Biden executive order imposes new rules for AI.

https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977

Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.

This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.

The EU is also considering legislation:
https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/

pt

"pete...@gmail.com" <petertrei@gmail.com> writes:
>Biden executive order imposes new rules for AI.
>
>https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>
>Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>
>This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.

In real life, yes.

In Science Fiction, _The Forbin Project_ and _The Two Faces of Tomorrow_ both
explore the potential downsides of AI.

On 11/1/2023 4:24 PM, Scott Lurndal wrote:
> "pete...@gmail.com" <petertrei@gmail.com> writes:
>> Biden executive order imposes new rules for AI.
>>
>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>
>> Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>>
>> This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>
> In real life, yes.
>
> In Science Fiction, _The Forbin Project_ and _The Two Faces of Tomorrow_ both
> explore the potential downsides of AI.

And the Terminator ?

Lynn

On Wednesday, November 1, 2023 at 8:50:07 PM UTC, pete...@gmail.com wrote:
> Biden executive order imposes new rules for AI.
>
> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>
> Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>
> This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>
> The EU is also considering legislation:
> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>
> pt
"Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.

I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.

On Thursday, 2 November 2023 at 05:23:47 UTC, Andrew McDowell wrote:
> On Wednesday, November 1, 2023 at 8:50:07 PM UTC, pete...@gmail.com wrote:
> > Biden executive order imposes new rules for AI.
> >
> > https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
> >
> > Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
> >
> > This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
> >
> > The EU is also considering legislation:
> > https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
> >
> > pt
> "Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.
>
> I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.

The issue isn't about AI being "smart" but about
how much decision-making it's doing on its own.

The matter of bias is of AI being trained to reproduce
decisions that conventionally involve human judgement,
by copying human results, and in practice, and with
AI technically not going wrong, if the human decisions
are prejudiced then the AI will reproduce the prejudices
in its own decisions. Supposing that we want to avoid
that, it takes work to avoid it. For human decision makers,
advice isn't to not have prejudices, that is probably
impossible and also is a form of discrimination itself
if prejudiced people aren't allowed to be decision makers.
Instead, good practice is to inspect a decision for prejudice
after you make the decision, and then correct the decision.
I went on a course: I know this.

On Thu, 2 Nov 2023 03:20:52 -0700 (PDT), Robert Carnegie
<rja.carnegie@excite.com> wrote:

>On Thursday, 2 November 2023 at 05:23:47 UTC, Andrew McDowell wrote:
>> On Wednesday, November 1, 2023 at 8:50:07?PM UTC, pete...@gmail.com wrote:
>> > Biden executive order imposes new rules for AI.
>> >
>> > https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>> >
>> > Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>> >
>> > This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>> >
>> > The EU is also considering legislation:
>> > https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>> >
>> > pt
>> "Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.
>>
>> I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.
>
>The issue isn't about AI being "smart" but about
>how much decision-making it's doing on its own.
>
>The matter of bias is of AI being trained to reproduce
>decisions that conventionally involve human judgement,
>by copying human results, and in practice, and with
>AI technically not going wrong, if the human decisions
>are prejudiced then the AI will reproduce the prejudices
>in its own decisions. Supposing that we want to avoid
>that, it takes work to avoid it. For human decision makers,
>advice isn't to not have prejudices, that is probably
>impossible and also is a form of discrimination itself
>if prejudiced people aren't allowed to be decision makers.
>Instead, good practice is to inspect a decision for prejudice
>after you make the decision, and then correct the decision.
>I went on a course: I know this.

Which is fine if you can do it, but it is often very hard to do.
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"

On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
<petertrei@gmail.com> wrote:

>Biden executive order imposes new rules for AI.
>
>https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>
>Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>
>This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>
>The EU is also considering legislation:
>https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/

Just tryin' to get /ahead/ of the curve. For once.

As noted by someone else, see the Terminator series (or read the
Forbin novels or even /Tron/) for what happens when you /don't/ have
adequate controls in place /before/ the AI tries to take over.
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"

On 11/1/23 18:23, Lynn McGuire wrote:
> On 11/1/2023 4:24 PM, Scott Lurndal wrote:
>> "pete...@gmail.com" <petertrei@gmail.com> writes:
>>> Biden executive order imposes new rules for AI.
>>>
>>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>>
>>> Essentially, anyone training a model with more than a certain
>>> complexity, needs to at register it with the government, and report
>>> safety testing results.
>>>
>>> This is the first case I can recall where governments have taken
>>> action out of fear that a computer system can be 'too smart'.
>>
>> In real life, yes.
>>
>> In Science Fiction, _The Forbin Project_ and _The Two Faces of
>> Tomorrow_ both
>> explore the potential downsides of AI.
>
> And the Terminator ?
>
> Lynn
>
And I Have No Mouth and I Must Scream
--
Dave Scruggs
Captain, Boulder Creek Fire (Retired)
Sr. Software Engineer (Retired, mostly)

