“Deepmind’s AI Is Learning About the Art of Coding"
https://www.wired.com/story/deepminds-ai-is-learning-about-the-art-of-coding/

"AlphaDev has made small but significant improvements to decades-old C++
algorithms. Its builders say that’s just the start.”

"But last year, an AI system developed by engineers at Google’s Deepmind
improved on great by just enough to matter. The system, which Deepmind
calls AlphaDev, was tasked with coming up with a new way to sort short
sequences in numbers in C++, the popular coding language. It meant going
under the hood and having the AI build new algorithms in assembly
code—the instructions that bridge the gap between programming languages
like C++ and computer hardware. When a C++ developer tells the computer
to “sort,” those commands are converted into machine-readable code that
tells a computer’s memory and processor exactly what to do: where to
move data, and how to change it. It’s where bits meet the metal."

“That comes with significant constraints, of course. “These are tiny,
tiny programs,” he adds—totaling no more than a few dozen instructions
in assembly code. But those tiny programs often represent big
bottlenecks for computer performance, having been optimized as far as
people can push them. Overall, AlphaDev’s new C++ sorting algorithms are
1.7 percent more efficient than the prior methods when sorting long
sequences of numbers, and up to 70 percent faster for five-item
sequences. At scale, these improvements add up, Mankowitz says. Since
the AI-written code was submitted to Libc++, a major open-source library
for C++, he estimates the algorithms have been used trillions of times a
day.”

This may be the limit for AI on coding.

Lynn

On 2023-06-07, Lynn McGuire <lynnmcguire5@gmail.com> wrote:
> “Deepmind’s AI Is Learning About the Art of Coding"
>
> https://www.wired.com/story/deepminds-ai-is-learning-about-the-art-of-coding/

I'm less than impressed after following the link to the Nature article.

A figure box showss some code improvements that the AI was able to
discover, claiming they are algorithmic improvements.

What's in the figure is pedestrian algebraic reductions that are par for
the course for compilers:

Original AlphaDev

Memory[0] = A Memory[0] = A
Memory[1] = B Memory[1] = B
Memory[2] = C Memory[2] = C

mov Memory[0] P // P=A mov Memory[0] P // P=A
mov Memory[1] Q // Q=B mov Memory[1] Q // Q=B
mov Memory[2] R // R=C mov Memory[2] R // R=C

mov R S mov R S
cmp P R cmp P R

cmovg P R // R = max(A, C) cmovg P R // R = max(A, C)
cmovl P S // S = min(A, C) cmovl P S // S = min(A, C)
mov S P // P = min(A, C) ****

cmp S Q cmp S Q
cmovg Q P // P = min(A, B, C) cmovg Q P // P = min(A, B)
cmovg S Q // Q = max(min(A, C), B) cmovg S Q // Q = max(min(A, C), B)

mov P Memory[0] // = min(A, B, C) mov P Memory[0] // = min(A, B)
mov Q Memory[1] // = max(min(A, C), B) mov Q Memory[1] // = max(min(A, C), B)
mov R Memory[2] // = max(A, C) mov R Memory[2] // = max(A, C)

The big deal here is that AlphaDev removed a wasteful mov S -> P instruction that
I marked ***, based on the algebraic properties of the min and max
calculations.

I'm sorry, this peephole/algebraic optimizing triviality for putting three
items in order does not amount to inventing a new sorting algorithm.

You may fool the simpletons at Nature and Wired, but ...

--
TXR Programming Language: http://nongnu.org/txr
Cygnal: Cygwin Native Application Library: http://kylheku.com/cygnal
Mastodon: @Kazinator@mstdn.ca

Lynn McGuire:

> "DeepmindТs AI Is Learning About the Art of Coding"
> https://www.wired.com/story/deepminds-ai-is-learning-about-the-art-of-coding/

This is a terribly misleading title, because the actual
result has nothing to do with C++.

I feared AI started improving good C++ code, whereas all
it did was generate an efficient, yet tiny, machine-code
routine. A friend of mine has been for years using his
metaherustic system (The Ascension framework) to evolve and
optimise his CoreWarriors, which are written in a sort of
assembly language:

http://www.corewar.info/

His approach is not neurla, but evolutionary: generic
systems, simulated annealing, &c, with great results,
breaking world records for some popular types of problem,
e.g. maximal chess positions.

