From Newsgroup: comp.lang.prolog

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 | | x1 | | x4 |
| 0 0 1 0 | | x2 | = | x3 |
| 0 1 0 0 | | x3 | | x2 |
| 1 0 0 0 | | x4 | | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.
--- Synchronet 3.21f-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

You just escaped AI dooms day. Humanity has
reset all internet and computers as a last resort
to prevent AGI developing, by an electromagnetic

pulse. You are stuck in G|+ttinger Wald and hunted
down a deer by your bare hands, the deer still
confused and tame because tourists were feeding it.

Now you have no knife, what do you do:

Chimpanzees Have Entered The Stone Age https://www.youtube.com/watch?v=wPXX2I_uYjc

So we are just apes with internet.

Bye

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 |-a-a | x1 |-a-a-a-a | x4 |
| 0 0 1 0 |-a-a | x2 |-a =-a | x3 |
| 0 1 0 0 |-a-a | x3 |-a-a-a-a | x2 |
| 1 0 0 0 |-a-a | x4 |-a-a-a-a | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.21f-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

His last paper , that I am refering to, his
last opus , could have some merrits nevertheless.
Like if you are interested in intuitionistic logic,

and ever wondered what makes (a1 -> (a2 -> .. (an -> b)
..)) different from ((.. (b -> a1) .. -> an-1) -> an).
Kind of loophole to encode lists? But transformer

doesn't work with lists, its not that a transformer
would directly learn:

reverse(X, Y) :- reverse(X, [], Y).

reverse([], X, X).
reverse([X|Y], Z, T) :- reverse(Y, [X|Z], T).

Because transformers don't use LISP consing
representation. Neither for the input token stream
nor for the output token stream,

that ChatGPT shows tokens gradually appearing in
his UX is only a Gimmick.

Bye

Mild Shock schrieb:

Hi,

You just escaped AI dooms day. Humanity has
reset all internet and computers as a last resort
to prevent AGI developing, by an electromagnetic

pulse. You are stuck in G|+ttinger Wald and hunted
down a deer by your bare hands, the deer still
confused and tame because tourists were feeding it.

Now you have no knife, what do you do:

Chimpanzees Have Entered The Stone Age https://www.youtube.com/watch?v=wPXX2I_uYjc

So we are just apes with internet.

Bye

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 |-a-a | x1 |-a-a-a-a | x4 |
| 0 0 1 0 |-a-a | x2 |-a =-a | x3 |
| 0 1 0 0 |-a-a | x3 |-a-a-a-a | x2 |
| 1 0 0 0 |-a-a | x4 |-a-a-a-a | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.21f-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

How it started:

Epstein Island LEGO Set
Build the legendary island, assemble the pool, and get
the party started with all your kids. This set captures
the intensity and atmosphere of the island. https://www.youtube.com/watch?v=ttwO_eeqkWc

How its going:

The EU approach to age verification
The European Commission's approach to age verification
focuses on developing user-friendly, privacy-preserving
solutions alongside Member States. https://digital-strategy.ec.europa.eu/en/policies/eu-age-verification

Bye

Mild Shock schrieb:

Hi,

You just escaped AI dooms day. Humanity has
reset all internet and computers as a last resort
to prevent AGI developing, by an electromagnetic

pulse. You are stuck in G|+ttinger Wald and hunted
down a deer by your bare hands, the deer still
confused and tame because tourists were feeding it.

Now you have no knife, what do you do:

Chimpanzees Have Entered The Stone Age https://www.youtube.com/watch?v=wPXX2I_uYjc

So we are just apes with internet.

Bye

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 |-a-a | x1 |-a-a-a-a | x4 |
| 0 0 1 0 |-a-a | x2 |-a =-a | x3 |
| 0 1 0 0 |-a-a | x3 |-a-a-a-a | x2 |
| 1 0 0 0 |-a-a | x4 |-a-a-a-a | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.21f-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Lets get emotional! While Varoufakis painted
the picture of cloud capital. That might have
mobilized "The Internationale", or another

more defensive less motolotov throwing song:

Pink Floyd - Run Like Hell (Live)
https://www.youtube.com/watch?v=lKgOe1Rl8YY

Now since Athropic is teaming with xAI, we
might ask do we see the next OneDrive of Prolog
on the horizon. Even a tame Erlang dream:

populate the Web with clever Prolog agents!
https://trinity.elfenbenstornet.se/

Might have a nasty Prolog as SaaS aspect!
As long as we talk about services and not
assets, we might miss something. Who owns

the present and future LLMs/LRMs?

Bye

Mild Shock schrieb:

Hi,

You just escaped AI dooms day. Humanity has
reset all internet and computers as a last resort
to prevent AGI developing, by an electromagnetic

pulse. You are stuck in G|+ttinger Wald and hunted
down a deer by your bare hands, the deer still
confused and tame because tourists were feeding it.

Now you have no knife, what do you do:

Chimpanzees Have Entered The Stone Age https://www.youtube.com/watch?v=wPXX2I_uYjc

So we are just apes with internet.

Bye

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 |-a-a | x1 |-a-a-a-a | x4 |
| 0 0 1 0 |-a-a | x2 |-a =-a | x3 |
| 0 1 0 0 |-a-a | x3 |-a-a-a-a | x2 |
| 1 0 0 0 |-a-a | x4 |-a-a-a-a | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Even the Buddos are cluless, while Tarau might
indeed appear in the anals of the Borg, as a
notable human being, seeing connections.

While the Buddos are the man mountains of
Janathan Swists Gulliver's Travel, creating
huge egg montains, replaying some rewriting

school inventions. They might nevertheless be
strapped down by Liliputians:

Gulliver captzured by the Liliputians https://www.lookandlearn.com/history-images/M301092/Scene-from-Gullivers-Travels

But who are these Liliputians? Well just
toying around with a deep seek v4 derivate in
LM Studio, a model that came out 9 days ago.

Etc.. etc.. it shows more text, all generated
on a laptop that was even only $1000 since
end of year 2025, there were some discounts.

