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

Groups > alt.folklore.computers > #234648 > unrolled thread

Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You

Back to article view | Back to alt.folklore.computers

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Anonymous User <noreply@dirge.harmsk.com> - 2026-04-14 16:52 -0400
    Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Jonathan Lamothe <jonathan@jlamothe.net> - 2026-04-14 17:31 -0400
      Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Bill Findlay <findlaybill@blueyonder.co.uk> - 2026-04-15 23:33 +0100
        Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Lawrence D’Oliveiro <ldo@nz.invalid> - 2026-04-15 22:57 +0000
          Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed   Cybersecurity Forever - And Why It Should Scare You Bill Findlay <findlaybill@blueyonder.co.uk> - 2026-04-16 00:15 +0100
          Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Lawrence D’Oliveiro <ldo@nz.invalid> - 2026-04-15 23:20 +0000
    Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You John Ames <commodorejohn@gmail.com> - 2026-04-14 14:52 -0700
      Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Fritz Wuehler <fritz@spamexpire-202604.rodent.frell.theremailer.net> - 2026-04-15 07:55 +0200
      Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You antispam@fricas.org (Waldek Hebisch) - 2026-04-16 03:43 +0000
        Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Jonathan Lamothe <jonathan@jlamothe.net> - 2026-04-16 00:08 -0400
          Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Lawrence D’Oliveiro <ldo@nz.invalid> - 2026-04-16 05:47 +0000
          Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You antispam@fricas.org (Waldek Hebisch) - 2026-04-16 12:51 +0000
            Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Jonathan Lamothe <jonathan@jlamothe.net> - 2026-04-16 11:34 -0400
              Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You antispam@fricas.org (Waldek Hebisch) - 2026-04-16 20:50 +0000
            Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Lawrence D’Oliveiro <ldo@nz.invalid> - 2026-04-16 23:14 +0000
              Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Jonathan Lamothe <jonathan@jlamothe.net> - 2026-04-16 21:23 -0400
              Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed   Cybersecurity Forever - And Why It Should Scare You Bill Findlay <findlaybill@blueyonder.co.uk> - 2026-04-17 13:06 +0100
        Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Anthk <anthk@disroot.org> - 2026-04-25 13:14 +0000
          Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Jonathan Lamothe <jonathan@jlamothe.net> - 2026-04-25 14:26 -0400
            Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You The True Melissa <thetruemelissa@gmail.com> - 2026-04-25 14:33 -0400
              Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Jonathan Lamothe <jonathan@jlamothe.net> - 2026-04-26 09:45 -0400
                Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You The True Melissa <thetruemelissa@gmail.com> - 2026-04-26 10:10 -0400
                  Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You oldernow <oldernow@dev.null> - 2026-04-26 14:22 +0000
                  Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Jonathan Lamothe <jonathan@jlamothe.net> - 2026-04-26 22:51 -0400
                    Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Charlie Gibbs <cgibbs@kltpzyxm.invalid> - 2026-04-28 04:33 +0000
            Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Nuno Silva <nunojsilva@invalid.invalid> - 2026-04-26 00:54 +0100
          Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Lawrence D’Oliveiro <ldo@nz.invalid> - 2026-04-26 01:08 +0000
    Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You Fritz Wuehler <fritz@spamexpire-202604.rodent.frell.theremailer.net> - 2026-04-15 03:35 +0200

Page 1 of 2  [1] 2  Next page →

#234648 — Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You

Anonymous wrote:

> Singularity point has been reached. Once code begins writing code we are 
there.
>
> https://www.youtube.com/watch?v=X_4rKVXev8k

This PC guy is kind of a dumbshit.

Code has been "writing code" since the early 80's.  AIX, VAX/VMS, AT&T
Unix, all "wrote code" based on operational state and runtime requirements
on a regular basis every day.

[toc] | [next] | [standalone]

#234650

Anonymous User <noreply@dirge.harmsk.com> writes:

> Anonymous wrote:
>
>> Singularity point has been reached. Once code begins writing code we are 
> there.
>>
>> https://www.youtube.com/watch?v=X_4rKVXev8k
>
> This PC guy is kind of a dumbshit.
>
> Code has been "writing code" since the early 80's.  AIX, VAX/VMS, AT&T
> Unix, all "wrote code" based on operational state and runtime requirements
> on a regular basis every day.
>

One could argue that we've had "code writing code" since the first
compiler.  I think the bigger issue is the tendency of LLMs to generate
plausible looking nonsense that can be difficult to identify as such.