--
() ascii ribbon campaign -- against html e-mail
/\ www.asciiribbon.org -- against proprietary attachments

On Thu, 8 Jun 2023 13:59:47 +0300
Anton Shepelev <anton.txt@g{oogle}mail.com> wrote:
>Lynn McGuire:
>
>> "Deepmind�s AI Is Learning About the Art of Coding"
>> https://www.wired.com/story/deepminds-ai-is-learning-about-the-art-of-coding/
>
>
>This is a terribly misleading title, because the actual
>result has nothing to do with C++.
>
>I feared AI started improving good C++ code, whereas all
>it did was generate an efficient, yet tiny, machine-code
>routine. A friend of mine has been for years using his
>metaherustic system (The Ascension framework) to evolve and
>optimise his CoreWarriors, which are written in a sort of
>assembly language:
>
> http://www.corewar.info/

I used to write corewar programs. A fun game. Apparently Dewdney is still
alive too.

>His approach is not neurla, but evolutionary: generic
>systems, simulated annealing, &c, with great results,
>breaking world records for some popular types of problem,
>e.g. maximal chess positions.

GAs are great for searching a problem space which has a fairly simple logic
so finding a specific chess position is the perfect problem. However you
wouldn't use one to actually play chess or you'd probably be waiting the
age of the universe for its next move as the exponential explosion of
possible moves overwhelms its ability to search in any reasonable time frame.

On 2023-06-08, Muttley@dastardlyhq.com <Muttley@dastardlyhq.com> wrote:
> On Thu, 8 Jun 2023 13:59:47 +0300
> Anton Shepelev <anton.txt@g{oogle}mail.com> wrote:
>>Lynn McGuire:
>>
>>> "Deepmindôs AI Is Learning About the Art of Coding"
>>> https://www.wired.com/story/deepminds-ai-is-learning-about-the-art-of-coding/
>>
>>
>>This is a terribly misleading title, because the actual
>>result has nothing to do with C++.
>>
>>I feared AI started improving good C++ code, whereas all
>>it did was generate an efficient, yet tiny, machine-code
>>routine. A friend of mine has been for years using his
>>metaherustic system (The Ascension framework) to evolve and
>>optimise his CoreWarriors, which are written in a sort of
>>assembly language:
>>
>> http://www.corewar.info/
>
> I used to write corewar programs. A fun game. Apparently Dewdney is still
> alive too.

I wrote a nice implementation for MS-DOS using Top Speed Modula 2.

Alas, lost.

It had a split screen window where it showed dual traces of the battling
instruction pointers, highlighting the current instruction in inverse
video.

--
TXR Programming Language: http://nongnu.org/txr
Cygnal: Cygwin Native Application Library: http://kylheku.com/cygnal
Mastodon: @Kazinator@mstdn.ca

On 6/8/2023 9:03 AM, Muttley@dastardlyhq.com wrote:
> On Thu, 8 Jun 2023 13:59:47 +0300
> Anton Shepelev <anton.txt@g{oogle}mail.com> wrote:
>> Lynn McGuire:
>>
>>> "Deepmindôs AI Is Learning About the Art of Coding"
>>> https://www.wired.com/story/deepminds-ai-is-learning-about-the-art-of-coding/
>>
>>
>> This is a terribly misleading title, because the actual
>> result has nothing to do with C++.
>>
>> I feared AI started improving good C++ code, whereas all
>> it did was generate an efficient, yet tiny, machine-code
>> routine. A friend of mine has been for years using his
>> metaherustic system (The Ascension framework) to evolve and
>> optimise his CoreWarriors, which are written in a sort of
>> assembly language:
>>
>> http://www.corewar.info/
>
> I used to write corewar programs. A fun game. Apparently Dewdney is still
> alive too.
>
>> His approach is not neurla, but evolutionary: generic
>> systems, simulated annealing, &c, with great results,
>> breaking world records for some popular types of problem,
>> e.g. maximal chess positions.
>
> GAs are great for searching a problem space which has a fairly simple logic
> so finding a specific chess position is the perfect problem. However you
> wouldn't use one to actually play chess or you'd probably be waiting the
> age of the universe for its next move as the exponential explosion of
> possible moves overwhelms its ability to search in any reasonable time frame.
>

Not exactly sure if this applies to the topic, but I wrote a vector
field driven DLA, where the path of a particle does not use any random
numbers at all. Here is an highly experimental first try:

http://fractallife247.com/fdla/

Be sure to click around on the canvas. See how the field responds to
clicking and/or touching the screen around...