On Thursday, November 2, 2023 at 12:12:06 PM UTC-4, Paul S Person wrote:
> On Thu, 2 Nov 2023 03:20:52 -0700 (PDT), Robert Carnegie
> <rja.ca...@excite.com> wrote:
>
> >On Thursday, 2 November 2023 at 05:23:47 UTC, Andrew McDowell wrote:
> >> On Wednesday, November 1, 2023 at 8:50:07?PM UTC, pete...@gmail.com wrote:
> >> > Biden executive order imposes new rules for AI.
> >> >
> >> > https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
> >> >
> >> > Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
> >> >
> >> > This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
> >> >
> >> > The EU is also considering legislation:
> >> > https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
> >> >
> >> > pt
> >> "Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.
> >>
> >> I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.
> >
> >The issue isn't about AI being "smart" but about
> >how much decision-making it's doing on its own.
> >
> >The matter of bias is of AI being trained to reproduce
> >decisions that conventionally involve human judgement,
> >by copying human results, and in practice, and with
> >AI technically not going wrong, if the human decisions
> >are prejudiced then the AI will reproduce the prejudices
> >in its own decisions. Supposing that we want to avoid
> >that, it takes work to avoid it. For human decision makers,
> >advice isn't to not have prejudices, that is probably
> >impossible and also is a form of discrimination itself
> >if prejudiced people aren't allowed to be decision makers.
> >Instead, good practice is to inspect a decision for prejudice
> >after you make the decision, and then correct the decision.
> >I went on a course: I know this.
> Which is fine if you can do it, but it is often very hard to do.

I like The Economist's cartoon take on the current EU AI Safety Summit:
https://www.economist.com/the-world-this-week/2023/11/02/kals-cartoon

pt

On Thursday, November 2, 2023 at 1:58:01 PM UTC-4, BCFD36 wrote:
> On 11/1/23 18:23, Lynn McGuire wrote:
> > On 11/1/2023 4:24 PM, Scott Lurndal wrote:
> >> "pete...@gmail.com" <pete...@gmail.com> writes:
> >>> Biden executive order imposes new rules for AI.
> >>>
> >>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
> >>>
> >>> Essentially, anyone training a model with more than a certain
> >>> complexity, needs to at register it with the government, and report
> >>> safety testing results.
> >>>
> >>> This is the first case I can recall where governments have taken
> >>> action out of fear that a computer system can be 'too smart'.
> >>
> >> In real life, yes.
> >>
> >> In Science Fiction, _The Forbin Project_ and _The Two Faces of
> >> Tomorrow_ both
> >> explore the potential downsides of AI.
> >
> > And the Terminator ?
> >
> > Lynn
> >
> And I Have No Mouth and I Must Scream

"An Intelligent Creation Thinks For Itself, And The Results Are Bad." is
probably one of the oldest tropes in SF. One could argue that the
Adam & Eve story fits, as well as Frankenstein and R.U.R.

pt

On 11/2/2023 3:46 PM, pete...@gmail.com wrote:
> On Thursday, November 2, 2023 at 1:58:01 PM UTC-4, BCFD36 wrote:
>> On 11/1/23 18:23, Lynn McGuire wrote:
>>> On 11/1/2023 4:24 PM, Scott Lurndal wrote:
>>>> "pete...@gmail.com" <pete...@gmail.com> writes:
>>>>> Biden executive order imposes new rules for AI.
>>>>>
>>>>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>>>>
>>>>> Essentially, anyone training a model with more than a certain
>>>>> complexity, needs to at register it with the government, and report
>>>>> safety testing results.
>>>>>
>>>>> This is the first case I can recall where governments have taken
>>>>> action out of fear that a computer system can be 'too smart'.
>>>>
>>>> In real life, yes.
>>>>
>>>> In Science Fiction, _The Forbin Project_ and _The Two Faces of
>>>> Tomorrow_ both
>>>> explore the potential downsides of AI.
>>>
>>> And the Terminator ?
>>>
>>> Lynn
>>>
>> And I Have No Mouth and I Must Scream
>
> "An Intelligent Creation Thinks For Itself, And The Results Are Bad." is
> probably one of the oldest tropes in SF. One could argue that the
> Adam & Eve story fits, as well as Frankenstein and R.U.R.
>
> pt

To God, we humans are self replicating automatons.