-- 
Regards,
Jonathan Lamothe
https://jlamothe.net

[toc] | [prev] | [next] | [standalone]

#234697

On 14 Apr 2026, Jonathan Lamothe wrote
(in article <87a4v52pql.fsf@posteo.de>):

> One could argue that we've had "code writing code" since the first
> compiler. I think the bigger issue is the tendency of LLMs to generate
> plausible looking nonsense that can be difficult to identify as such.

Today I asked:

why is lithium 6 a fermion

Reply:

Lithium-6 is classified as a fermion because it has
an odd number of nucleons(three protons and three neutrons),
which gives it half-integer spin.

I then asked:

why does lithium 6 have half-integer spin

And got the reply:

Lithium-6 has half-integer spin because it consists of
an odd number of particles (3 protons and 3 neutrons),
resulting in a net spin of 1

-- 
Bill Findlay

[toc] | [prev] | [next] | [standalone]

#234698

On Wed, 15 Apr 2026 23:33:50 +0100, Bill Findlay wrote:

> why is lithium 6 a fermion

“Fermion” means the object has an odd wave function, which, as far as
I can recall from undergrad physics, has nothing to do with its spin.

The opposite of “fermion” is “boson”, which means it has an even wave
function. The particles that transmit the fundamental forces (e.g.
photons for the electromagnetic force) are bosons.

Fermions obey the Pauli Exclusion Principle, bosons don’t. This
(simplistically) means that two fermions cannot occupy the same space
at the same time.

All matter is made out of fermions. I suppose this is by definition,
really; two objects that could occupy the same space at the same time
would not be considered “material”.

[toc] | [prev] | [next] | [standalone]

#234699 — Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You

On 15 Apr 2026, Lawrence D´Oliveiro wrote
(in article <10rp552$192to$15@dont-email.me>):

> On Wed, 15 Apr 2026 23:33:50 +0100, Bill Findlay wrote:
>
> > why is lithium 6 a fermion
>
> "Fermion" means the object has an odd wave function, which, as far as
> I can recall from undergrad physics, has nothing to do with its spin.
>
> The opposite of "fermion" is "boson", which means it has an even wave
> function. The particles that transmit the fundamental forces (e.g.
> photons for the electromagnetic force) are bosons.
>
> Fermions obey the Pauli Exclusion Principle, bosons don´t. This
> (simplistically) means that two fermions cannot occupy the same space
> at the same time.
>
> All matter is made out of fermions. I suppose this is by definition,
> really; two objects that could occupy the same space at the same time
> would not be considered "material".

Just utterly priceless.

-- 
Bill Findlay

[toc] | [prev] | [next] | [standalone]

#234700

On Wed, 15 Apr 2026 22:57:38 -0000 (UTC), I wrote:

> On Wed, 15 Apr 2026 23:33:50 +0100, Bill Findlay wrote:
>
>> why is lithium 6 a fermion
>
> “Fermion” means the object has an odd wave function, which, as far
> as I can recall from undergrad physics, has nothing to do with its
> spin.

OK, I was wrong about that spin business
<https://en.wikipedia.org/wiki/Fermion>. Seems the lithium-6 nucleus
is a boson, after all. However, note this bit:

    The fermionic or bosonic behaviour of a composite particle is only
    observed when the constituent particles remain far apart. When
    they are close together, the spatial structure becomes important
    and the composite particles behave according to their constituent
    makeup.

Which sounds counterintuitive, but that’s quantum physics for you ...

[toc] | [prev] | [next] | [standalone]

#234651

On Tue, 14 Apr 2026 16:52:06 -0400
Anonymous User <noreply@dirge.harmsk.com> wrote:

> > Singularity point has been reached. Once code begins writing code
> > we are there.
> >
> > https://www.youtube.com/watch?v=X_4rKVXev8k  
> 
> This PC guy is kind of a dumbshit.
> 
> Code has been "writing code" since the early 80's.  AIX, VAX/VMS, AT&T
> Unix, all "wrote code" based on operational state and runtime
> requirements on a regular basis every day.