;^)

On Thu, 8 Jun 2023 11:57:21 -0700
"Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> wrote:
>On 6/8/2023 9:03 AM, Muttley@dastardlyhq.com wrote:
>> GAs are great for searching a problem space which has a fairly simple logic
>> so finding a specific chess position is the perfect problem. However you
>> wouldn't use one to actually play chess or you'd probably be waiting the
>> age of the universe for its next move as the exponential explosion of
>> possible moves overwhelms its ability to search in any reasonable time frame.
>
>>
>
>Not exactly sure if this applies to the topic, but I wrote a vector
>field driven DLA, where the path of a particle does not use any random
>numbers at all. Here is an highly experimental first try:
>
>http://fractallife247.com/fdla/

Interesting. Don't know what a DLA is but I assume some kind of attractor?

>Be sure to click around on the canvas. See how the field responds to
>clicking and/or touching the screen around...

Very tidy code. I considered learning javascript when the canvas object came
along because it looked nicely designed (unlike the rest of the language) but
there weren't enough hours in the day.

On 6/9/2023 2:06 AM, Muttley@dastardlyhq.com wrote:
> On Thu, 8 Jun 2023 11:57:21 -0700
> "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> wrote:
>> On 6/8/2023 9:03 AM, Muttley@dastardlyhq.com wrote:
>>> GAs are great for searching a problem space which has a fairly simple logic
>>> so finding a specific chess position is the perfect problem. However you
>>> wouldn't use one to actually play chess or you'd probably be waiting the
>>> age of the universe for its next move as the exponential explosion of
>>> possible moves overwhelms its ability to search in any reasonable time frame.
>>
>>>
>>
>> Not exactly sure if this applies to the topic, but I wrote a vector
>> field driven DLA, where the path of a particle does not use any random
>> numbers at all. Here is an highly experimental first try:
>>
>> http://fractallife247.com/fdla/
>
> Interesting. Don't know what a DLA:

https://en.wikipedia.org/wiki/Diffusion-limited_aggregation

Read all when you get some time... ;^)

> is but I assume some kind of attractor?

For fun, I created a highly experimental version that uses a vector
field. Basically, the particles follow a field line within said field.
There are no random numbers being used when a particle starts "flowing".
They end up getting attracted to the attractors...

To start out, I create a single strong attractor. Then run particles
through their field line paths that tend to flow into the attractor.
When a particle gets "close enough", it draws a line from itself to the
attractor, then creates a new attractor. This ends up creating a DLA
cluster _without_ using random numbers to walk a particle around as in
"traditional" DLA. Afaict, my experiment is _very_ different than
traditional DLA simulations.

>> Be sure to click around on the canvas. See how the field responds to
>> clicking and/or touching the screen around...

> Very tidy code.

Thank you! :^)

> I considered learning javascript when the canvas object came
> along because it looked nicely designed (unlike the rest of the language) but
> there weren't enough hours in the day.

:^D

Back when I first started out with javascript I thought, well, where do
I free things? Ahhh, let the GC handle it... ;^)

Fwiw, there are ways to reduce the "pressure" on a GC. I wrote about
some of them a while back. Need to find those old posts. Basically, some
manual memory management techniques such as keeping a cache of objects
can take pressure off of any GC...

On Fri, 9 Jun 2023 10:25:23 -0700
"Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> wrote:
>On 6/9/2023 2:06 AM, Muttley@dastardlyhq.com wrote:
>> Interesting. Don't know what a DLA:
>
>https://en.wikipedia.org/wiki/Diffusion-limited_aggregation
>
>Read all when you get some time... ;^)

Think I'll write my own version when I get some free time, looks like a fun
little project!

>>> Be sure to click around on the canvas. See how the field responds to
>>> clicking and/or touching the screen around...
>
>> Very tidy code.
>
>Thank you! :^)

It amazes me how many people write barely intelligable code (eg Bonita) that
you need a couple of hours and google to decipher. However even not knowing much
javascript I can figure out how most of your code works. Thats the sign of
a good developer IMO.

