Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > sci.logic > #342670

New branding Accelerated Linear Algebra [OpenXLA C++] (Was: I am already using this method for 20 years)

Cross-posted to 3 groups.

Show all headers | View raw

Hi,

Its amazing what can be done with linear
algebra. Want graph isomorphism formulated
in linear algebra? Try this:

      AX = XB

Didn't try yet. But there is currently a rebranding
going on. The fancy term "tensor" seems to be
not extremly accessible for the many engineers

that anyway know linear algebra already. So we
find for eample now TensorFlow mutated into:

XLA (Accelerated Linear Algebra)
is an open source compiler for machine learning. The
XLA compiler takes models from popular frameworks
such as PyTorch, TensorFlow, and JAX, and optimizes
the models for high-performance execution across
different hardware platforms including GPUs, CPUs,
and ML accelerators.
https://openxla.org/?hl=de

So Local AI will again change the landscape, and
it might be no joke that Python disappears. It
could be really the case that C++ becomes

more dominant. Not sure whats going on exactly.

Bye

Mild Shock schrieb:
> Hi,
> 
> I am already using this method for 20 years,
> just copied the Google Page Rank idea a couple
> of decades ago. Its described here:
> 
> We solve this equation exactly in that we
> calculate R = (E-L)-1*H. The matrix inversion
> (E-L)-1 is only computed once for the web site
> and is stored across the pages of the web site.
> May 14th, 2006 - Web Site Ranking
> https://www.herbrand.ai/izytab/doclet/home/03_support/01_browse/05_rank.html 
> 
> 
> While the matrix inversion was very primitive
> in the beginning. I use a couple of ideas to further
> develop it. From using sparsity to using tiles,
> 
> and doing it with multiple works, and worker
> queues. You find everything behind the scenes
> now, I did some explorations. Just image what
> 
> experience Google has and its TPUs to compute
> with matrices, their page rank was what made
> them the first attractive search engine, that
> 
> could provide information retrieval that made sense.
> 
> Bye
> 
> P.S.: Disclaimer, I don't say its possible so
> easily to replicate the new Gemini AI, by simply
> tapping into Local AI via AI Accelerators.
> 
> Its more to get a glimpse how the AI Effloreszenz
> could happen, did happen. I guess mostly it hapoened
> by not by getting obsessed with sequential algorithms
> 
> and all the Stony Brook nonsense, but by adopting a
> linear algebra view point, with very controlled
> doses of parallelism and of course the lock step
> 
> innovation in better hardware.
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Maybe a probem with the Logic Programming community
>> is the fact, that they are totally clueless about
>> linear algebra. Something tells me, we I could
>>
>> compute the transitive closure via matrix inversion:
>>
>> (I-M)^(-1) = I + M + M^2 + M^3 + ...
>>
>> I am pretty sure there are a few Lean or even Rocq
>> theorems already formalized, showing some of these
>> identities and their numerical perks.
>>
>> Still I find here long discussions (**), mareveling
>> how to compute the transitive closure, in a
>> sequential setting. But already when M is sparse
>>
>> things are pretty speedy with linear algebra. Matrices
>> are perfect for AI Accelerators that are specialized
>> for linear algebra, since it is one of the most
>>
>> parallelizable computations in numerical linear algebra.
>> GPU and distributed HPC performance is excellent,
>> but one has possibly to consider AI Accelerators nowadays
>>
>> on laptop and micro servers, for so called Local AI.
>>
>> Bye
>>
>> P.S.: (**)
>> Wanna see a bunch of clueless people associated to Stony Brook or not:
>> https://discourse.prolog-lang.org/t/subgraph-isomorphism-prolog-vs-conventional/173/7 
>>
>>
>> P.P.S.: I am currently looking into combinatorial search,
>> but transitive closure is of course also a candidate to
>> delegate to an AI accelerator when doing Local AI.
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> Woa! I could have told you so. Although hired
>>> by EricGT to provoke me on SWI-Prolog discourse,
>>> and then EricGT kick me out of SWI-Prolog discourse,
>>>
>>> on grounds "not listening to staff". The same
>>> Julio Di Egidio who invented nonsense stuff on
>>> SWI-Prolog discouse, to annoy me.
>>>
>>> Had his lesson teached about the usual code of
>>> conduct in politically correct forums, where one
>>> has to adopt a "performative" way of telling
>>>
>>> technical thruths, either by not telling them,
>>> because the community has already lost their
>>> moral compas, and is only producing hot air,
>>>
>>> or just quitting the community. The lesson was
>>> already posted in 2023:
>>>
>>> Théo Zimmermann
>>> For anyone who might have followed this discussion or any of the 
>>> recent discussions in which Julio Di Egidio was a participant and 
>>> which turned sour, we (administrators of the Coq Zulip and members of 
>>> the Code of Conduct enforcement team) want to let everyone know that 
>>> we do not consider this behavior acceptable (regardless of its 
>>> technical content). A prior warning had been addressed to Julio in 
>>> August 2023. Following this warning and recent new reports, we have 
>>> decided to send a new warning accompanied by a temporary ban, that 
>>> could be lifted with a statement that the offender would follow our 
>>> rules in the future.
>>>
>>> Bye
>>>
>>> Julio Di Egidio schrieb:
>>>> On 10/07/2025 14:31, Julio Di Egidio wrote:
>>>>> On 10/07/2025 14:01, Julio Di Egidio wrote:
>>>>>> The end of the world is nigh:
>>>>>>
>>>>>> <https://rocq-prover.zulipchat.com/#narrow/channel/237656-Rocq-devs-.26-plugin-devs/topic/status.20and.20future.20of.20the.20phase.20split/near/528029103> 
>>>>>>
>>>>>>
>>>>>>  From interactive proof assistant to completely upside-down and
>>>>>> completely broken, and not just on that at that point of course.
>>>>>>
>>>>>> And the fucking shamelessness...
>>>>>
>>>>> But we must thank MS for the nail in that coffin, too: they can't
>>>>> be satisfied with just a Lean broken by design, they must own the
>>>>> whole compartment: only poisoned meatballs for the public...
>>>>
>>>> Follow up on Rocq and the appointed pieces of ungodly shit:
>>>>
>>>> Now if I try to login:
>>>>
>>>> << Account suspended: No constructive purpose to their actions other 
>>>> than creating dissent within the community. >>
>>>>
>>>> You pieces of FRAUDULENT UNGODLY SHIT, the defamation added to the
>>>> abuse, abuse of power as well as personal.
>>>>
>>>> And the last blow, just to mention the icing on the cake, if the
>>>> great people they have been onboarding:
>>>>
>>>> << FYI: I have no animosity towards you. >>
>>>> (Dear Thomas I-am-a-retarded-Nazi Lamiaux, FYI I don't give a shit
>>>> whether you smile or not while shitting all over the place, you
>>>> fucking nazi-retarded pieces of shit, you and the whole gang...)
>>>>
>>>> YOU UNGODLY FRAUDULENT AND RETARDED PIECES OF SHIT, now it's all
>>>> Microsoft and Lean and Lean 2 and sub-Lean and proto-Lean... with
>>>> Microsoft's world-famous best practices, top guidance and impeccable
>>>> products tailored to their customers' real needs...
>>>>
>>>> Indeed you managers and maintainers and marketers and the fucking
>>>> corporate ladder and the whole fucking pyramid indeed: on behalf
>>>> of real humanity, I wish you all a rapid ass cancer!!
>>>>
>>>> Bottom line:
>>>>
>>>> Fuck you. And fuck you especially for 20 years of Coq that you have
>>>> managed to destroy in less than 5 years, you pieces of pyramidic shit.
>>>>
>>>> *Plonk*
>>>>
>>>> Julio
>>>>
>>>
>>
> 