The laptop has the Windows Copilot+ specs.
The secrete sauce? Some general matrix
multiplications (GEMM) tucked in your iGPU:

What is Xe Matrix eXtensions (XMX)? https://www.intel.com/content/www/us/en/support/articles/000091112/graphics.html

Bye


Mild Shock schrieb:

Hi,

Lets get emotional! While Varoufakis painted
the picture of cloud capital. That might have
mobilized "The Internationale", or another

more defensive less motolotov throwing song:

Pink Floyd - Run Like Hell (Live)
https://www.youtube.com/watch?v=lKgOe1Rl8YY

Now since Athropic is teaming with xAI, we
might ask do we see the next OneDrive of Prolog
on the horizon. Even a tame Erlang dream:

populate the Web with clever Prolog agents! https://trinity.elfenbenstornet.se/

Might have a nasty Prolog as SaaS aspect!
As long as we talk about services and not
assets, we might miss something. Who owns

the present and future LLMs/LRMs?

Bye

Mild Shock schrieb:

Hi,

You just escaped AI dooms day. Humanity has
reset all internet and computers as a last resort
to prevent AGI developing, by an electromagnetic

pulse. You are stuck in G|+ttinger Wald and hunted
down a deer by your bare hands, the deer still
confused and tame because tourists were feeding it.

Now you have no knife, what do you do:

Chimpanzees Have Entered The Stone Age
https://www.youtube.com/watch?v=wPXX2I_uYjc

So we are just apes with internet.

Bye

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 |-a-a | x1 |-a-a-a-a | x4 |
| 0 0 1 0 |-a-a | x2 |-a =-a | x3 |
| 0 1 0 0 |-a-a | x3 |-a-a-a-a | x2 |
| 1 0 0 0 |-a-a | x4 |-a-a-a-a | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Interesting, the Copilot+ minimal
requirement is 40 TOPS. Now Wiki is
mumbling something of 35 TOPS + 15%,

for the new A18 chip, that is found
in iPhones and the new Mac Neo. The
new Mac Neo is only $500 , half of

my discount AI laptop, and is selling
like hotcakes. I should try it, see
what AI workloads it can do locally.

Bye

P.S.: The Mac Neo is jokingly called
Ne-Meow in this hands on video by Bijan
Bowen, showing some vibe web coding:

MacBook Neo Local AI Test
https://www.youtube.com/watch?v=75PFpW9SOL0

Mild Shock schrieb:

Hi,

Even the Buddos are cluless, while Tarau might
indeed appear in the anals of the Borg, as a
notable human being, seeing connections.

While the Buddos are the man mountains of
Janathan Swists Gulliver's Travel, creating
huge egg montains, replaying some rewriting

school inventions. They might nevertheless be
strapped down by Liliputians:

Gulliver captzured by the Liliputians https://www.lookandlearn.com/history-images/M301092/Scene-from-Gullivers-Travels

But who are these Liliputians? Well just
toying around with a deep seek v4 derivate in
LM Studio, a model that came out 9 days ago.

Etc.. etc.. it shows more text, all generated
on a laptop that was even only $1000 since
end of year 2025, there were some discounts.

The laptop has the Windows Copilot+ specs.
The secrete sauce? Some general matrix
multiplications (GEMM) tucked in your iGPU:

What is Xe Matrix eXtensions (XMX)? https://www.intel.com/content/www/us/en/support/articles/000091112/graphics.html

Bye

Mild Shock schrieb:

Hi,

Lets get emotional! While Varoufakis painted
the picture of cloud capital. That might have
mobilized "The Internationale", or another

more defensive less motolotov throwing song:

Pink Floyd - Run Like Hell (Live)
https://www.youtube.com/watch?v=lKgOe1Rl8YY

Now since Athropic is teaming with xAI, we
might ask do we see the next OneDrive of Prolog
on the horizon. Even a tame Erlang dream:

populate the Web with clever Prolog agents!
https://trinity.elfenbenstornet.se/

Might have a nasty Prolog as SaaS aspect!
As long as we talk about services and not
assets, we might miss something. Who owns

the present and future LLMs/LRMs?

Bye

Mild Shock schrieb:

Hi,

You just escaped AI dooms day. Humanity has
reset all internet and computers as a last resort
to prevent AGI developing, by an electromagnetic

pulse. You are stuck in G|+ttinger Wald and hunted
down a deer by your bare hands, the deer still
confused and tame because tourists were feeding it.

Now you have no knife, what do you do:

Chimpanzees Have Entered The Stone Age
https://www.youtube.com/watch?v=wPXX2I_uYjc

So we are just apes with internet.

Bye

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 |-a-a | x1 |-a-a-a-a | x4 |
| 0 0 1 0 |-a-a | x2 |-a =-a | x3 |
| 0 1 0 0 |-a-a | x3 |-a-a-a-a | x2 |
| 1 0 0 0 |-a-a | x4 |-a-a-a-a | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog


Hi,

I guess, you can read off how to do it here,
i.e. avoid TDR (DXGI_ERROR_DEVICE_HUNG 0x887A0006).
The below compute toys example has a similar

approach, just like my -C-WAM and Hack GPU backend,
that is based on a Instruction Set Architecture (ISA):

Conway's Game of Life
https://compute.toys/view/2777

He adds a spin to his ISA, that he uses
some gas as computational resource limiter:

for (var gas = 0; gas < 60; gas++) {
if (vm.done == 1u) { break; }
if (vm.pc >= 33u) { vm.done = 1u; break; }

let instr = rom[vm.pc]; vm.pc++;
switch (instr) {
case OP_HALT: { vm.done = 1u; }
case OP_LIT: { vm.stack[vm.sp] = rom[vm.pc]; vm.pc++; vm.sp++; }
case OP_ADD: { vm.sp-=2; vm.stack[vm.sp] =
vm.stack[vm.sp] + vm.stack[vm.sp+1]; vm.sp++; }
case OP_SUB: { vm.sp-=2; vm.stack[vm.sp] =
vm.stack[vm.sp] - vm.stack[vm.sp+1]; vm.sp++; }

Etc..

Now need to do the same for my Hack GPU backend
somehow. Will see how this works out.

But the https://compute.toys/ will do the repeated
invokation of the shader. And the screen syncing
and some pacing to get 60 FPS. So I have to look

at the GitHub source of compute toys as well, in
case it is open source, to at least find a code
template for the repeated dispatch.

Bye

Mild Shock schrieb:

Hi,

You just escaped AI dooms day. Humanity has
reset all internet and computers as a last resort
to prevent AGI developing, by an electromagnetic

pulse. You are stuck in G|+ttinger Wald and hunted
down a deer by your bare hands, the deer still
confused and tame because tourists were feeding it.