Lynn

On 11/2/2023 11:14 AM, Paul S Person wrote:
> On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
> <petertrei@gmail.com> wrote:
>
>> Biden executive order imposes new rules for AI.
>>
>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>
>> Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>>
>> This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>>
>> The EU is also considering legislation:
>> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>
> Just tryin' to get /ahead/ of the curve. For once.
>
> As noted by someone else, see the Terminator series (or read the
> Forbin novels or even /Tron/) for what happens when you /don't/ have
> adequate controls in place /before/ the AI tries to take over.

I still want to know how the various entities are going to control all
of the hobbyist programmers in their garages.

Lynn

In article <ui151s$2dbel$3@dont-email.me>,
Lynn McGuire <lynnmcguire5@gmail.com> wrote:
>On 11/2/2023 11:14 AM, Paul S Person wrote:
>> On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
>> <petertrei@gmail.com> wrote:
>>
>>> Biden executive order imposes new rules for AI.
>>>
>>>
>https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>>
>>> Essentially, anyone training a model with more than a certain
>complexity, needs to at register it with the government, and report
>safety testing results.
>>>
>>> This is the first case I can recall where governments have taken
>action out of fear that a computer system can be 'too smart'.
>>>
>>> The EU is also considering legislation:
>>> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>>
>> Just tryin' to get /ahead/ of the curve. For once.
>>
>> As noted by someone else, see the Terminator series (or read the
>> Forbin novels or even /Tron/) for what happens when you /don't/ have
>> adequate controls in place /before/ the AI tries to take over.
>
>I still want to know how the various entities are going to control all
>of the hobbyist programmers in their garages.
>
>Lynn
>

Shades of Asimov's "The Dead Past".
--
columbiaclosings.com
What's not in Columbia anymore..

On Thursday, November 2, 2023 at 5:31:32 PM UTC-4, Lynn McGuire wrote:
> On 11/2/2023 3:46 PM, pete...@gmail.com wrote:
> > On Thursday, November 2, 2023 at 1:58:01 PM UTC-4, BCFD36 wrote:
> >> On 11/1/23 18:23, Lynn McGuire wrote:
> >>> On 11/1/2023 4:24 PM, Scott Lurndal wrote:
> >>>> "pete...@gmail.com" <pete...@gmail.com> writes:
> >>>>> Biden executive order imposes new rules for AI.
> >>>>>
> >>>>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
> >>>>>
> >>>>> Essentially, anyone training a model with more than a certain
> >>>>> complexity, needs to at register it with the government, and report
> >>>>> safety testing results.
> >>>>>
> >>>>> This is the first case I can recall where governments have taken
> >>>>> action out of fear that a computer system can be 'too smart'.
> >>>>
> >>>> In real life, yes.
> >>>>
> >>>> In Science Fiction, _The Forbin Project_ and _The Two Faces of
> >>>> Tomorrow_ both
> >>>> explore the potential downsides of AI.
> >>>
> >>> And the Terminator ?
> >>>
> >>> Lynn
> >>>
> >> And I Have No Mouth and I Must Scream
> >
> > "An Intelligent Creation Thinks For Itself, And The Results Are Bad." is
> > probably one of the oldest tropes in SF. One could argue that the
> > Adam & Eve story fits, as well as Frankenstein and R.U.R.
> >
> > pt
> To God, we humans are self replicating automatons.

....and disobedient ones.

That was my point.

pt

On Thursday, November 2, 2023 at 5:41:26 PM UTC-4, Lynn McGuire wrote:
> On 11/2/2023 11:14 AM, Paul S Person wrote:
> > On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
> > <pete...@gmail.com> wrote:
> >
> >> Biden executive order imposes new rules for AI.
> >>
> >> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
> >>
> >> Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
> >>
> >> This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
> >>
> >> The EU is also considering legislation:
> >> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
> >
> > Just tryin' to get /ahead/ of the curve. For once.
> >
> > As noted by someone else, see the Terminator series (or read the
> > Forbin novels or even /Tron/) for what happens when you /don't/ have
> > adequate controls in place /before/ the AI tries to take over.
> I still want to know how the various entities are going to control all
> of the hobbyist programmers in their garages.

To a large extent, they can't.

This essay discusses the problem:

"We Have No Moat, And Neither Does OpenAI"
https://www.semianalysis.com/p/google-we-have-no-moat-and-neither

Its an internal Google document, in which the author discusses a technique
called 'Low Rank Adaption', which (to simplify) make training decomposable;
so small systems can train for a specific, 'small' task, but the training can then
be added to other LLMs, to combine their power.

