In the discussion of AI generated sort algorithms, I gave the link to
the news release where the authors talked about applying their ideas to
hardware design. Now another group has done that.

https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/

I haven't read the linked to actual paper yet, but it appears we are on
the way toward AI assisted hardware design.

--
- Stephen Fuld
(e-mail address disguised to prevent spam)

Stephen Fuld <sfuld@alumni.cmu.edu.invalid> schrieb:
> In the discussion of AI generated sort algorithms, I gave the link to
> the news release where the authors talked about applying their ideas to
> hardware design. Now another group has done that.
>
> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/

ChatGPT? OMG.

When applied to anything that requires technical expertise, ChatGPT
is known for giving answers which are well written, sound plausible
and are wrong.

Remember that lawyer who let ChatGPT write a legal document
for him, which made up case law out of whole cloth?
https://www.bbc.com/news/world-us-canada-65735769

On Wednesday, June 21, 2023 at 10:18:56 AM UTC-5, Stephen Fuld wrote:
> In the discussion of AI generated sort algorithms, I gave the link to
> the news release where the authors talked about applying their ideas to
> hardware design. Now another group has done that.
>
> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/
>
> I haven't read the linked to actual paper yet, but it appears we are on
> the way toward AI assisted hardware design.
<
It occurs to me that this could actually benefit computer science.
<
Imaging if Hennessey an Patterson taught an AI to create computer
architectures and it came up with a design that was significantly
better, but looked nothing like any RISC architecture !?!
<
The benefits are 2 fold:
a) a better processor
b) the breaking of the "our RISC is the only way" mold.
>
>
> --
> - Stephen Fuld
> (e-mail address disguised to prevent spam)

On 6/21/2023 10:56 AM, Thomas Koenig wrote:
> Stephen Fuld <sfuld@alumni.cmu.edu.invalid> schrieb:
>> In the discussion of AI generated sort algorithms, I gave the link to
>> the news release where the authors talked about applying their ideas to
>> hardware design. Now another group has done that.
>>
>> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/
>
> ChatGPT? OMG.
>
> When applied to anything that requires technical expertise, ChatGPT
> is known for giving answers which are well written, sound plausible
> and are wrong.
>
> Remember that lawyer who let ChatGPT write a legal document
> for him, which made up case law out of whole cloth?
> https://www.bbc.com/news/world-us-canada-65735769

Yes, but . . .

Upon reading the actual paper, it is quite nuanced. First of all,
many/most of the errors are detected when the output Verilog is compiled
and simulated. This was an iterative process between Chat GPT and an
experienced designer, and the authors make it clear that is the current
state of the art. They propose this as a "coworker" who can offload
some of the tasks from the human designer.

ChatGPT is certainly not ready to replace human designers, but it can
make the human more productive. As such I regard it as a good starting
point, and look forward to further developments.

--
- Stephen Fuld
(e-mail address disguised to prevent spam)

MitchAlsup wrote:
> On Wednesday, June 21, 2023 at 10:18:56 AM UTC-5, Stephen Fuld wrote:
>> In the discussion of AI generated sort algorithms, I gave the link to
>> the news release where the authors talked about applying their ideas to
>> hardware design. Now another group has done that.
>>
>> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/
>>
>> I haven't read the linked to actual paper yet, but it appears we are on
>> the way toward AI assisted hardware design.
> <
> It occurs to me that this could actually benefit computer science.
> <
> Imaging if Hennessey an Patterson taught an AI to create computer
> architectures and it came up with a design that was significantly
> better, but looked nothing like any RISC architecture !?!
> <
> The benefits are 2 fold:
> a) a better processor
> b) the breaking of the "our RISC is the only way" mold.

What, you mean more non-riscy than my two destination registers?

By the way, since one can't copyright AI text output then
also one shouldn't be able to patent AI hardware designs.

Which, now that I think of it, the no-AI-copyright rule might invalidate
licenses for any software that incorporates AI generated software.

https://en.wikipedia.org/wiki/Artificial_intelligence_and_copyright

On 6/21/2023 6:51 PM, EricP wrote:
> MitchAlsup wrote:
>> On Wednesday, June 21, 2023 at 10:18:56 AM UTC-5, Stephen Fuld wrote:
>>> In the discussion of AI generated sort algorithms, I gave the link to
>>> the news release where the authors talked about applying their ideas
>>> to hardware design. Now another group has done that.
>>> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/
>>> I haven't read the linked to actual paper yet, but it appears we are
>>> on the way toward AI assisted hardware design.
>> <
>> It occurs to me that this could actually benefit computer science.
>> <
>> Imaging if Hennessey an Patterson taught an AI to create computer
>> architectures and it came up with a design that was significantly
>> better, but looked nothing like any RISC architecture !?!
>> <
>> The benefits are 2 fold:
>> a) a better processor
>> b) the breaking of the "our RISC is the only way" mold.
>
> What, you mean more non-riscy than my two destination registers?
>
> By the way, since one can't copyright AI text output then
> also one shouldn't be able to patent AI hardware designs.
>
> Which, now that I think of it, the no-AI-copyright rule might invalidate
> licenses for any software that incorporates AI generated software.
>
> https://en.wikipedia.org/wiki/Artificial_intelligence_and_copyright
>

As I see it, there are two major cases of AI that could be used here:
Generative AI, where the code is derived (indirectly) from the training
data and thus (inherently) derivative;
Genetic Algorithms, where what is generated is typically ex-nihilo.

