you tell from my
>>>> posts here:)) 3 years later in early 2000 Amit moved to be Head of
>>>> Research at an up-and-coming search engine company called Google,
>>>> rewrote the IR system, and the rest is history. AI was a pet interest
>>>> of Amit's throughout his 15 years as research head that he pushed as
>>>> much as possible; his goal was the ship's computer on original Star
>>>> Trek - something we're getting close to!
>>>>
>>>> Overall, at Google and elsewhere, I view the start of exponential
>>>> growth of AI to be around 2000, sparked by first time availability of
>>>> very large data sets (whether text or images or eventually financial)
>>>> fast, cheap computer resources, and the realization that you didn't
>>>> need formal rules - the data was enough. It started reasonably slowly
>>>> but the exponential growth has not stopped yet. However, I will say
>>>> that right now many people are overestimating current AI capabilities
>>>> and underestimating the remaining problems (eg reliability). Lots of
>>>> hype!
>>>>
>>>> Chris
>>>
>>> Having been a software developer since 1975, I am really wondering about the
>>> AI hype. The AI thing so far seems to be a major expansion of the old Eliza
>>> program. Of course, I am really outdated, writing in Fortran and C++
>>> nowadays. I have about two million lines of F77 / C++ code that I am
>>> shepherding around the place for several customers.
>>> https://en.wikipedia.org/wiki/ELIZA
>>>
>>> Lynn
>>
>> I agree. I am also not super impressed. I've tried the common ones and for
>> my use cases I find that I still need to proof read, change and rewrite.
>> The same goes for code.
>>
>> So I can just as well come up with the text/code myself, with the added
>> advantage that I then know exactly what I did.
>>
>> But based on what I am reading, there are ninjas who seem to have fused
>> with their favourite AI of choice and feel they are much more productive.
>>
>> That said, I think the progress from Eliza is great, and if the progress
>> continues so that I don't have to do what I have to do today, then it's a
>> valuable tool for corporate BS and for simple coding tasks (parse file,
>> convert to csv, give me a basic web site setup in framework X).
>
> The progress of AI has been truly impressive the last couple of
> decades. But one of the major problems is that general AI is
> extremely hard to evaluate well. There are many different criteria
> that need to be used and humans disagree deeply on almost every
> one of them. Eg, what does it mean to be fair?
>
> Ellen Voorhees was the TREC project manager at NIST for decades until
> she retired last year. TREC has been the leading text evaluation
> workshop/conference since the early 1990s. Her last couple of years at
> NIST were spent trying to come up with tasks upon which general AI
> could be evaluated. She did not succeed. Just too big of a mismatch
> between the breadth of what "intelligence" means and the human
> disagreements on the narrow criteria that can be evaluated in a task.
> It was just too messy. (I heard about it often; we've been married
> for over 40 years.)
>
> One of the problems with the general public's opinion of AI is that
> the fluency of the new models is just too good! That sort of fluency
> in humans is only achievable by adult experts but the generative AI
> models are masking all their understanding gaps, which may be equivalent
> to a 10 year old child.
>
> To my mind, a major proof that we have reached the area of true learning,
> is the performance of AlphaGo and then AlphaZero 7 or 8 years ago. The
> massive data there was their own game simulations and it was a very restricted
> world of chess, go, and other games but to start with just the rules, come up
> with their own criteria for "good moves", and become better than anybody
> else in the world in less than a day is impressive! This is the type
> of learning that is going into general Ai models of today.
>
> Chris
>

Yes, if we're talking narrow, focused use cases such as chess and go,
massive progress indeed. Depending on how you define it, you could add
great progress when it comes to self driving cars, robots, image
recognition and voice recognition too.

On 30/03/2024 06.12, D wrote:
>
>
> On Fri, 29 Mar 2024, Chris Buckley wrote:
>
>> On 2024-03-29, D <nospam@example.net> wrote:
>>>
>>>
>>> On Thu, 28 Mar 2024, Lynn McGuire wrote:

>>>> Having been a software developer since 1975, I am really wondering about the
>>>> AI hype.  The AI thing so far seems to be a major expansion of the old Eliza
>>>> program.