Its something like in the Matrix, when Neo(?) says 'teach me to fly a helicopter', and
suddenly the skills, knowledge, and muscle memory are all loaded into his brain.

This means you don't *need* to have access to thousands of processor hours; things
can be built up gradually, by many small systems.

pt

On 11/2/2023 5:01 PM, pete...@gmail.com wrote:
> On Thursday, November 2, 2023 at 5:41:26 PM UTC-4, Lynn McGuire wrote:
>> On 11/2/2023 11:14 AM, Paul S Person wrote:
>>> On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
>>> <pete...@gmail.com> wrote:
>>>
>>>> Biden executive order imposes new rules for AI.
>>>>
>>>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>>>
>>>> Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>>>>
>>>> This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>>>>
>>>> The EU is also considering legislation:
>>>> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>>>
>>> Just tryin' to get /ahead/ of the curve. For once.
>>>
>>> As noted by someone else, see the Terminator series (or read the
>>> Forbin novels or even /Tron/) for what happens when you /don't/ have
>>> adequate controls in place /before/ the AI tries to take over.
>> I still want to know how the various entities are going to control all
>> of the hobbyist programmers in their garages.
>
> To a large extent, they can't.
>
> This essay discusses the problem:
>
> "We Have No Moat, And Neither Does OpenAI"
> https://www.semianalysis.com/p/google-we-have-no-moat-and-neither
>
> Its an internal Google document, in which the author discusses a technique
> called 'Low Rank Adaption', which (to simplify) make training decomposable;
> so small systems can train for a specific, 'small' task, but the training can then
> be added to other LLMs, to combine their power.
>
> Its something like in the Matrix, when Neo(?) says 'teach me to fly a helicopter', and
> suddenly the skills, knowledge, and muscle memory are all loaded into his brain.
>
> This means you don't *need* to have access to thousands of processor hours; things
> can be built up gradually, by many small systems.
>
> pt

They are already building bases for the various AI implementations and
passing them around. You can get various stages of AIs to customize now.

Lynn

On Thursday, 2 November 2023 at 16:12:06 UTC, Paul S Person wrote:
> On Thu, 2 Nov 2023 03:20:52 -0700 (PDT), Robert Carnegie
> <rja.ca...@excite.com> wrote:
>
> >On Thursday, 2 November 2023 at 05:23:47 UTC, Andrew McDowell wrote:
> >> On Wednesday, November 1, 2023 at 8:50:07?PM UTC, pete...@gmail.com wrote:
> >> > Biden executive order imposes new rules for AI.
> >> >
> >> > https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
> >> >
> >> > Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
> >> >
> >> > This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
> >> >
> >> > The EU is also considering legislation:
> >> > https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
> >> >
> >> > pt
> >> "Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.
> >>
> >> I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.
> >
> >The issue isn't about AI being "smart" but about
> >how much decision-making it's doing on its own.
> >
> >The matter of bias is of AI being trained to reproduce
> >decisions that conventionally involve human judgement,
> >by copying human results, and in practice, and with
> >AI technically not going wrong, if the human decisions
> >are prejudiced then the AI will reproduce the prejudices
> >in its own decisions. Supposing that we want to avoid
> >that, it takes work to avoid it. For human decision makers,
> >advice isn't to not have prejudices, that is probably
> >impossible and also is a form of discrimination itself
> >if prejudiced people aren't allowed to be decision makers.
> >Instead, good practice is to inspect a decision for prejudice
> >after you make the decision, and then correct the decision.
> >I went on a course: I know this.
>
> Which is fine if you can do it, but it is often very hard to do.

It's important. But I don't know if AI can do it, currently.