Now you have no knife, what do you do:

Chimpanzees Have Entered The Stone Age https://www.youtube.com/watch?v=wPXX2I_uYjc

So we are just apes with internet.

Bye

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 | | x1 | | x4 |
| 0 0 1 0 | | x2 | = | x3 |
| 0 1 0 0 | | x3 | | x2 |
| 1 0 0 0 | | x4 | | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

These novel GPUs , that are part of AI Laptops,
feature Unified Memory Architecture (UMA).
In the case of my Ryzen the main memory is 32 GB,

and the GPU can access 16 GB. It is a design where
CPUs and GPUs share a single coherent memory space,
eliminating the need for separate host and

accelerator memory. Nevertheless the WebGPU API,
works with some copy and synchronization semantics
and buffer abstractions, which is not a big loss.

But all buffers reside in the same AI Lapop
RAM. And the MIMD architecture allows instructions
where seperate threads access the same

memory location, either for read or for write.
There is also a data type atomic(T), which
features operation such as AtomicAdd() etc.. etc..

BTW, I made a GitHub project of my exploration:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

BTW, pi-WAM is nevertheless optimized to have
no memory contention. On the other hand pi-WAM
is happy to access large memory areas.

Bye

BTW: The grandmother of these novel GPUs
is NVIDIAs Volta which already appeared in
2017, meanwhile we have 2026.

See also:

Starting with the NVIDIA Volta architecture,
Independent Thread Scheduling allows full
concurrency between threads, regardless of warp. https://forums.developer.nvidia.com/t/back-to-simd/311983

The AMD Radeon 860M is an integrated graphics
processor that does not have its own dedicated
VRAM. Instead, it dynamically shares up to 50%
of your total system RAM with the CPU in a
standard Windows configuration. https://www.amd.com/en/blogs/2025/faqs-amd-variable-graphics-memory-vram-ai-model-sizes-quantization-mcp-more.html

Ryann Likunov schrieb:

Mild Shock wrote:

Hi,

I guess, you can read off how to do it here,
i.e. avoid TDR (DXGI_ERROR_DEVICE_HUNG 0x887A0006). The below compute
toys example has a similar

approach, just like my -C-WAM and Hack GPU backend,
that is based on a Instruction Set Architecture (ISA):

how come they are not able to map the local RAM
as gpu arrays for using them as local AI resource

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens
inside the L1 and L2 caches of the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try
yourself, in case you find an AI Laptop
with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

I already wrote pi-WAM is designed to
not use memory contention. In particalur
it also uses local variables like pc and

accu of the idependent thread states,
which seems to be also pretty speedy.

Bye

Tanner Babadzhan schrieb:

Mild Shock wrote:

These novel GPUs , that are part of AI Laptops, feature Unified Memory
Architecture (UMA).
In the case of my Ryzen the main memory is 32 GB,

you are in error talking bullshit, the bottleneck there is the max 4

GB/s,

4 times by paralleling, however the proper gddr5/6 gpu arrays goes up to 4,000 GB/s by parallel design

GDDR7 (2025rCo2026 standard)
Max Bandwidth: 1,792 GB/s (RTX 5090, 32GB)

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

But I do grouping of Hack VMs for my pi-WAM in
32 wide work groups. And the there are 4096
/ 32 = 128 such work groups.

Traditionally work groups were executed lockstep:

In GPU architecture, a warp (or wavefront in AMD
terminology) is the fundamental unit of execution,
typically comprising 32 scalar threads. Warps
execute in a SIMT (Single Instruction, Multiple
Thread) fashion, where all 32 threads execute
the same instruction in synchronized lockstep
over different data

New GPUs offer independent thread scheduling:

Modern GPU architectures (such as NVIDIA's Volta
and later) feature Independent Thread Scheduling,
maintaining independent execution states even for
threads within the same warp. This allows the GPU
to yield and resume threads dynamically, essentially
acting as a hardware-level MIMD processor running
on SIMD execution lanes.

I guess MIMD drastically increases the arithmetic
bandwidth for control flow based WGSL code, while
some group arrangement can also increase

the memory bandwidth. By kind of concurrently
flushing L1/L2 caches and reloading L1/L2 caches,
creating some simple sequential systolic computing.

At least I have used a memory layout where threads
from a workgroup are adjacent. Standard processors
continuously fetch data from memory, they suffer from the

"Von Neumann bottleneck". Systolic computing bypasses
this by feeding data into an array of Processing
Elements (PEs) in a wave-like flow.

Bye

Mild Shock schrieb:

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens
inside the L1 and L2 caches of the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try
yourself, in case you find an AI Laptop
with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

I already wrote pi-WAM is designed to
not use memory contention. In particalur
it also uses local variables like pc and

accu of the idependent thread states,
which seems to be also pretty speedy.

Bye

Tanner Babadzhan schrieb:

Mild Shock wrote:

These novel GPUs , that are part of AI Laptops, feature Unified Memory
Architecture (UMA).
In the case of my Ryzen the main memory is 32 GB,

you are in error talking bullshit, the bottleneck there is the max 4

GB/s,

4 times by paralleling, however the proper gddr5/6 gpu arrays goes up to 4,000 GB/s by parallel design

GDDR7 (2025rCo2026 standard)
-a-a-a-a-a Max Bandwidth: 1,792 GB/s (RTX 5090, 32GB)

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Interestingly you can half the AI Laptop
Budget now. The new Mac Neo is only $500 ,
half of my discount AI laptop. Here is

some measurement if this AI Laptop zoo,
including the Mac Neo dwarf (in milliseconds):

AI Laptop example63
Ryzen 542.0
Neo 950.0
Yoga 1475.0
Think 2798.0

AI Laptop example64
Ryzen 1141.0
Neo 2022.0
Yoga 2118.0
Think 7439.0

Since my test doesn't use much memory, it also
fits into into the Mac Neo dwarf 8 GB, and the
Mac Neo dwarf is extrem since the GPU can access

all of 8 GB, not only 50% and the memory is
integrated directly into the processor die,
it is still DRAM not yet HBM. The ARM instruction

set has been streamlined to better support
common GPU interfaces. The CPU acts as a pure
scheduler via native kernel extensions,

appending command buffers straight into memory
queues that the GPU's command processor
reads without a middleman.

Bye

Mild Shock schrieb:

Hi,

But I do grouping of Hack VMs for my pi-WAM in
32 wide work groups. And the there are 4096
/ 32 = 128 such work groups.