>>> That said, I think the progress from Eliza is great, and if the progress
>>> continues so that I don't have to do what I have to do today, then it's a
>>> valuable tool for corporate BS and for simple coding tasks (parse file,
>>> convert to csv, give me a basic web site setup in framework X).
>>
>> The progress of AI has been truly impressive the last couple of
>> decades.  But one of the major problems is that general AI is
>> extremely hard to evaluate well. There are many different criteria
>> that need to be used and humans disagree deeply on almost every
>> one of them.  Eg, what does it mean to be fair?

>> One of the problems with the general public's opinion of AI is that
>> the fluency of the new models is just too good!  That sort of fluency
>> in humans is only achievable by adult experts but the generative AI
>> models are masking all their understanding gaps, which may be equivalent
>> to a 10 year old child.
>>
>> To my mind, a major proof that we have reached the area of true learning,
>> is the performance of AlphaGo and then AlphaZero 7 or 8 years ago.  The
>> massive data there was their own game simulations and it was a very restricted
>> world of chess, go, and other games but to start with just the rules, come up
>> with their own criteria for "good moves", and become better than anybody
>> else in the world in less than a day is impressive!  This is the type
>> of learning that is going into general Ai models of today.

> Yes, if we're talking narrow, focused use cases such as chess and go, massive progress indeed. Depending on how you define it, you could add great progress when it comes to self driving cars, robots, image recognition and voice recognition too.

If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
Douglas Hofstader commented that "AI is sometimes just whatever hasn't
been done yet."

--
Michael F. Stemper
Galatians 3:28

On Sat, 30 Mar 2024 08:27:28 -0500, "Michael F. Stemper"
<michael.stemper@gmail.com> wrote:

<snippo>

>If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>been done yet."

That dovetails with something I've run into: when a bit of AI research
produces something /useful/ it ceases to be "AI" and is called
something else. "AI" is always used for pure research, unpolluted by
tawdry commercialization.
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"

On Sat, 30 Mar 2024, Paul S Person wrote:

> On Sat, 30 Mar 2024 08:27:28 -0500, "Michael F. Stemper"
> <michael.stemper@gmail.com> wrote:
>
> <snippo>
>
>> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>> been done yet."
>
> That dovetails with something I've run into: when a bit of AI research
> produces something /useful/ it ceases to be "AI" and is called
> something else. "AI" is always used for pure research, unpolluted by
> tawdry commercialization.
>

Agreed. I read the same somewhere and I agree.

"Michael F. Stemper" <michael.stemper@gmail.com> writes:
>On 30/03/2024 06.12, D wrote:
>>
>>
>> On Fri, 29 Mar 2024, Chris Buckley wrote:
>>
>>> On 2024-03-29, D <nospam@example.net> wrote:
>>>>
>>>>
>>>> On Thu, 28 Mar 2024, Lynn McGuire wrote:
>
>>>>> Having been a software developer since 1975, I am really wondering about the
>>>>> AI hype.  The AI thing so far seems to be a major expansion of the old Eliza
>>>>> program.
>
>>>> That said, I think the progress from Eliza is great, and if the progress
>>>> continues so that I don't have to do what I have to do today, then it's a
>>>> valuable tool for corporate BS and for simple coding tasks (parse file,
>>>> convert to csv, give me a basic web site setup in framework X).
>>>
>>> The progress of AI has been truly impressive the last couple of
>>> decades.  But one of the major problems is that general AI is
>>> extremely hard to evaluate well. There are many different criteria
>>> that need to be used and humans disagree deeply on almost every
>>> one of them.  Eg, what does it mean to be fair?
>
>>> One of the problems with the general public's opinion of AI is that
>>> the fluency of the new models is just too good!  That sort of fluency
>>> in humans is only achievable by adult experts but the generative AI
>>> models are masking all their understanding gaps, which may be equivalent
>>> to a 10 year old child.
>>>
>>> To my mind, a major proof that we have reached the area of true learning,
>>> is the performance of AlphaGo and then AlphaZero 7 or 8 years ago.  The
>>> massive data there was their own game simulations and it was a very restricted
>>> world of chess, go, and other games but to start with just the rules, come up
>>> with their own criteria for "good moves", and become better than anybody
>>> else in the world in less than a day is impressive!  This is the type
>>> of learning that is going into general Ai models of today.
>
>> Yes, if we're talking narrow, focused use cases such as chess and go, massive progress indeed. Depending on how you define it, you could add great progress when it comes to self driving cars, robots, image recognition and voice recognition too.
>
>If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>been done yet."