Back to sci.logic | Previous | Next — Previous in thread | Next in thread | Find similar | Unroll thread

Thread

Coq/Rocq and how to completely break it Julio Di Egidio <julio@diegidio.name> - 2025-07-10 14:01 +0200
  Re: Coq/Rocq and how to completely break it Julio Di Egidio <julio@diegidio.name> - 2025-07-10 14:31 +0200
    Maybe AGI should take over proving (Was: Coq/Rocq and how to completely break it) Mild Shock <janburse@fastmail.fm> - 2025-07-15 18:31 +0200
      Having 2544 issues is probably a bad sign (Was: Maybe AGI should take over proving) Mild Shock <janburse@fastmail.fm> - 2025-07-15 18:34 +0200
        The Signal Collapse gives us System Uncertainty (Was: Having 2544 issues is probably a bad sign) Mild Shock <janburse@fastmail.fm> - 2025-07-15 19:01 +0200
    Re: Coq/Rocq and how to completely break it Julio Di Egidio <julio@diegidio.name> - 2025-11-30 09:52 +0100
      Re: Coq/Rocq and how to completely break it Julio Di Egidio <julio@diegidio.name> - 2025-11-30 10:00 +0100
      SWI Agent Provocateur gets kicked out of Rocq? (Re: Coq/Rocq and how to completely break it) Mild Shock <janburse@fastmail.fm> - 2025-11-30 10:28 +0100
        How to get GigaLIPS for Transitive Closure (Was: SWI Agent Provocateur gets kicked out of Rocq?) Mild Shock <janburse@fastmail.fm> - 2025-11-30 11:20 +0100
          I am already using this method for 20 years [AI Effloreszenz] (Was: How to get GigaLIPS for Transitive Closure) Mild Shock <janburse@fastmail.fm> - 2025-11-30 11:32 +0100
            New branding Accelerated Linear Algebra [OpenXLA C++] (Was: I am already using this method for 20 years) Mild Shock <janburse@fastmail.fm> - 2025-11-30 11:43 +0100
        When put in virtual jail, ask for removal of your content (Was: SWI Agent Provocateur gets kicked out of Rocq?) Mild Shock <janburse@fastmail.fm> - 2025-11-30 12:55 +0100

csiph-web