Lynn McGuire <lynnmcguire5@gmail.com> writes:
>On 11/2/2023 5:01 PM, pete...@gmail.com wrote:
>> On Thursday, November 2, 2023 at 5:41:26 PM UTC-4, Lynn McGuire wrote:
>>> On 11/2/2023 11:14 AM, Paul S Person wrote:
>>>> On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
>>>> <pete...@gmail.com> wrote:
>>>>
>>>>> Biden executive order imposes new rules for AI.
>>>>>
>>>>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>>>>
>>>>> Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>>>>>
>>>>> This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>>>>>
>>>>> The EU is also considering legislation:
>>>>> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>>>>
>>>> Just tryin' to get /ahead/ of the curve. For once.
>>>>
>>>> As noted by someone else, see the Terminator series (or read the
>>>> Forbin novels or even /Tron/) for what happens when you /don't/ have
>>>> adequate controls in place /before/ the AI tries to take over.
>>> I still want to know how the various entities are going to control all
>>> of the hobbyist programmers in their garages.
>>
>> To a large extent, they can't.
>>
>> This essay discusses the problem:
>>
>> "We Have No Moat, And Neither Does OpenAI"
>> https://www.semianalysis.com/p/google-we-have-no-moat-and-neither
>>
>> Its an internal Google document, in which the author discusses a technique
>> called 'Low Rank Adaption', which (to simplify) make training decomposable;
>> so small systems can train for a specific, 'small' task, but the training can then
>> be added to other LLMs, to combine their power.
>>
>> Its something like in the Matrix, when Neo(?) says 'teach me to fly a helicopter', and
>> suddenly the skills, knowledge, and muscle memory are all loaded into his brain.
>>
>> This means you don't *need* to have access to thousands of processor hours; things
>> can be built up gradually, by many small systems.
>>
>> pt
>
>They are already building bases for the various AI implementations and
>passing them around. You can get various stages of AIs to customize now.

It is useful to remember that what we have today is machine
learning. It's not artificial intelligence by any stretch
of the imagination.

The leap from the former to the latter hasn't yet happened
and probably won't absent self-awareness.

On Thu, 2 Nov 2023 15:26:51 -0700 (PDT), Robert Carnegie
<rja.carnegie@excite.com> wrote:

>On Thursday, 2 November 2023 at 16:12:06 UTC, Paul S Person wrote:
>> On Thu, 2 Nov 2023 03:20:52 -0700 (PDT), Robert Carnegie
>> <rja.ca...@excite.com> wrote:
>>
>> >On Thursday, 2 November 2023 at 05:23:47 UTC, Andrew McDowell wrote:
>> >> On Wednesday, November 1, 2023 at 8:50:07?PM UTC, pete...@gmail.com wrote:
>> >> > Biden executive order imposes new rules for AI.
>> >> >
>> >> > https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>> >> >
>> >> > Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>> >> >
>> >> > This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>> >> >
>> >> > The EU is also considering legislation:
>> >> > https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>> >> >
>> >> > pt
>> >> "Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.
>> >>
>> >> I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.
>> >
>> >The issue isn't about AI being "smart" but about
>> >how much decision-making it's doing on its own.
>> >
>> >The matter of bias is of AI being trained to reproduce
>> >decisions that conventionally involve human judgement,
>> >by copying human results, and in practice, and with
>> >AI technically not going wrong, if the human decisions
>> >are prejudiced then the AI will reproduce the prejudices
>> >in its own decisions. Supposing that we want to avoid
>> >that, it takes work to avoid it. For human decision makers,
>> >advice isn't to not have prejudices, that is probably
>> >impossible and also is a form of discrimination itself
>> >if prejudiced people aren't allowed to be decision makers.
>> >Instead, good practice is to inspect a decision for prejudice
>> >after you make the decision, and then correct the decision.
>> >I went on a course: I know this.
>>
>> Which is fine if you can do it, but it is often very hard to do.
>
>It's important. But I don't know if AI can do it, currently.

It's been a while, but an article on this in /Science News/ revealed
that, when an AI is clearly reproducing the racism of the system (the
AI in question was intended to make non-racist decisions on who to
grant bail to and who to hold on to), the only solution is to retrain
it with part of the original data missing and see if a less-biased AI
can be developed.

An AI trained on actual prior decisions that turns out to be racist,
BTW, is very strong evidence for structural racism.
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"

On Thu, 2 Nov 2023 16:41:16 -0500, Lynn McGuire
<lynnmcguire5@gmail.com> wrote:

>On 11/2/2023 11:14 AM, Paul S Person wrote:
>> On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
>> <petertrei@gmail.com> wrote:
>>
>>> Biden executive order imposes new rules for AI.
>>>
>>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>>
>>> Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>>>
>>> This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>>>
>>> The EU is also considering legislation:
>>> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>>
>> Just tryin' to get /ahead/ of the curve. For once.
>>
>> As noted by someone else, see the Terminator series (or read the
>> Forbin novels or even /Tron/) for what happens when you /don't/ have
>> adequate controls in place /before/ the AI tries to take over.
>
>I still want to know how the various entities are going to control all
>of the hobbyist programmers in their garages.