But none of what is called "AI" today is really AI. It's just machine learning
and pattern matching with a massive database of patterns. Eliza writ large.

On 30/03/2024 14.33, Scott Lurndal wrote:
> "Michael F. Stemper" <michael.stemper@gmail.com> writes:
>> On 30/03/2024 06.12, D wrote:

>>> Yes, if we're talking narrow, focused use cases such as chess and go, massive progress indeed. Depending on how you define it, you could add great progress when it comes to self driving cars, robots, image recognition and voice recognition too.
>>
>> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>> been done yet."
>
> But none of what is called "AI" today is really AI. It's just machine learning
> and pattern matching with a massive database of patterns. Eliza writ large.

Agreed. If it learns more than word patterns, I might be impressed. But
right now, it's just textual manipulation. When the words start to have
meaning to the program(s), then we'll really be approaching the Singularity.

As far as I can tell, though, it's not aware that there's conflicting
information out there. If you fed a bunch of Electric Universe text into
one of these programs, it'd be just as happy parroting it back as it
would GR. It wouldn't realize that the two "systems" are in conflict
with each other.

ObSF: _Stand on Zanzibar_, in which Shalmaneser wouldn't accept new information
that conflicted with what it had stored unless you said "What I tell you
three times is true."

--
Michael F. Stemper
There's no "me" in "team". There's no "us" in "team", either.

Paul S Person <psperson@old.netcom.invalid> wrote:
>That dovetails with something I've run into: when a bit of AI research
>produces something /useful/ it ceases to be "AI" and is called
>something else. "AI" is always used for pure research, unpolluted by
>tawdry commercialization.

This is a longstanding tradition. When I was in grad school, expert systems
were AI, but by the time I was out, expert systems weren't AI anymore.
(In the end expert systems didn't turn out to be useful either but that
is a separate issue).

But now we have come to a weird point in time when "AI" seems to be synonomous
with "machine learning systems."
--scott
--
"C'est un Nagra. C'est suisse, et tres, tres precis."

On 3/30/2024 9:27 AM, Michael F. Stemper wrote:
> On 30/03/2024 06.12, D wrote:
>>
>>
>> On Fri, 29 Mar 2024, Chris Buckley wrote:
>>
>>> On 2024-03-29, D <nospam@example.net> wrote:
>>>>
>>>>
>>>> On Thu, 28 Mar 2024, Lynn McGuire wrote:
>
>>>>> Having been a software developer since 1975, I am really wondering
>>>>> about the
>>>>> AI hype.  The AI thing so far seems to be a major expansion of the
>>>>> old Eliza
>>>>> program.
>
>>>> That said, I think the progress from Eliza is great, and if the
>>>> progress
>>>> continues so that I don't have to do what I have to do today, then
>>>> it's a
>>>> valuable tool for corporate BS and for simple coding tasks (parse file,
>>>> convert to csv, give me a basic web site setup in framework X).
>>>
>>> The progress of AI has been truly impressive the last couple of
>>> decades.  But one of the major problems is that general AI is
>>> extremely hard to evaluate well. There are many different criteria
>>> that need to be used and humans disagree deeply on almost every
>>> one of them.  Eg, what does it mean to be fair?
>
>>> One of the problems with the general public's opinion of AI is that
>>> the fluency of the new models is just too good!  That sort of fluency
>>> in humans is only achievable by adult experts but the generative AI
>>> models are masking all their understanding gaps, which may be equivalent
>>> to a 10 year old child.
>>>
>>> To my mind, a major proof that we have reached the area of true
>>> learning,
>>> is the performance of AlphaGo and then AlphaZero 7 or 8 years ago.  The
>>> massive data there was their own game simulations and it was a very
>>> restricted
>>> world of chess, go, and other games but to start with just the rules,
>>> come up
>>> with their own criteria for "good moves", and become better than anybody
>>> else in the world in less than a day is impressive!  This is the type
>>> of learning that is going into general Ai models of today.
>
>> Yes, if we're talking narrow, focused use cases such as chess and go,
>> massive progress indeed. Depending on how you define it, you could add
>> great progress when it comes to self driving cars, robots, image
>> recognition and voice recognition too.
>
> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
> been done yet."