Traditionally work groups were executed lockstep:

In GPU architecture, a warp (or wavefront in AMD
terminology) is the fundamental unit of execution,
typically comprising 32 scalar threads. Warps
execute in a SIMT (Single Instruction, Multiple
Thread) fashion, where all 32 threads execute
the same instruction in synchronized lockstep
over different data

New GPUs offer independent thread scheduling:

Modern GPU architectures (such as NVIDIA's Volta
and later) feature Independent Thread Scheduling,
maintaining independent execution states even for
threads within the same warp. This allows the GPU
to yield and resume threads dynamically, essentially
acting as a hardware-level MIMD processor running
on SIMD execution lanes.

I guess MIMD drastically increases the arithmetic
bandwidth for control flow based WGSL code, while
some group arrangement can also increase

the memory bandwidth. By kind of concurrently
flushing L1/L2 caches and reloading L1/L2 caches,
creating some simple sequential systolic computing.

At least I have used a memory layout where threads
from a workgroup are adjacent. Standard processors
continuously fetch data from memory, they suffer from the

"Von Neumann bottleneck". Systolic computing bypasses
-athis by feeding data into an array of Processing
Elements (PEs) in a wave-like flow.

Bye

Mild Shock schrieb:

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens
inside the L1 and L2 caches of the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try
yourself, in case you find an AI Laptop
with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

I already wrote pi-WAM is designed to
not use memory contention. In particalur
it also uses local variables like pc and

accu of the idependent thread states,
which seems to be also pretty speedy.

Bye

Tanner Babadzhan schrieb:

Mild Shock wrote:

These novel GPUs , that are part of AI Laptops, feature Unified

Memory

Architecture (UMA).
In the case of my Ryzen the main memory is 32 GB,

you are in error talking bullshit, the bottleneck there is the max

4 GB/s,

4 times by paralleling, however the proper gddr5/6 gpu arrays goes

up to

4,000 GB/s by parallel design

GDDR7 (2025rCo2026 standard)
-a-a-a-a-a Max Bandwidth: 1,792 GB/s (RTX 5090, 32GB)

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hey Dumbwit,

just run it on your RTX 5070 trash. I am
software developer, not a hardware
develper. I don't care what hardware

people use. Here is the software:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots:

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

You see in the screenshots with a
Ryzen AI 7 350 w/ Radeon 860M that the
results are:

1 Shader 4096 Shaders
542.0 ms 1141.0 ms

What does your RTX 5070 trash deliver?
Just redo the experiment on your hardware.
If the figures are better, well good for

you. If the figures are worse, well I wouldn't
care less. You are just wasting everbodies
bandwidth with your idiotic posts, and being

lazy, instead of replicating the experiment
on your RTX 5070 trash.

Bye

Olin Bagramov schrieb:

Mild Shock wrote:

a wider von Neuann Neck. You can try yourself, in case you find an AI
Laptop with similary specs as the Radeon 860M.

idiot, that's nothing in llm, you are wasting your time

compare with this, if you want proper llm

HBM3e: The current flagship memory for the H200 and B200 series, offering the highest bandwidth (~8 TB/s) required for trillion-parameter models.

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens inside the L1 and L2 caches of
the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try yourself, in case you find an AI
Laptop with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

L1 and L2 are small in size and slow, then the 6 stages pipelining

destroys the neural AI/llm algorithm; compare that with 4,000 GB/s
arrays gddr7 for a graphic card, then talk. You are not good at numbers,
are you
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hey Dumbwit,

We are waiting : 1 Month, 3 Months,
12 Months ... Mostlikely the idiot even
doesn't own a RTX 5070. And if he owns

a RTX 5070 he might struggle with setting
up HTTPS, so that the browser gives you
a WebGPU adapter.

Well the good news is, you don't need
a browser. You could also run it with
node.js. Just use node.js dawn.

See also

Llamas on the Web: Memory-Efficient,
Performance-Portable, and Multi-Precision
LLM Inference with WebGPU
Reese Levine et al. -- 20 May 2026
Figure 2: Breakdown of the LlamaWeb llama.cpp WebGPU
backend and its different paths for executing on GPUs. https://arxiv.org/abs/2605.20706

But I didn't prepare some node.js code
on my GitHub. I also dont use some WASM (*)
helpers, its just pure HTML that taps

into WebGPU via JavaScript inside a HTML page.

Bye

(*) Compute Toys seems to use WASM
to support Slang besides WGSL.

Mild Shock schrieb:

Hey Dumbwit,

just run it on your RTX 5070 trash. I am
software developer, not a hardware
develper. I don't care what hardware

people use. Here is the software:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots:

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

You see in the screenshots with a
Ryzen AI 7 350 w/ Radeon 860M that the
results are:

1 Shader-a-a-a-a-a 4096 Shaders
542.0 ms-a-a-a-a-a 1141.0 ms

What does your RTX 5070 trash deliver?
Just redo the experiment on your hardware.
If the figures are better, well good for

you. If the figures are worse, well I wouldn't
care less. You are just wasting everbodies
bandwidth with your idiotic posts, and being

lazy, instead of replicating the experiment
on your RTX 5070 trash.

Bye

Olin Bagramov schrieb:

Mild Shock wrote:

a wider von Neuann Neck. You can try yourself, in case you find an AI
Laptop with similary specs as the Radeon 860M.

idiot, that's nothing in llm, you are wasting your time

compare with this, if you want proper llm

HBM3e: The current flagship memory for the H200 and B200 series,

offering

the highest bandwidth (~8 TB/s) required for trillion-parameter models.

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens inside the L1 and L2 caches of
the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try yourself, in case you find an AI
Laptop with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

L1 and L2 are small in size and slow, then the 6 stages pipelining

destroys the neural AI/llm algorithm; compare that with 4,000 GB/s
arrays gddr7 for a graphic card, then talk. You are not good at numbers,
are you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hey Dumbwit,

We are still waiting for result. You only
post gibberish:

/* Gibberish I */

idiot, embedding the graphic card gpu into
the cpu, you already have there
the bottleneck, low speeds in ai, disregard the ram size allocated to

the graphic card.

Could you show us the bottleneck, is it
in the same room as us. Whats your proof?
During my testing the CPU just waits:

await outputBuffer.mapAsync(GPUMapMode.READ); https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L181C1-L181C54

Whats your point ultra moron?