As always, the devil is in the details.
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"

On 11/2/2023 2:41 PM, Lynn McGuire wrote:
> On 11/2/2023 11:14 AM, Paul S Person wrote:
>> On Wed, 1 Nov 2023 13:50:04 -0700 (PDT), "pete...@gmail.com"
>> <petertrei@gmail.com> wrote:
>>
>>> Biden executive order imposes new rules for AI.
>>>
>>> https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>>>
>>> Essentially, anyone training a model with more than a certain
>>> complexity, needs to at register it with the government, and report
>>> safety testing results.
>>>
>>> This is the first case I can recall where governments have taken
>>> action out of fear that a computer system can be 'too smart'.
>>>
>>> The EU is also considering legislation:
>>> https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>>
>> Just tryin' to get /ahead/ of the curve. For once.
>>
>> As noted by someone else, see the Terminator series (or read the
>> Forbin novels or even /Tron/) for what happens when you /don't/ have
>> adequate controls in place /before/ the AI tries to take over.
>
> I still want to know how the various entities are going to control all
> of the hobbyist programmers in their garages.
>
Outlaw garages, obviously!

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

On Friday, November 3, 2023 at 4:08:01 PM UTC, Paul S Person wrote:
> On Thu, 2 Nov 2023 15:26:51 -0700 (PDT), Robert Carnegie
> <rja.ca...@excite.com> wrote:
>
> >On Thursday, 2 November 2023 at 16:12:06 UTC, Paul S Person wrote:
> >> On Thu, 2 Nov 2023 03:20:52 -0700 (PDT), Robert Carnegie
> >> <rja.ca...@excite.com> wrote:
> >>
> >> >On Thursday, 2 November 2023 at 05:23:47 UTC, Andrew McDowell wrote:
> >> >> On Wednesday, November 1, 2023 at 8:50:07?PM UTC, pete...@gmail.com wrote:
> >> >> > Biden executive order imposes new rules for AI.
> >> >> >
> >> >> > https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
> >> >> >
> >> >> > Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
> >> >> >
> >> >> > This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
> >> >> >
> >> >> > The EU is also considering legislation:
> >> >> > https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
> >> >> >
> >> >> > pt
> >> >> "Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.
> >> >>
> >> >> I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.
> >> >
> >> >The issue isn't about AI being "smart" but about
> >> >how much decision-making it's doing on its own.
> >> >
> >> >The matter of bias is of AI being trained to reproduce
> >> >decisions that conventionally involve human judgement,
> >> >by copying human results, and in practice, and with
> >> >AI technically not going wrong, if the human decisions
> >> >are prejudiced then the AI will reproduce the prejudices
> >> >in its own decisions. Supposing that we want to avoid
> >> >that, it takes work to avoid it. For human decision makers,
> >> >advice isn't to not have prejudices, that is probably
> >> >impossible and also is a form of discrimination itself
> >> >if prejudiced people aren't allowed to be decision makers.
> >> >Instead, good practice is to inspect a decision for prejudice
> >> >after you make the decision, and then correct the decision.
> >> >I went on a course: I know this.
> >>
> >> Which is fine if you can do it, but it is often very hard to do.
> >
> >It's important. But I don't know if AI can do it, currently.
> It's been a while, but an article on this in /Science News/ revealed
> that, when an AI is clearly reproducing the racism of the system (the
> AI in question was intended to make non-racist decisions on who to
> grant bail to and who to hold on to), the only solution is to retrain
> it with part of the original data missing and see if a less-biased AI
> can be developed.
>
> An AI trained on actual prior decisions that turns out to be racist,
> BTW, is very strong evidence for structural racism.
> --
> "Here lies the Tuscan poet Aretino,
> Who evil spoke of everyone but God,
> Giving as his excuse, 'I never knew him.'"
Some comments on this:

In the world we were originally sold where e.g. gender or colour of skin make no difference to any other characteristics, and it is sexist or racist to believe otherwise, fairness is easy; just ignore the protected characteristics. If you are arguing that this is unfair, you are arguing that the protected characteristic is in fact correlated with other characteristics, or there would be no way for the AI to distinguish between the groups in order to be unfair, even if that was its sole purpose.

Given the admission that there are correlations to the protected characteristics, there at least three definitions of fairness that people find natural. Unfortunately, it is mathematically impossible to satisfy all three definitions at the same time, unless there are no correlations with the protected characteristic (from https://courses.grainger.illinois.edu/ECE448/sp2023/slides/lec05.pdf) Your AI can only be fair (in all three ways) if society is already fair.