Its worth remembering that when Hofstader wrote that, AI was in the
doldrums, with expert systems stagnating.

My wife, who was working on an expert system a the time (XCON, a tool
for configuring Vax systems at DEC), recalls that then people would
resist even putting the term 'Artificial Intelligence' on their resumes,
since it could well lead to rejections.

pt

On 2024-03-30, Michael F. Stemper <michael.stemper@gmail.com> wrote:
> On 30/03/2024 14.33, Scott Lurndal wrote:
>> "Michael F. Stemper" <michael.stemper@gmail.com> writes:
>>> On 30/03/2024 06.12, D wrote:
>
>>>> Yes, if we're talking narrow, focused use cases such as chess and go, massive progress indeed. Depending on how you define it, you could add great progress when it comes to self driving cars, robots, image recognition and voice recognition too.
>>>
>>> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>>> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>>> been done yet."
>>
>> But none of what is called "AI" today is really AI. It's just machine learning
>> and pattern matching with a massive database of patterns. Eliza writ large.
>
> Agreed. If it learns more than word patterns, I might be impressed. But
> right now, it's just textual manipulation. When the words start to have
> meaning to the program(s), then we'll really be approaching the Singularity.
>
> As far as I can tell, though, it's not aware that there's conflicting
> information out there. If you fed a bunch of Electric Universe text into
> one of these programs, it'd be just as happy parroting it back as it
> would GR. It wouldn't realize that the two "systems" are in conflict
> with each other.

All you folks are redefining AI once again, twisting it to mean
something unrecognizable (to me at least). It is not Artificial
Consciousness; it is not Artificial Embodiment. It is Artificial
*Intelligence*.

Is a fish intelligent? Are some dogs more intelligent than others?

I included my paragraph on AlphaZero for a reason. AlphaZero is given
the rules of the game but no other game info. It took 34 hours of
playing itself at Go to become the best in the world. (The even more
impressive MuZero doesn't even have the rules of the game programmed
in.)

It is not just pattern matching. By 50 moves of Go with strong
probability the board is in a position that it has never simulated,
and games to completion go 200+ moves. Every move and stone on the
board is important. AlphaZero is learning high level concepts on its
own. It then uses those high level concepts to only consider
reasonable moves. (It does several orders of magnitude fewer
simulations at game time than other programs, at least at chess.)

How is this not intelligence? What would AlphaZero have to do to show
intelligence in your opinion?