/* Gibberish II */

this imbecile doesnt know what ai and llm is, nor using it in coding,

programming etc, an idiot. He is doing graphics, what a fool. AI graphic
cards are not for graphics, cretin. What a fool.

Could you show us where I use graphics?
The screenshots? The screenshots are only
timings shown. Like here:

document.getElementById("result").innerText =
i32s[0].toString() + " /* "+Math.round(performance.now() - start)+" ms */"; https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L184C1-L185C67

Whats your point ultra moron?

It seems you are highly confused Dumbwit!
Go see a doctor as fast as you can.

Bye

Mild Shock schrieb:

Hey Dumbwit,

We are waiting : 1 Month, 3 Months,
12 Months ... Mostlikely the idiot even
doesn't own a RTX 5070. And if he owns

a RTX 5070 he might struggle with setting
up HTTPS, so that the browser gives you
a WebGPU adapter.

Well the good news is, you don't need
a browser. You could also run it with
node.js. Just use node.js dawn.

See also

Llamas on the Web: Memory-Efficient,
Performance-Portable, and Multi-Precision
LLM Inference with WebGPU
Reese Levine et al. -- 20 May 2026
Figure 2: Breakdown of the LlamaWeb llama.cpp WebGPU
backend and its different paths for executing on GPUs. https://arxiv.org/abs/2605.20706

But I didn't prepare some node.js code
on my GitHub. I also dont use some WASM (*)
helpers, its just pure HTML that taps

into WebGPU via JavaScript inside a HTML page.

Bye

(*) Compute Toys seems to use WASM
to support Slang besides WGSL.

Mild Shock schrieb:

Hey Dumbwit,

just run it on your RTX 5070 trash. I am
software developer, not a hardware
develper. I don't care what hardware

people use. Here is the software:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots:

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

You see in the screenshots with a
Ryzen AI 7 350 w/ Radeon 860M that the
results are:

1 Shader-a-a-a-a-a 4096 Shaders
542.0 ms-a-a-a-a-a 1141.0 ms

What does your RTX 5070 trash deliver?
Just redo the experiment on your hardware.
If the figures are better, well good for

you. If the figures are worse, well I wouldn't
care less. You are just wasting everbodies
bandwidth with your idiotic posts, and being

lazy, instead of replicating the experiment
on your RTX 5070 trash.

Bye

Olin Bagramov schrieb:

Mild Shock wrote:

a wider von Neuann Neck. You can try yourself, in case you find an AI >> -a>> Laptop with similary specs as the Radeon 860M.

idiot, that's nothing in llm, you are wasting your time

compare with this, if you want proper llm

HBM3e: The current flagship memory for the H200 and B200 series,

offering

the highest bandwidth (~8 TB/s) required for trillion-parameter

models.

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens inside the L1 and L2 caches of >> -a>> the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try yourself, in case you find an AI >> -a>> Laptop with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

L1 and L2 are small in size and slow, then the 6 stages pipelining

destroys the neural AI/llm algorithm; compare that with 4,000 GB/s
arrays gddr7 for a graphic card, then talk. You are not good at
numbers, are you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Insist on what? Your stupidity? I only
see 404 Brain not Found in your case.
Who cares about 10GB/s, the facts are here:

1 Shader 4096 Shaders
542.0 ms 1141.0 ms

Means with 4096 shaders and the problem at
hand, we still didn't reach the GPU
Knee in the case of a Ryzen AI 7 350

w/ Radeon 860M. If you take another
hardware, you might see another GPU
saturation in the function

f(N) = time used for N shaders.

Bye

P.S.: So whats YOUR hardware and GPU
saturation? Its all open source:

Here is the software:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots (of the timings):

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

i must insist, memory arrays on AI gpu cards
are not for graphics, idiot.

the proof? amazing a prolog guy dont even know what is going on in background, here the speed for embedded gpu/cpu are for instructions
timing, not AI, hence say 10GB/s, which is nothing for running llm AI.

Mild Shock schrieb:

Hey Dumbwit,

We are still waiting for result. You only
post gibberish:

/* Gibberish I */

idiot, embedding the graphic card gpu into
the cpu, you already have there
the bottleneck, low speeds in ai, disregard the ram size allocated to

the graphic card.

Could you show us the bottleneck, is it
in the same room as us. Whats your proof?
During my testing the CPU just waits:

-a-a-a-a-a-a-a await outputBuffer.mapAsync(GPUMapMode.READ); https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L181C1-L181C54

Whats your point ultra moron?

/* Gibberish II */

this imbecile doesnt know what ai and llm is, nor using it in coding,

programming etc, an idiot. He is doing graphics, what a fool. AI graphic cards are not for graphics, cretin. What a fool.

Could you show us where I use graphics?
The screenshots? The screenshots are only
timings shown. Like here:

-a-a-a-a-a-a-a document.getElementById("result").innerText = i32s[0].toString() + " /* "+Math.round(performance.now() - start)+" ms */"; https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L184C1-L185C67

Whats your point ultra moron?

It seems you are highly confused Dumbwit!
Go see a doctor as fast as you can.

Bye

Mild Shock schrieb:

Hey Dumbwit,

We are waiting : 1 Month, 3 Months,
12 Months ... Mostlikely the idiot even
doesn't own a RTX 5070. And if he owns

a RTX 5070 he might struggle with setting
up HTTPS, so that the browser gives you
a WebGPU adapter.

Well the good news is, you don't need
a browser. You could also run it with
node.js. Just use node.js dawn.

See also

Llamas on the Web: Memory-Efficient,
Performance-Portable, and Multi-Precision
LLM Inference with WebGPU
Reese Levine et al. -- 20 May 2026
Figure 2: Breakdown of the LlamaWeb llama.cpp WebGPU
backend and its different paths for executing on GPUs.
https://arxiv.org/abs/2605.20706

But I didn't prepare some node.js code
on my GitHub. I also dont use some WASM (*)
helpers, its just pure HTML that taps

into WebGPU via JavaScript inside a HTML page.

Bye

(*) Compute Toys seems to use WASM
to support Slang besides WGSL.