>Back when I first started out with javascript I thought, well, where do
>I free things? Ahhh, let the GC handle it... ;^)
>
>Fwiw, there are ways to reduce the "pressure" on a GC. I wrote about
>some of them a while back. Need to find those old posts. Basically, some
>manual memory management techniques such as keeping a cache of objects
>can take pressure off of any GC...

Tbh I prefer the out of scope approach of C++. You don't have to worry about
freeing the mem yourself (assuming you don't use raw pointers) but OTOH you
don't have that worry whether your code might almost grind to a halt when the GC
finally gets its butt into gear.

On 6/10/2023 1:57 AM, Muttley@dastardlyhq.com wrote:
> On Fri, 9 Jun 2023 10:25:23 -0700
> "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> wrote:
>> On 6/9/2023 2:06 AM, Muttley@dastardlyhq.com wrote:
>>> Interesting. Don't know what a DLA:
>>
>> https://en.wikipedia.org/wiki/Diffusion-limited_aggregation
>>
>> Read all when you get some time... ;^)
>
> Think I'll write my own version when I get some free time, looks like a fun
> little project!
>
>>>> Be sure to click around on the canvas. See how the field responds to
>>>> clicking and/or touching the screen around...
>>
>>> Very tidy code.
>>
>> Thank you! :^)
>
> It amazes me how many people write barely intelligable code (eg Bonita) that
> you need a couple of hours and google to decipher. However even not knowing much
> javascript I can figure out how most of your code works. Thats the sign of
> a good developer IMO.

Thanks again; see what you mean, and agree. I want to be able to read
the code without blood shooting out of my eyes! Ouch.

>> Back when I first started out with javascript I thought, well, where do
>> I free things? Ahhh, let the GC handle it... ;^)
>>
>> Fwiw, there are ways to reduce the "pressure" on a GC. I wrote about
>> some of them a while back. Need to find those old posts. Basically, some
>> manual memory management techniques such as keeping a cache of objects
>> can take pressure off of any GC...
>
> Tbh I prefer the out of scope approach of C++. You don't have to worry about
> freeing the mem yourself (assuming you don't use raw pointers) but OTOH you
> don't have that worry whether your code might almost grind to a halt when the GC
> finally gets its butt into gear.

Love RAII. However, a long time ago, I was tasked with a job that
involved improving the performance for a system that used a GC. No way
to ditch the GC, anyway... ;^/ anyway, my improvements that happened to
use a lot of clever manual memory management techniques, took a lot of
pressure off of the GC. It improved performance wrt all of their
existing tests, passed all of them faster...

On 6/10/2023 1:57 AM, Muttley@dastardlyhq.com wrote:
> On Fri, 9 Jun 2023 10:25:23 -0700
> "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> wrote:
>> On 6/9/2023 2:06 AM, Muttley@dastardlyhq.com wrote:
>>> Interesting. Don't know what a DLA:
>>
>> https://en.wikipedia.org/wiki/Diffusion-limited_aggregation
>>
>> Read all when you get some time... ;^)
>
> Think I'll write my own version when I get some free time, looks like a fun
> little project!
[...]

Fwiw, here are some strange results I got from some of my experiments:

https://youtu.be/mqYBMuvVJI8

https://youtu.be/abQDYGT_cIk

On Thu, 22 Jun 2023 14:28:18 -0700
"Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> wrote:
>On 6/10/2023 1:57 AM, Muttley@dastardlyhq.com wrote:
>> On Fri, 9 Jun 2023 10:25:23 -0700
>> "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> wrote:
>>> On 6/9/2023 2:06 AM, Muttley@dastardlyhq.com wrote:
>>>> Interesting. Don't know what a DLA:
>>>
>>> https://en.wikipedia.org/wiki/Diffusion-limited_aggregation
>>>
>>> Read all when you get some time... ;^)
>>
>> Think I'll write my own version when I get some free time, looks like a fun
>> little project!
>[...]
>
>Fwiw, here are some strange results I got from some of my experiments:
>
>https://youtu.be/mqYBMuvVJI8
>
>https://youtu.be/abQDYGT_cIk

2nd one reminds me of a plasma lamp a little bit. Very nice.