Chris

On 3/30/2024 7:50 PM, Chris Buckley wrote:
> On 2024-03-30, Michael F. Stemper <michael.stemper@gmail.com> wrote:
>> On 30/03/2024 14.33, Scott Lurndal wrote:
>>> "Michael F. Stemper" <michael.stemper@gmail.com> writes:
>>>> On 30/03/2024 06.12, D wrote:
>>
>>>>> Yes, if we're talking narrow, focused use cases such as chess and go, massive progress indeed. Depending on how you define it, you could add great progress when it comes to self driving cars, robots, image recognition and voice recognition too.
>>>>
>>>> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>>>> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>>>> been done yet."
>>>
>>> But none of what is called "AI" today is really AI. It's just machine learning
>>> and pattern matching with a massive database of patterns. Eliza writ large.
>>
>> Agreed. If it learns more than word patterns, I might be impressed. But
>> right now, it's just textual manipulation. When the words start to have
>> meaning to the program(s), then we'll really be approaching the Singularity.
>>
>> As far as I can tell, though, it's not aware that there's conflicting
>> information out there. If you fed a bunch of Electric Universe text into
>> one of these programs, it'd be just as happy parroting it back as it
>> would GR. It wouldn't realize that the two "systems" are in conflict
>> with each other.
>
> All you folks are redefining AI once again, twisting it to mean
> something unrecognizable (to me at least). It is not Artificial
> Consciousness; it is not Artificial Embodiment. It is Artificial
> *Intelligence*.
>
> Is a fish intelligent? Are some dogs more intelligent than others?
>
> I included my paragraph on AlphaZero for a reason. AlphaZero is given
> the rules of the game but no other game info. It took 34 hours of
> playing itself at Go to become the best in the world. (The even more
> impressive MuZero doesn't even have the rules of the game programmed
> in.)
>
> It is not just pattern matching. By 50 moves of Go with strong
> probability the board is in a position that it has never simulated,
> and games to completion go 200+ moves. Every move and stone on the
> board is important. AlphaZero is learning high level concepts on its
> own. It then uses those high level concepts to only consider
> reasonable moves. (It does several orders of magnitude fewer
> simulations at game time than other programs, at least at chess.)
>
> How is this not intelligence? What would AlphaZero have to do to show
> intelligence in your opinion?
>
I would say that is highly advanced number crunching. It can run
predictive simulations much farther into the "future" than past systems
but that is just more advanced computation hardware. But the main
problem with claiming successful "artificial intelligence" is you first
have to _define_ "intelligence" in some kind of measurable, objective
manner. As far as I know we can't.

--
I've done good in this world. Now I'm tired and just want to be a cranky
dirty old man.

On 2024-03-31, Dimensional Traveler <dtravel@sonic.net> wrote:
> On 3/30/2024 7:50 PM, Chris Buckley wrote:
>> On 2024-03-30, Michael F. Stemper <michael.stemper@gmail.com> wrote:
>>> On 30/03/2024 14.33, Scott Lurndal wrote:
>>>> "Michael F. Stemper" <michael.stemper@gmail.com> writes:
>>>>> On 30/03/2024 06.12, D wrote:
>>>
>>>>>> Yes, if we're talking narrow, focused use cases such as chess and go, massive progress indeed. Depending on how you define it, you could add great progress when it comes to self driving cars, robots, image recognition and voice recognition too.
>>>>>
>>>>> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>>>>> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>>>>> been done yet."
>>>>
>>>> But none of what is called "AI" today is really AI. It's just machine learning
>>>> and pattern matching with a massive database of patterns. Eliza writ large.
>>>
>>> Agreed. If it learns more than word patterns, I might be impressed. But
>>> right now, it's just textual manipulation. When the words start to have
>>> meaning to the program(s), then we'll really be approaching the Singularity.
>>>
>>> As far as I can tell, though, it's not aware that there's conflicting
>>> information out there. If you fed a bunch of Electric Universe text into
>>> one of these programs, it'd be just as happy parroting it back as it
>>> would GR. It wouldn't realize that the two "systems" are in conflict
>>> with each other.
>>
>> All you folks are redefining AI once again, twisting it to mean
>> something unrecognizable (to me at least). It is not Artificial
>> Consciousness; it is not Artificial Embodiment. It is Artificial
>> *Intelligence*.
>>
>> Is a fish intelligent? Are some dogs more intelligent than others?
>>
>> I included my paragraph on AlphaZero for a reason. AlphaZero is given
>> the rules of the game but no other game info. It took 34 hours of
>> playing itself at Go to become the best in the world. (The even more
>> impressive MuZero doesn't even have the rules of the game programmed
>> in.)
>>
>> It is not just pattern matching. By 50 moves of Go with strong
>> probability the board is in a position that it has never simulated,
>> and games to completion go 200+ moves. Every move and stone on the
>> board is important. AlphaZero is learning high level concepts on its
>> own. It then uses those high level concepts to only consider
>> reasonable moves. (It does several orders of magnitude fewer
>> simulations at game time than other programs, at least at chess.)
>>
>> How is this not intelligence? What would AlphaZero have to do to show
>> intelligence in your opinion?
>>
> I would say that is highly advanced number crunching. It can run
> predictive simulations much farther into the "future" than past systems
> but that is just more advanced computation hardware. But the main
> problem with claiming successful "artificial intelligence" is you first
> have to _define_ "intelligence" in some kind of measurable, objective
> manner. As far as I know we can't.