Mild Shock schrieb:

Hey Dumbwit,

just run it on your RTX 5070 trash. I am
software developer, not a hardware
develper. I don't care what hardware

people use. Here is the software:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots:

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

You see in the screenshots with a
Ryzen AI 7 350 w/ Radeon 860M that the
results are:

1 Shader-a-a-a-a-a 4096 Shaders
542.0 ms-a-a-a-a-a 1141.0 ms

What does your RTX 5070 trash deliver?
Just redo the experiment on your hardware.
If the figures are better, well good for

you. If the figures are worse, well I wouldn't
care less. You are just wasting everbodies
bandwidth with your idiotic posts, and being

lazy, instead of replicating the experiment
on your RTX 5070 trash.

Bye

Olin Bagramov schrieb:

Mild Shock wrote:

a wider von Neuann Neck. You can try yourself, in case you find

an AI

Laptop with similary specs as the Radeon 860M.

idiot, that's nothing in llm, you are wasting your time

compare with this, if you want proper llm

HBM3e: The current flagship memory for the H200 and B200 series,

offering

the highest bandwidth (~8 TB/s) required for trillion-parameter

models.

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens inside the L1 and L2

caches of

the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try yourself, in case you find

an AI

Laptop with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

L1 and L2 are small in size and slow, then the 6 stages pipelining

destroys the neural AI/llm algorithm; compare that with 4,000 GB/s
arrays gddr7 for a graphic card, then talk. You are not good at
numbers, are you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi Dumbwit,

You are a real moron right.

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

What does LIPS mean. Look it up!
Nothing to do with bytes (B).

Bye

Hint: The speed of a Prolog implementation is
sometimes quoted in LIPS - logical inferences
per second.

See also:

Prolog basics CSc 372, Fall 2006 Prolog, Slide 1
W. H. Mitchell (whm@msweng.com) https://www2.cs.arizona.edu/classes/cs372/fall06/prolog.sli.pdf

Mild Shock schrieb:

Hi,

Insist on what? Your stupidity? I only
see 404 Brain not Found in your case.
Who cares about 10GB/s, the facts are here:

1 Shader-a-a-a-a-a 4096 Shaders
542.0 ms-a-a-a-a-a 1141.0 ms

Means with 4096 shaders and the problem at
hand, we still didn't reach the GPU
Knee in the case of a Ryzen AI 7 350

w/ Radeon 860M. If you take another
hardware, you might see another GPU
saturation in the function

f(N) = time used for N shaders.

Bye

P.S.: So whats YOUR hardware and GPU
saturation? Its all open source:

Here is the software:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots (of the timings):

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

i must insist, memory arrays on AI gpu cards
are not for graphics, idiot.

the proof? amazing a prolog guy dont even know what is going on in
background, here the speed for embedded gpu/cpu are for instructions
timing, not AI, hence say 10GB/s, which is nothing for running llm AI.

Mild Shock schrieb:

Hey Dumbwit,

We are still waiting for result. You only
post gibberish:

/* Gibberish I */

idiot, embedding the graphic card gpu into
the cpu, you already have there
the bottleneck, low speeds in ai, disregard the ram size allocated

to the graphic card.

Could you show us the bottleneck, is it
in the same room as us. Whats your proof?
During my testing the CPU just waits:

-a-a-a-a-a-a-a-a await outputBuffer.mapAsync(GPUMapMode.READ);
https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L181C1-L181C54


Whats your point ultra moron?

/* Gibberish II */

this imbecile doesnt know what ai and llm is, nor using it in

coding, programming etc, an idiot. He is doing graphics, what a fool.
AI graphic cards are not for graphics, cretin. What a fool.

Could you show us where I use graphics?
The screenshots? The screenshots are only
timings shown. Like here:

-a-a-a-a-a-a-a-a document.getElementById("result").innerText =
i32s[0].toString() + " /* "+Math.round(performance.now() - start)+" ms
*/";
https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L184C1-L185C67


Whats your point ultra moron?

It seems you are highly confused Dumbwit!
Go see a doctor as fast as you can.

Bye

Mild Shock schrieb:

Hey Dumbwit,

We are waiting : 1 Month, 3 Months,
12 Months ... Mostlikely the idiot even
doesn't own a RTX 5070. And if he owns

a RTX 5070 he might struggle with setting
up HTTPS, so that the browser gives you
a WebGPU adapter.

Well the good news is, you don't need
a browser. You could also run it with
node.js. Just use node.js dawn.

See also

Llamas on the Web: Memory-Efficient,
Performance-Portable, and Multi-Precision
LLM Inference with WebGPU
Reese Levine et al. -- 20 May 2026
Figure 2: Breakdown of the LlamaWeb llama.cpp WebGPU
backend and its different paths for executing on GPUs.
https://arxiv.org/abs/2605.20706

But I didn't prepare some node.js code
on my GitHub. I also dont use some WASM (*)
helpers, its just pure HTML that taps

into WebGPU via JavaScript inside a HTML page.

Bye

(*) Compute Toys seems to use WASM
to support Slang besides WGSL.

Mild Shock schrieb:

Hey Dumbwit,

just run it on your RTX 5070 trash. I am
software developer, not a hardware
develper. I don't care what hardware

people use. Here is the software:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots:

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

You see in the screenshots with a
Ryzen AI 7 350 w/ Radeon 860M that the
results are:

1 Shader-a-a-a-a-a 4096 Shaders
542.0 ms-a-a-a-a-a 1141.0 ms

What does your RTX 5070 trash deliver?
Just redo the experiment on your hardware.
If the figures are better, well good for

you. If the figures are worse, well I wouldn't
care less. You are just wasting everbodies
bandwidth with your idiotic posts, and being

lazy, instead of replicating the experiment
on your RTX 5070 trash.

Bye

Olin Bagramov schrieb:

Mild Shock wrote:

a wider von Neuann Neck. You can try yourself, in case you find

an AI

Laptop with similary specs as the Radeon 860M.

idiot, that's nothing in llm, you are wasting your time

compare with this, if you want proper llm

HBM3e: The current flagship memory for the H200 and B200 series,

offering

the highest bandwidth (~8 TB/s) required for trillion-parameter

models.

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens inside the L1 and L2

caches of

the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try yourself, in case you find

an AI

Laptop with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

L1 and L2 are small in size and slow, then the 6 stages

pipelining destroys the neural AI/llm algorithm; compare that with
4,000 GB/s arrays gddr7 for a graphic card, then talk. You are not
good at numbers, are you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi Dumbwit,

Nobody is interested in discussing AI,
I am discussing prolog on a AI laptop.

Who cares that they use the AI marketing,
when they have splendid GPUs?