Also, to the surprise of precisely nobody who's been paying any damn
attention to the "AI" playbook, it turns out the whole "found a million
zillion super-complicated bugs, but they live in Canada, you wouldn't
know them, also we totally *do* have an everything-proof-shield-proof-
sword and a real actual wizard staff that does magic, but they're in
the treehouse and you're not allowed up there" report is, to put it
politely, massively overstated PR hoopla:

https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-claude-mythos-isnt-a-sentient-super-hacker-its-a-sales-pitch-claims-of-thousands-of-severe-zero-days-rely-on-just-198-manual-reviews

...which will doubtless come as a *total* surprise to everyone who's
spent the last three years parroting everything Dario Amodei says un-
questioned, immediately forgetting about any given claim when the next
one farts out of his mouth, and never bothering to go back and check
whether any of his apocalyptic Real Soon Now predictions turned out to
be ludicrous bullshit (spoiler: the answer is "very much yes.")

Meanwhile, the money's already drying up, datacenters are behind
schedule or not even started, the big players have already started on
the "service gets worse" stage of enshittification, OpenAI just killed
its most cost-intensive service mere months after announcing a billion-
dollar deal with Disney for it, and its CFO was making uncomfortable
noises about their prospects for an IPO (which would involve opening
the books for public inspection) before getting the vaudeville hook.

All very healthy and normal!

https://www.wheresyoured.at/the-ai-industry-is-lying-to-you/
https://www.wheresyoured.at/the-subprime-ai-crisis-is-here/
https://www.wheresyoured.at/openai-cfo-news/

[toc] | [prev] | [next] | [standalone]

#234658

John Ames posted:

> On Tue, 14 Apr 2026 16:52:06 -0400
> Anonymous User <noreply@dirge.harmsk.com> wrote:
>
>>> Singularity point has been reached. Once code begins writing code
>>> we are there.
>>>
>>> https://www.youtube.com/watch?v=X_4rKVXev8k
>>
>> This PC guy is kind of a dumbshit.
>>
>> Code has been "writing code" since the early 80's.  AIX, VAX/VMS, AT&T
>> Unix, all "wrote code" based on operational state and runtime
>> requirements on a regular basis every day.
>
> Also, to the surprise of precisely nobody who's been paying any damn
> attention to the "AI" playbook, it turns out the whole "found a million
> zillion super-complicated bugs, but they live in Canada, you wouldn't
> know them, also we totally *do* have an everything-proof-shield-proof-
> sword and a real actual wizard staff that does magic, but they're in
> the treehouse and you're not allowed up there" report is, to put it
> politely, massively overstated PR hoopla:
>
> https://www.tomshardware.com/tech-industry/artificial-
intelligence/anthropics-claude-mythos-isnt-a-sentient-super-hacker-its-a-
sales-pitch-claims-of-thousands-of-severe-zero-days-rely-on-just-198-
manual-reviews
>
> ...which will doubtless come as a *total* surprise to everyone who's
> spent the last three years parroting everything Dario Amodei says un-
> questioned, immediately forgetting about any given claim when the next
> one farts out of his mouth, and never bothering to go back and check
> whether any of his apocalyptic Real Soon Now predictions turned out to
>
> Meanwhile, the money's already drying up, datacenters are behind
> schedule or not even started, the big players have already started on
> the "service gets worse" stage of enshittification, OpenAI just killed
> its most cost-intensive service mere months after announcing a billion-
> dollar deal with Disney for it, and its CFO was making uncomfortable
> noises about their prospects for an IPO (which would involve opening
> the books for public inspection) before getting the vaudeville hook.
>
> All very healthy and normal!
>
> https://www.wheresyoured.at/the-ai-industry-is-lying-to-you/
> https://www.wheresyoured.at/the-subprime-ai-crisis-is-here/
> https://www.wheresyoured.at/openai-cfo-news/

Good observations!