Why? Weren't you around for that almost interminable discussion on
the definitions of "life"? "Life" is an high level concept that
humans have many, very different, definitions for that don't agree
with each other. They agree on determining whether something is alive
on the vast majority of cases, but no measurable, objective definition
exists. That's life.

"Intelligence" is yet another, even more difficult, high level concept
that there will never be a single agreed upon definition in English.
Just too high a level of concept and too many fuzzy English words
(other high level concepts) around. But that doesn't mean we can't
come up with useful definitions for our purposes, just as we do with
"life".

You need to be very careful when talking about measurable, objective
manner. That implies automated, implying a program, say Eval_I, could
do it. Would you be satisfied with an intelligence evaluation of a
program that examined possible solutions of whatever problem and
picked the one that scored highest on Eval_I?

This is not just a theoretical problem. For instance, the premier
Information Retrieval conference has a workshop this summer on
evaluating information retrieval systems using GPT as the decider
of "correct" answers. The problems and difficulties are obvious.

Chris

On 2024-03-30, Michael F. Stemper <michael.stemper@gmail.com> wrote:
> On 30/03/2024 14.33, Scott Lurndal wrote:
>> "Michael F. Stemper" <michael.stemper@gmail.com> writes:
>>> On 30/03/2024 06.12, D wrote:
>
>>>> Yes, if we're talking narrow, focused use cases such as chess and go, massive progress indeed. Depending on how you define it, you could add great progress when it comes to self driving cars, robots, image recognition and voice recognition too.
>>>
>>> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden Braid_,
>>> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>>> been done yet."
>>
>> But none of what is called "AI" today is really AI. It's just machine learning
>> and pattern matching with a massive database of patterns. Eliza writ large.
>
> Agreed. If it learns more than word patterns, I might be impressed. But
> right now, it's just textual manipulation. When the words start to have
> meaning to the program(s), then we'll really be approaching the Singularity.
>
> As far as I can tell, though, it's not aware that there's conflicting
> information out there. If you fed a bunch of Electric Universe text into
> one of these programs, it'd be just as happy parroting it back as it
> would GR. It wouldn't realize that the two "systems" are in conflict
> with each other.

Without the narrow world of AlphaZero and games, where what is happening
and not happening is much clearer, I'd have to argue from my experiences
with text and GPT. I think we all agree that that is much less convincing
and satisfactory. Nonetheless, I'll give a brief argument, though my
previous post on AlphaZero is the better argument.

I've been doing research in simple machine learning and pattern
matching of text for over 40 years. That's pretty much the definition
of statistical information retrieval. I know what the limits of
simple LLM's are. People have been trying to add AI and knowledge
structures on top of statistical ir for many decades, but unsuccessfully
(for general search.)

And the level of performance of simple LLM's and pattern matching is
pretty bad. It's far below what you've seen every day from Google Search
for the past 20 years. Google Search (at least in the 200X's) got its
increased performance by adding human input, mostly indirect. It did a very
good job at using human links between web pages and using human click-through
data (what pages humans actually click on after searching.)

This poor performance of simple LLMs and pattern matching continues to
this day. A couple of years ago I participated in a workshop
"comparison" designed to look at the problems of retrieval in a
rapidly changing domain (Covid research papers). Performance of
simple LLMs (including from researchers from places like Google) wasn't
significantly better than year 2000 systems.

So what has changed from simple LLMs to LLMs like GPT-4? Basically
deep learning: being able to make use of multiple layers of processing,
where high level concepts are defined and recognized and then operated
on by higher layers of processing (with feedback from those higher levels
used at the lower levels.) That's pretty much how I define "intelligence":
the ability to recognize and reason with high level (abstract) concepts.

Note that when you say "it's just machine learning and pattern recognition",
I don't really disagree. But that's because my definition of
"machine learning" includes "machine intelligence" as an almost
proper subset. GPT-4 (or AlphaZero) is learning new abstract concepts
and operating with them. What more do you require from machine
intelligence?