Only morons like you!

Bye

Wilfred Durmanov schrieb:

yet another obsolete it-supporter, without proper
education, doesnt know what AI stands there for.
Maybe you should look round to have your face
properly rearranged

Mild Shock schrieb:

Hi Dumbwit,

You are a real moron right.

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

What does LIPS mean. Look it up!
Nothing to do with bytes (B).

Bye

Hint: The speed of a Prolog implementation is
sometimes quoted in LIPS - logical inferences
per second.

See also:

Prolog basics CSc 372, Fall 2006 Prolog, Slide 1
W. H. Mitchell (whm@msweng.com) https://www2.cs.arizona.edu/classes/cs372/fall06/prolog.sli.pdf

Mild Shock schrieb:

Hi,

Insist on what? Your stupidity? I only
see 404 Brain not Found in your case.
Who cares about 10GB/s, the facts are here:

1 Shader-a-a-a-a-a 4096 Shaders
542.0 ms-a-a-a-a-a 1141.0 ms

Means with 4096 shaders and the problem at
hand, we still didn't reach the GPU
Knee in the case of a Ryzen AI 7 350

w/ Radeon 860M. If you take another
hardware, you might see another GPU
saturation in the function

f(N) = time used for N shaders.

Bye

P.S.: So whats YOUR hardware and GPU
saturation? Its all open source:

Here is the software:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots (of the timings):

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

i must insist, memory arrays on AI gpu cards

are not for graphics, idiot.

the proof? amazing a prolog guy dont even know what is going on in
background, here the speed for embedded gpu/cpu are for instructions
timing, not AI, hence say 10GB/s, which is nothing for running llm AI.

Mild Shock schrieb:

Hey Dumbwit,

We are still waiting for result. You only
post gibberish:

/* Gibberish I */

idiot, embedding the graphic card gpu into
the cpu, you already have there
the bottleneck, low speeds in ai, disregard the ram size allocated

to the graphic card.

Could you show us the bottleneck, is it
in the same room as us. Whats your proof?
During my testing the CPU just waits:

-a-a-a-a-a-a-a-a await outputBuffer.mapAsync(GPUMapMode.READ);
https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L181C1-L181C54


Whats your point ultra moron?

/* Gibberish II */

this imbecile doesnt know what ai and llm is, nor using it in

coding, programming etc, an idiot. He is doing graphics, what a fool.
AI graphic cards are not for graphics, cretin. What a fool.

Could you show us where I use graphics?
The screenshots? The screenshots are only
timings shown. Like here:

-a-a-a-a-a-a-a-a document.getElementById("result").innerText =
i32s[0].toString() + " /* "+Math.round(performance.now() - start)+"
ms */";
https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/package.html#L184C1-L185C67


Whats your point ultra moron?

It seems you are highly confused Dumbwit!
Go see a doctor as fast as you can.

Bye

Mild Shock schrieb:

Hey Dumbwit,

We are waiting : 1 Month, 3 Months,
12 Months ... Mostlikely the idiot even
doesn't own a RTX 5070. And if he owns

a RTX 5070 he might struggle with setting
up HTTPS, so that the browser gives you
a WebGPU adapter.

Well the good news is, you don't need
a browser. You could also run it with
node.js. Just use node.js dawn.

See also

Llamas on the Web: Memory-Efficient,
Performance-Portable, and Multi-Precision
LLM Inference with WebGPU
Reese Levine et al. -- 20 May 2026
Figure 2: Breakdown of the LlamaWeb llama.cpp WebGPU
backend and its different paths for executing on GPUs.
https://arxiv.org/abs/2605.20706

But I didn't prepare some node.js code
on my GitHub. I also dont use some WASM (*)
helpers, its just pure HTML that taps

into WebGPU via JavaScript inside a HTML page.

Bye

(*) Compute Toys seems to use WASM
to support Slang besides WGSL.

Mild Shock schrieb:

Hey Dumbwit,

just run it on your RTX 5070 trash. I am
software developer, not a hardware
develper. I don't care what hardware

people use. Here is the software:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Here are the screenshots:

11.4 Giga Lips with a Budget Laptop
https://medium.com/2989/899b0d5c027b

You see in the screenshots with a
Ryzen AI 7 350 w/ Radeon 860M that the
results are:

1 Shader-a-a-a-a-a 4096 Shaders
542.0 ms-a-a-a-a-a 1141.0 ms

What does your RTX 5070 trash deliver?
Just redo the experiment on your hardware.
If the figures are better, well good for

you. If the figures are worse, well I wouldn't
care less. You are just wasting everbodies
bandwidth with your idiotic posts, and being

lazy, instead of replicating the experiment
on your RTX 5070 trash.

Bye

Olin Bagramov schrieb:

Mild Shock wrote:

a wider von Neuann Neck. You can try yourself, in case you find >>>>> an AI
Laptop with similary specs as the Radeon 860M.

idiot, that's nothing in llm, you are wasting your time

compare with this, if you want proper llm

HBM3e: The current flagship memory for the H200 and B200 series, >>>>> offering
the highest bandwidth (~8 TB/s) required for trillion-parameter >>>>> models.

Hi,

You are a fucking moron, arent you?

Most of the stuff in my pi-WAM happens inside the L1 and L2

caches of

the GPU.
Which is faster than normal RAM and has

a wider von Neuann Neck. You can try yourself, in case you find >>>>> an AI
Laptop with similary specs as the Radeon 860M.

The example is open source:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

L1 and L2 are small in size and slow, then the 6 stages

pipelining destroys the neural AI/llm algorithm; compare that with
4,000 GB/s arrays gddr7 for a graphic card, then talk. You are not
good at numbers, are you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog


Hi,

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Budget Laptop means ~1000 USD.
Thats not big money. The Mac Neo is
even better only ~500 USD.

Its still cheaper than a RTX 5070,
which is around ~4000 USD. So if you
have some benchmark data for RTX 5070,

the LIPS, you still pay 4x times more
for each LIPS.

Questions ultra moron?

Bye

Junior Romagna schrieb:

Mild Shock wrote:

Hi Dumbwit,

Nobody is interested in discussing AI,
I am discussing prolog on a AI laptop.

Who cares that they use the AI marketing,
when they have splendid GPUs?

what an idiot, so you put big money in AI laptop do do graphics you

can do

on old 8088 PC etc. Thanks, you just proved you are so fucking stoopid

i can see now why morons like you are using crappy scripting languages
like prolog. No brain required, the prolog is doing the brain for you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

can do on old 8088 PC etc.