I would think in the former, the copyright status is inherently murky
unless the original model was trained exclusively using data that is
freely usable within certain terms (say, training a generative AI
exclusively on public-domain works; as opposed to "strip mining the
internet" for content).

For most forms of media, this would effectively limit it to stuff from
before around 1920 or so. Could maybe stretch it to allow using stuff
under the MIT, BSD, and some of the Creative Commons licenses.

For Genetic Algorithms, presumably the copyright ownership would likely
fall to whoever uses the algorithm, as they are already going to need to
put in a lot of work in the process:
Fiddling with the design criteria;
Trying to coerce the output into something "actually usable";

...

Well, and a genetic algorithm isn't going to spit out big gobs of code
fully-formed, but would usually be applied to more specific and narrowly
defined tasks (such as tweaking a hash function or a branch predictor
design or similar).

In premise, one could potentially set an genetic algorithm loose on
tweaking or optimizing an instruction set; but this would be far more
involved. Most likely this would take the form of it being able to add
or remove possible instructions from the listing, with the compiler
being able to dynamically adapt to the addition or elimination of
instructions; with either a "mandatory core set", or a way to
"terminate" branches if it becomes no longer possible to compile programs.

In such a case, likely one wouldn't start with an entirely blank-slate,
but possibly a "large and orthogonal" space from which particular
instructions or encodings could be defined (and some way to model the
cost of the CPU that would result, ...).

But, then again, starting from a similar basis, one could parameterize
various parts of an ISA and already have a pretty good idea what it
would look like, without needing to get any "AI" involved.

Say, if we know things like:
Number and size of GPRs;
Whether other types of registers exist (and their sizes);
What types of encodings are available;
2R, 3R, Immed Sizes, ...
Whether the ISA is Load/Store or Reg/Mem;
Whether it has condition codes;
Whether it has predication;
Whether or not it is a VLIW;
...

Then, the actual instruction listing becomes almost secondary. One could
derive most of its behavioral properties from an abstract description.

Say, for example, say we describe an ISA like:
32 GPRs, 64-bit, with a Zero Register;
Has 3R, 3RI (Imm12/Disp12), 2RI Imm20/Disp20;
Is Load/Store;
Lacks condition codes;
Lacks predication;
Is not a VLIW;
Fixed 32-bit instruction words;
...

Then... Do we even really need to describe the listing?...
One can shuffle the listing around, or move the bit-fields around within
the instruction word, and it isn't going to make *that* much difference
to the end result.

One can further constrain that better options are to try to avoid "dog
chewing" the register and immediate fields, and then find that the
possibility space narrows considerably. One can likely already guess the
possible encoding schemes.

One could also add other constraints, say:
No "new" ISA features that don't have precedent in a mainstream ISA from
at least 25 years ago (so if any question ever comes up, one can be
like, "Hey, look here, prior art!");
....

....

Though, in related news, I am almost half tempted to consider reviving
the idea of a 2-wide version of BJX2 with a 4R2W register file for sake
of optimizing it to fit more easily into an XC7S50 than is possible with
the current 3-wide core.

It is possible that if I redesign a few parts some of the past
limitations could be reduced.

Say, allowing:
ALU | MEM

Though, in this case, both indexed-addressing ops and 128-bit Load/Store
would still need to eat both lanes.

But, not allowing "ALU|MEM" in the origin 2-wide WEX was a serious
hindrance to its usefulness (but was required given how the MEM ops were
mapped onto the register ports).

....