Chris

On 3/30/2024 11:11 PM, Dimensional Traveler wrote:
> On 3/30/2024 7:50 PM, Chris Buckley wrote:
>> On 2024-03-30, Michael F. Stemper <michael.stemper@gmail.com> wrote:
>>> On 30/03/2024 14.33, Scott Lurndal wrote:
>>>> "Michael F. Stemper" <michael.stemper@gmail.com> writes:
>>>>> On 30/03/2024 06.12, D wrote:
>>>
>>>>>> Yes, if we're talking narrow, focused use cases such as chess and
>>>>>> go, massive progress indeed. Depending on how you define it, you
>>>>>> could add great progress when it comes to self driving cars,
>>>>>> robots, image recognition and voice recognition too.
>>>>>
>>>>> If I recall correctly, in _Go:del, Escher, Bach: An Eternal Golden
>>>>> Braid_,
>>>>> Douglas Hofstader commented that "AI is sometimes just whatever hasn't
>>>>> been done yet."
>>>>
>>>> But none of what is called "AI" today is really AI.  It's just
>>>> machine learning
>>>> and pattern matching with a massive database of patterns.   Eliza
>>>> writ large.
>>>
>>> Agreed. If it learns more than word patterns, I might be impressed. But
>>> right now, it's just textual manipulation. When the words start to have
>>> meaning to the program(s), then we'll really be approaching the
>>> Singularity.
>>>
>>> As far as I can tell, though, it's not aware that there's conflicting
>>> information out there. If you fed a bunch of Electric Universe text into
>>> one of these programs, it'd be just as happy parroting it back as it
>>> would GR. It wouldn't realize that the two "systems" are in conflict
>>> with each other.
>>
>> All you folks are redefining AI once again, twisting it to mean
>> something unrecognizable (to me at least). It is not Artificial
>> Consciousness; it is not Artificial Embodiment. It is Artificial
>> *Intelligence*.
>>
>> Is a fish intelligent? Are some dogs more intelligent than others?
>>
>> I included my paragraph on AlphaZero for a reason.  AlphaZero is given
>> the rules of the game but no other game info. It took 34 hours of
>> playing itself at Go to become the best in the world. (The even more
>> impressive MuZero doesn't even have the rules of the game programmed
>> in.)
>>
>> It is not just pattern matching. By 50 moves of Go with strong
>> probability the board is in a position that it has never simulated,
>> and games to completion go 200+ moves. Every move and stone on the
>> board is important. AlphaZero is learning high level concepts on its
>> own.  It then uses those high level concepts to only consider
>> reasonable moves. (It does several orders of magnitude fewer
>> simulations at game time than other programs, at least at chess.)
>>
>> How is this not intelligence? What would AlphaZero have to do to show
>> intelligence in your opinion?
>>
> I would say that is highly advanced number crunching.  It can run
> predictive simulations much farther into the "future" than past systems
> but that is just more advanced computation hardware.  But the main
> problem with claiming successful "artificial intelligence" is you first
> have to _define_ "intelligence" in some kind of measurable, objective
> manner.  As far as I know we can't.

There's a saying, attributed to Lenin: "Quantity has a quality all its own."

I go for the duck test.

If it looks like a duck, quacks like a duck, and acts like a duck,
then its a duck.

pt

On 30 Mar 2024 21:40:36 -0000, kludge@panix.com (Scott Dorsey) wrote:

>Paul S Person <psperson@old.netcom.invalid> wrote:
>>That dovetails with something I've run into: when a bit of AI research
>>produces something /useful/ it ceases to be "AI" and is called
>>something else. "AI" is always used for pure research, unpolluted by
>>tawdry commercialization.
>
>This is a longstanding tradition. When I was in grad school, expert systems
>were AI, but by the time I was out, expert systems weren't AI anymore.
>(In the end expert systems didn't turn out to be useful either but that
>is a separate issue).
>
>But now we have come to a weird point in time when "AI" seems to be synonomous
>with "machine learning systems."

Well, we are living in an age of semantic goo ...
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"