I too have gone on a course, in this case an online course on Machine Learning based on Microsoft's cloud offerings. If you pick your favoured definition of fairness, since the machinery in the background is just an optimisation routine, you can get it to optimise predictive accuracy subject to a fairness constraint perfectly well. The catch is that this will decrease the accuracy to an undetermined account (which you can estimate by comparing different solutions).

IMHO in almost all cases, getting the predictions right is good for society - it means the right people are chosen for the right job, the right treatment is given for the right disease, resources are used most efficiently (so less CO2 - e.g. burning fuel is expensive so reducing cost reduces CO2) and so on. Therefore I suggest that the colour-blind approach of ignoring the protected characteristics will often be the most efficient and therefore preferable answers, since this does not have to try and satisfy political constraints at the same time.

On Friday, 3 November 2023 at 17:29:24 UTC, Andrew McDowell wrote:
> In the world we were originally sold where e.g. gender or colour of skin make no difference to any other characteristics, and it is sexist or racist to believe otherwise, fairness is easy; just ignore the protected characteristics. If you are arguing that this is unfair, you are arguing that the protected characteristic is in fact correlated with other characteristics, or there would be no way for the AI to distinguish between the groups in order to be unfair, even if that was its sole purpose.
>
> Given the admission that there are correlations to the protected characteristics, there at least three definitions of fairness that people find natural. Unfortunately, it is mathematically impossible to satisfy all three definitions at the same time, unless there are no correlations with the protected characteristic (from https://courses.grainger.illinois.edu/ECE448/sp2023/slides/lec05.pdf) Your AI can only be fair (in all three ways) if society is already fair.
>
> I too have gone on a course, in this case an online course on Machine Learning based on Microsoft's cloud offerings. If you pick your favoured definition of fairness, since the machinery in the background is just an optimisation routine, you can get it to optimise predictive accuracy subject to a fairness constraint perfectly well. The catch is that this will decrease the accuracy to an undetermined account (which you can estimate by comparing different solutions).
>
> IMHO in almost all cases, getting the predictions right is good for society - it means the right people are chosen for the right job, the right treatment is given for the right disease, resources are used most efficiently (so less CO2 - e.g. burning fuel is expensive so reducing cost reduces CO2) and so on. Therefore I suggest that the colour-blind approach of ignoring the protected characteristics will often be the most efficient and therefore preferable answers, since this does not have to try and satisfy political constraints at the same time.

A couple of reactions:

"Indirect discrimination" is where bias isn't based
on a "protected characteristic" - one on which
discrimination is not allowed - directly, but which
has a similar effect. I suppose the easiest example
is of a policy that treats part time workers differently
and so it indirectly treats women differently. So a
process that doesn't know a person's sex can still
be unfair.

Also, imagine a process that predicts outcomes
by clairvoyance, such as in bail or hiring decisions.
The clairvoyance will see worse outcomes for women
and minorities when they are treated prejudicially by
prejudiced individuals, such as being re-arrested in
the bail case, or being under-appreciated in their work.
This is the history, and it probably is the future to an
extent. And so, more realistically, a process which is
given data from which protected characteristics can
be inferred, and which then is allowed to do what it
finds best with the data, may be treating people
unfairly who, if they were allowed to change their
sex or race or other factor of prejudice (and if they
wanted to change it), would have better outcomes.

So, not only does this decision process need to
be made fair, but the rest of the world does, too.

On 11/3/2023 1:37 PM, Robert Carnegie wrote:
....
> So, not only does this decision process need to
> be made fair, but the rest of the world does, too.

Good luck with that. First, who's version of fairness are you going to
use ?

Lynn

On Fri, 3 Nov 2023 10:29:21 -0700 (PDT), Andrew McDowell
<mcdowell_ag@sky.com> wrote:

>On Friday, November 3, 2023 at 4:08:01?PM UTC, Paul S Person wrote:
>> On Thu, 2 Nov 2023 15:26:51 -0700 (PDT), Robert Carnegie
>> <rja.ca...@excite.com> wrote:
>>
>> >On Thursday, 2 November 2023 at 16:12:06 UTC, Paul S Person wrote:
>> >> On Thu, 2 Nov 2023 03:20:52 -0700 (PDT), Robert Carnegie
>> >> <rja.ca...@excite.com> wrote:
>> >>
>> >> >On Thursday, 2 November 2023 at 05:23:47 UTC, Andrew McDowell wrote:
>> >> >> On Wednesday, November 1, 2023 at 8:50:07?PM UTC, pete...@gmail.com wrote:
>> >> >> > Biden executive order imposes new rules for AI.
>> >> >> >
>> >> >> > https://abcnews.go.com/Business/biden-executive-order-imposes-new-rules-ai/story?id=104472977
>> >> >> >
>> >> >> > Essentially, anyone training a model with more than a certain complexity, needs to at register it with the government, and report safety testing results.
>> >> >> >
>> >> >> > This is the first case I can recall where governments have taken action out of fear that a computer system can be 'too smart'.
>> >> >> >
>> >> >> > The EU is also considering legislation:
>> >> >> > https://www.weforum.org/agenda/2023/06/european-union-ai-act-explained/
>> >> >> >
>> >> >> > pt
>> >> >> "Too smart" - by design or by chinese whispers - is something of a caricature of the problem. Most of the claims of damage are also claims that the damage is because the AI is getting the wrong answers. If people look at how the AI is making decisions and find something that is not politically correct, they don't claim that the AI is too smart and has found them out; they claim that the AI is biased, incorrect, and must be fed adjusted data until it gets an answer that is politically correct.
>> >> >>
>> >> >> I note also that if you want a career in AI, the easiest entry point for many people at the moment is probably not developing or applying AI, but regulating AI. Hence we do not see garages filled by people building robots, but conferences and meetings about the regulation of AI.
>> >> >
>> >> >The issue isn't about AI being "smart" but about
>> >> >how much decision-making it's doing on its own.
>> >> >
>> >> >The matter of bias is of AI being trained to reproduce
>> >> >decisions that conventionally involve human judgement,
>> >> >by copying human results, and in practice, and with
>> >> >AI technically not going wrong, if the human decisions
>> >> >are prejudiced then the AI will reproduce the prejudices
>> >> >in its own decisions. Supposing that we want to avoid
>> >> >that, it takes work to avoid it. For human decision makers,
>> >> >advice isn't to not have prejudices, that is probably
>> >> >impossible and also is a form of discrimination itself
>> >> >if prejudiced people aren't allowed to be decision makers.
>> >> >Instead, good practice is to inspect a decision for prejudice
>> >> >after you make the decision, and then correct the decision.
>> >> >I went on a course: I know this.
>> >>
>> >> Which is fine if you can do it, but it is often very hard to do.
>> >
>> >It's important. But I don't know if AI can do it, currently.
>> It's been a while, but an article on this in /Science News/ revealed
>> that, when an AI is clearly reproducing the racism of the system (the
>> AI in question was intended to make non-racist decisions on who to
>> grant bail to and who to hold on to), the only solution is to retrain
>> it with part of the original data missing and see if a less-biased AI
>> can be developed.
>>
>> An AI trained on actual prior decisions that turns out to be racist,
>> BTW, is very strong evidence for structural racism.
>> --
>> "Here lies the Tuscan poet Aretino,
>> Who evil spoke of everyone but God,
>> Giving as his excuse, 'I never knew him.'"
>Some comments on this:
>
>In the world we were originally sold where e.g. gender or colour of skin make no difference to any other characteristics, and it is sexist or racist to believe otherwise, fairness is easy; just ignore the protected characteristics. If you are arguing that this is unfair, you are arguing that the protected characteristic is in fact correlated with other characteristics, or there would be no way for the AI to distinguish between the groups in order to be unfair, even if that was its sole purpose.
>
>Given the admission that there are correlations to the protected characteristics, there at least three definitions of fairness that people find natural. Unfortunately, it is mathematically impossible to satisfy all three definitions at the same time, unless there are no correlations with the protected characteristic (from https://courses.grainger.illinois.edu/ECE448/sp2023/slides/lec05.pdf) Your AI can only be fair (in all three ways) if society is already fair.
>
>I too have gone on a course, in this case an online course on Machine Learning based on Microsoft's cloud offerings. If you pick your favoured definition of fairness, since the machinery in the background is just an optimisation routine, you can get it to optimise predictive accuracy subject to a fairness constraint perfectly well. The catch is that this will decrease the accuracy to an undetermined account (which you can estimate by comparing different solutions).
>
>IMHO in almost all cases, getting the predictions right is good for society - it means the right people are chosen for the right job, the right treatment is given for the right disease, resources are used most efficiently (so less CO2 - e.g. burning fuel is expensive so reducing cost reduces CO2) and so on. Therefore I suggest that the colour-blind approach of ignoring the protected characteristics will often be the most efficient and therefore preferable answers, since this does not have to try and satisfy political constraints at the same time.

I am glad to hear that these issues are being addressed, as reflected
in the online course you took.
--
"Here lies the Tuscan poet Aretino,
Who evil spoke of everyone but God,
Giving as his excuse, 'I never knew him.'"