BGB wrote:
> On 6/21/2023 6:51 PM, EricP wrote:
>> MitchAlsup wrote:
>>> On Wednesday, June 21, 2023 at 10:18:56 AM UTC-5, Stephen Fuld wrote:
>>>> In the discussion of AI generated sort algorithms, I gave the link
>>>> to the news release where the authors talked about applying their
>>>> ideas to hardware design. Now another group has done that.
>>>> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/
>>>>
>>>> I haven't read the linked to actual paper yet, but it appears we are
>>>> on the way toward AI assisted hardware design.
>>> <
>>> It occurs to me that this could actually benefit computer science.
>>> <
>>> Imaging if Hennessey an Patterson taught an AI to create computer
>>> architectures and it came up with a design that was significantly
>>> better, but looked nothing like any RISC architecture !?!
>>> <
>>> The benefits are 2 fold:
>>> a) a better processor
>>> b) the breaking of the "our RISC is the only way" mold.
>>
>> What, you mean more non-riscy than my two destination registers?
>>
>> By the way, since one can't copyright AI text output then
>> also one shouldn't be able to patent AI hardware designs.
>>
>> Which, now that I think of it, the no-AI-copyright rule might invalidate
>> licenses for any software that incorporates AI generated software.
>>
>> https://en.wikipedia.org/wiki/Artificial_intelligence_and_copyright
>>
>
> As I see it, there are two major cases of AI that could be used here:
> Generative AI, where the code is derived (indirectly) from the training
> data and thus (inherently) derivative;
> Genetic Algorithms, where what is generated is typically ex-nihilo.
>
>
> I would think in the former, the copyright status is inherently murky
> unless the original model was trained exclusively using data that is
> freely usable within certain terms (say, training a generative AI
> exclusively on public-domain works; as opposed to "strip mining the
> internet" for content).
> ...

In the US at least the USPO has ruled that patents are reserved for
humans and that they do not apply to AI generated products.

Certain countries, US and Europe for example, have ruled that AI generated
literary works, which covers computer programs, are not copyrightable.
They do not distinguish between types of AI.
Presumably they see copyright as a reward for individual human creativity.

LLM generative AI is pretty much by definition derivative of others works.

For software that seems to leave one in the situation where you cannot
protect any work the AI produced, and at the same time are liable for
any potential infringement committed by the AI.

Furthermore I'm speculating that having AI generated software might
taint a company's whole development process if the provenance of any
work product now can be called into question. In order for a company to
prove the authorship of any work either they must keep complete detailed
audit trails of the development process, which is totally impractical,
or to firewall-off any AI development to protect themselves.
Or ban AI generated software from all parts of the development process.

Companies such as Microsoft that are rushing to jam AI into github and
into their development suites might find they have made those products
unusable due to the mere potential for taint.

On Friday, June 23, 2023 at 10:31:53 AM UTC-5, EricP wrote:
> BGB wrote:
> > On 6/21/2023 6:51 PM, EricP wrote:
> >> MitchAlsup wrote:
> >>> On Wednesday, June 21, 2023 at 10:18:56 AM UTC-5, Stephen Fuld wrote:
> >>>> In the discussion of AI generated sort algorithms, I gave the link
> >>>> to the news release where the authors talked about applying their
> >>>> ideas to hardware design. Now another group has done that.
> >>>> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/
> >>>>
> >>>> I haven't read the linked to actual paper yet, but it appears we are
> >>>> on the way toward AI assisted hardware design.
> >>> <
> >>> It occurs to me that this could actually benefit computer science.
> >>> <
> >>> Imaging if Hennessey an Patterson taught an AI to create computer
> >>> architectures and it came up with a design that was significantly
> >>> better, but looked nothing like any RISC architecture !?!
> >>> <
> >>> The benefits are 2 fold:
> >>> a) a better processor
> >>> b) the breaking of the "our RISC is the only way" mold.
> >>
> >> What, you mean more non-riscy than my two destination registers?
> >>
> >> By the way, since one can't copyright AI text output then
> >> also one shouldn't be able to patent AI hardware designs.
> >>
> >> Which, now that I think of it, the no-AI-copyright rule might invalidate
> >> licenses for any software that incorporates AI generated software.
> >>
> >> https://en.wikipedia.org/wiki/Artificial_intelligence_and_copyright
> >>
> >
> > As I see it, there are two major cases of AI that could be used here:
> > Generative AI, where the code is derived (indirectly) from the training
> > data and thus (inherently) derivative;
> > Genetic Algorithms, where what is generated is typically ex-nihilo.
> >
> >
> > I would think in the former, the copyright status is inherently murky
> > unless the original model was trained exclusively using data that is
> > freely usable within certain terms (say, training a generative AI
> > exclusively on public-domain works; as opposed to "strip mining the
> > internet" for content).
> > ...
>
> In the US at least the USPO has ruled that patents are reserved for
> humans and that they do not apply to AI generated products.
>
> Certain countries, US and Europe for example, have ruled that AI generated
> literary works, which covers computer programs, are not copyrightable.
> They do not distinguish between types of AI.
> Presumably they see copyright as a reward for individual human creativity..
>
> LLM generative AI is pretty much by definition derivative of others works..
>
> For software that seems to leave one in the situation where you cannot
> protect any work the AI produced, and at the same time are liable for
> any potential infringement committed by the AI.
<
I see hardware in the same boat--given your first sentence. None of the
protection, all of the exposure.
>
> Furthermore I'm speculating that having AI generated software might
> taint a company's whole development process if the provenance of any
> work product now can be called into question. In order for a company to
> prove the authorship of any work either they must keep complete detailed
> audit trails of the development process, which is totally impractical,
> or to firewall-off any AI development to protect themselves.
> Or ban AI generated software from all parts of the development process.
>
> Companies such as Microsoft that are rushing to jam AI into github and
> into their development suites might find they have made those products
> unusable due to the mere potential for taint.