You are obviously clueless. You don't need
to look at old PCs, they only did KLips.

Nowadays SWI does MLips. So GLips is a leap!

Bye

Mild Shock schrieb:

Hi,

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Budget Laptop means ~1000 USD.
Thats not big money. The Mac Neo is
even better only ~500 USD.

Its still cheaper than a RTX 5070,
which is around ~4000 USD. So if you
have some benchmark data for RTX 5070,

the LIPS, you still pay 4x times more
for each LIPS.

Questions ultra moron?

Bye

Junior Romagna schrieb:

Mild Shock wrote:

Hi Dumbwit,

Nobody is interested in discussing AI,
I am discussing prolog on a AI laptop.

Who cares that they use the AI marketing,
when they have splendid GPUs?

what an idiot, so you put big money in AI laptop do do graphics you

can do

on old 8088 PC etc. Thanks, you just proved you are so fucking stoopid

i can see now why morons like you are using crappy scripting languages like prolog. No brain required, the prolog is doing the brain for you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Well I wrote:

At the end of 2025 we acquired a couple of AI
Laptops , that were still cheap, since RAM prices
had not yet rocketed.
https://github.com/Jean-Luc-Picard-2021/gigabudget

Here is what I payed, in 2025:

Brack.ch Swift Go 16 AI 1rCO359.00 CHF
inter Discount Lenovo Yoga 7 1rCO199.95 CHF
Lenovo Schweiz ThinkCentre Qualcomm 626.58 CHF

Now take the average price, its ca. 1000 CHF
The Swift Go 16 AI is the Ryzen AI 7 350 w/
Radeon 860M. Its an ultra thin laptop, and

the used Radeon 860M is a mobile grade GPU.
Whats your prololoproblem micro penis?

Bye

Loman Barabashev schrieb:

Mild Shock wrote:

Hi,

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Budget Laptop means ~1000 USD.
Thats not big money. The Mac Neo is even better only ~500 USD.

you stinking sack of rocks, you can't read your
own redundant links. Watch the prices for the Ryzen
AI laptops. Fucking idiot. Not worth the price
for what they give in AI. Those are embedded cpu/gpu
cretin, you don't need to read more.

Mild Shock schrieb:

Hi,

can do on old 8088 PC etc.

You are obviously clueless. You don't need
to look at old PCs, they only did KLips.

Nowadays SWI does MLips. So GLips is a leap!

Bye

Mild Shock schrieb:

Hi,

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Budget Laptop means ~1000 USD.
Thats not big money. The Mac Neo is
even better only ~500 USD.

Its still cheaper than a RTX 5070,
which is around ~4000 USD. So if you
have some benchmark data for RTX 5070,

the LIPS, you still pay 4x times more
for each LIPS.

Questions ultra moron?

Bye

Junior Romagna schrieb:

Mild Shock wrote:

Hi Dumbwit,

Nobody is interested in discussing AI,
I am discussing prolog on a AI laptop.

Who cares that they use the AI marketing,
when they have splendid GPUs?

what an idiot, so you put big money in AI laptop do do graphics you

can do

on old 8088 PC etc. Thanks, you just proved you are so fucking stoopid >> -a>
i can see now why morons like you are using crappy scripting languages >> -a> like prolog. No brain required, the prolog is doing the brain for you

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Micro Penis the putin payed troll, is now
on the payment list of overpriced NVIDIA as
well. He cannot get over the concept of a

Shoe String Budget:

"According to Darren AronofskyrCOs debut
1998 film Pi, rCLeverything around us can
be represented and understood through numbersrCY,
from the stock market to the rCLwax and wane
of caribou populationsrCY. But, ironically,
the filmrCOs own numerical facts feel particularly
inexplicable. It was made on a budget of just
$60,000, which is less a shoestring and more
just the aglet at the end of a frayed lace.
To put that in to perspective, 1999rCOs The
Matrix rCo similarly based around the dystopian
danger of AI rCo may have been slightly slicker
and sexier, but also was made for $60m, a
thousand times more buck for a thousand
times less bang. Plus, it means that its
profit margin was huge. As VarietyrCOs headline
following its release put it, rCyrCY-C=$1,000,000!rCY https://www.vice.com/en/article/pi-soundtrack-20-years-aronofsky-aphex-twin-autechre/

04 Bucephalus Bouncing Ball
https://www.youtube.com/watch?v=Ez6xDGTFswQ

Bye

Loman Barabashev schrieb:

I have a micro penis and my brain explodes

Mild Shock schrieb:

Hi,

Ok I was looking at this learning challenge,
producing vector (y1,y2,y3,y4) from a vector
(x1,x2,x3,x4), System R can do it via least square?

| 0 0 0 1 |-a-a | x1 |-a-a-a-a | x4 |
| 0 0 1 0 |-a-a | x2 |-a =-a | x3 |
| 0 1 0 0 |-a-a | x3 |-a-a-a-a | x2 |
| 1 0 0 0 |-a-a | x4 |-a-a-a-a | x1 |

How it started:

"multiplicative RNNs arises naturally from a
proof-theoretic interpretation of next-token
prediction as nested intuitionistic implication"
Paul Tarau - 2026
https://arxiv.org/abs/2601.19915

How its going:

"Dave uses a PDP-11 to train a real Neural
Network complete with Transformers and
Attention so you can see them at their most basic."
Mr. Taskmanager - 2026
https://www.youtube.com/watch?v=OUE3FSIk46g

We see Doctor Frankstein in action from
the Bronze Age of Computing, producing
a Humunkulus, the progenitor of todays

Bulgakov Shuriks in the Hyperscale Age!

Bye

P.S.: My impression neither cut to the core, that
this incredible transformer most likely
produced this deterministic attention:

| -1 | * | k | + | 5 | = | k' |

Or differently expressed y_k = x_{5-k}.

How did the transformer do it? It produced
a neural network with 1216 parameters, but
didn't use embeddings or polar encoding

of positions. But if we strip the noise
and denoise from the position encoding,
the denoise is done via softmax. We somehow

must get the above, right? I still need to
verify my claim! BTW: The PDP-11 assembly
from 1979 uses wider example not with n=4

but with n=8.

--- Synchronet 3.22a-Linux NewsLink 1.2