On 6/23/2023 10:30 AM, EricP wrote:
> BGB wrote:
>> On 6/21/2023 6:51 PM, EricP wrote:
>>> MitchAlsup wrote:
>>>> On Wednesday, June 21, 2023 at 10:18:56 AM UTC-5, Stephen Fuld wrote:
>>>>> In the discussion of AI generated sort algorithms, I gave the link
>>>>> to the news release where the authors talked about applying their
>>>>> ideas to hardware design. Now another group has done that.
>>>>> https://www.allaboutcircuits.com/news/exclusive-nyu-team-uses-chatgpt-to-design-processor-from-scratch/
>>>>> I haven't read the linked to actual paper yet, but it appears we
>>>>> are on the way toward AI assisted hardware design.
>>>> <
>>>> It occurs to me that this could actually benefit computer science.
>>>> <
>>>> Imaging if Hennessey an Patterson taught an AI to create computer
>>>> architectures and it came up with a design that was significantly
>>>> better, but looked nothing like any RISC architecture !?!
>>>> <
>>>> The benefits are 2 fold:
>>>> a) a better processor
>>>> b) the breaking of the "our RISC is the only way" mold.
>>>
>>> What, you mean more non-riscy than my two destination registers?
>>>
>>> By the way, since one can't copyright AI text output then
>>> also one shouldn't be able to patent AI hardware designs.
>>>
>>> Which, now that I think of it, the no-AI-copyright rule might invalidate
>>> licenses for any software that incorporates AI generated software.
>>>
>>> https://en.wikipedia.org/wiki/Artificial_intelligence_and_copyright
>>>
>>
>> As I see it, there are two major cases of AI that could be used here:
>> Generative AI, where the code is derived (indirectly) from the
>> training data and thus (inherently) derivative;
>> Genetic Algorithms, where what is generated is typically ex-nihilo.
>>
>>
>> I would think in the former, the copyright status is inherently murky
>> unless the original model was trained exclusively using data that is
>> freely usable within certain terms (say, training a generative AI
>> exclusively on public-domain works; as opposed to "strip mining the
>> internet" for content).
>> ...
>
> In the US at least the USPO has ruled that patents are reserved for
> humans and that they do not apply to AI generated products.
>
> Certain countries, US and Europe for example, have ruled that AI generated
> literary works, which covers computer programs, are not copyrightable.
> They do not distinguish between types of AI.
> Presumably they see copyright as a reward for individual human creativity.
>

Possibly, but then one has to have some way to "prove" either that a
work was entirely human generated or (fully or partially) AI generated,
if it is the case that humans may be using AI tools to help with some of
the work.

> LLM generative AI is pretty much by definition derivative of others works.
>

Pretty much.

> For software that seems to leave one in the situation where you cannot
> protect any work the AI produced, and at the same time are liable for
> any potential infringement committed by the AI.
>

IMHO, LLM style AI should not be used for software development for this
reason.

> Furthermore I'm speculating that having AI generated software might
> taint a company's whole development process if the provenance of any
> work product now can be called into question. In order for a company to
> prove the authorship of any work either they must keep complete detailed
> audit trails of the development process, which is totally impractical,
> or to firewall-off any AI development to protect themselves.
> Or ban AI generated software from all parts of the development process.
>
> Companies such as Microsoft that are rushing to jam AI into github and
> into their development suites might find they have made those products
> unusable due to the mere potential for taint.
>
>

Possibly.

Though, I can also note that many of this "hype" is around the
(inherently derivative) LLM or "Deep Learning" style AIs.

I would personally be more accepting of the use of genetic algorithms.

But, then again, a genetic algorithm isn't exactly going to "write your
code for you" (IME, it is typically more time and effort to try to get a
genetic algorithm to generate something usable, than it would have been
to just come up with something oneself).

Well, and say, one has a random-number generator that uses a form like:
seed=seed*0xEAC53AC9+0x1F;
value=(uint32_t)(seed>>32);
And, does it really matter if the choice of magic numbers were come up
with using human reasoning, through the use of a genetic algorithm, or
by just pulling digits randomly from a hat?...

Just, in this sense, a generic algo is a much more elaborate way of
pulling digits from a hat...