[toc] | [prev] | [next] | [standalone]

#234703

In alt.folklore.computers John Ames <commodorejohn@gmail.com> wrote:
> On Tue, 14 Apr 2026 16:52:06 -0400
> Anonymous User <noreply@dirge.harmsk.com> wrote:
> 
>> > Singularity point has been reached. Once code begins writing code
>> > we are there.
>> >
>> > https://www.youtube.com/watch?v=X_4rKVXev8k  
>> 
>> This PC guy is kind of a dumbshit.
>> 
>> Code has been "writing code" since the early 80's.  AIX, VAX/VMS, AT&T
>> Unix, all "wrote code" based on operational state and runtime
>> requirements on a regular basis every day.
> 
> Also, to the surprise of precisely nobody who's been paying any damn
> attention to the "AI" playbook, it turns out the whole "found a million
> zillion super-complicated bugs, but they live in Canada, you wouldn't
> know them, also we totally *do* have an everything-proof-shield-proof-
> sword and a real actual wizard staff that does magic, but they're in
> the treehouse and you're not allowed up there" report is, to put it
> politely, massively overstated PR hoopla:
> 
> https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-claude-mythos-isnt-a-sentient-super-hacker-its-a-sales-pitch-claims-of-thousands-of-severe-zero-days-rely-on-just-198-manual-reviews
> 
> ...which will doubtless come as a *total* surprise to everyone who's
> spent the last three years parroting everything Dario Amodei says un-
> questioned, immediately forgetting about any given claim when the next
> one farts out of his mouth, and never bothering to go back and check
> whether any of his apocalyptic Real Soon Now predictions turned out to
> be ludicrous bullshit (spoiler: the answer is "very much yes.")
> 
> Meanwhile, the money's already drying up, datacenters are behind
> schedule or not even started, the big players have already started on
> the "service gets worse" stage of enshittification, OpenAI just killed
> its most cost-intensive service mere months after announcing a billion-
> dollar deal with Disney for it, and its CFO was making uncomfortable
> noises about their prospects for an IPO (which would involve opening
> the books for public inspection) before getting the vaudeville hook.
> 
> All very healthy and normal!
> 
> https://www.wheresyoured.at/the-ai-industry-is-lying-to-you/
> https://www.wheresyoured.at/the-subprime-ai-crisis-is-here/
> https://www.wheresyoured.at/openai-cfo-news/

Zitron has some valid points.  But he keeps saying that LLM-s
will not work and that there should be crash.  He wrote that
in 2024, we have 2026 and no crash yet.  I agree that
financial part looks like a bubble which eventually is
likely to burst.

However, I think that LLM-s already can do some work like
first line helpdesk or "paper pushers" work where people
check that on the input there are appropriate documents
and push them for furthere processing.  IIUC LLM-based
machine translation while not perfect made substantial
progress.  There are potentially very lucrative markets
like autonomous weapons and mass surveilance.

Recent news about LLM-generated compiler IMO shows that
there is some potential for code generation.  ATM there
seem to be serious limitations, basically generated code
seem to be as good as available testsuite.  That may
limit applicability, but there is potential to serve as
"copyright anihilator".  First, LLM-s probably can create
non-GPL version of GPL-ed software with good testsuite.
Fuzzing and scraping of user forums may be able to create
testsuites for closed-source packages.

In slightly different spirit, classical copyright
legal theory says that information is _not_ subject
to copyright and only form is portected.  But LLM-s
have huge capability to change form.  One can imagine
LLM-based "clean room" pipeline, with one LLM extracting
information/specification from a copyrighted artifact(s)
and other LLM producing new form.

Considering cost, claimed cost of generated compiler was
$20000.  Even if real cost is higher (due to subsidized
price of LLM use) it may look attractive for businesses:
100 kloc size is likely to require more than 1 year
of developement, so it probably could cost $300000 to
create in USA.  From point of view of managers at
equal cost LLM which can be quickly put to use on demand
has advantage compared to hiring human developers which
needs time (and when dissatisfied humans may go away).
Also, according to report LLM took 2 weeks to create code.
It would be hard for humans to do similar work in this
time: too much for single developer and coordinating
large team could easily take more time than coding.

Raw costs of computations seem to go down.  As Deep Seek
showed better architecture can lower amout of computations
needed by a LLM at given quality.  Apparently LLM-s
struggle with some tasks that are easy for classic
algorithms, so hybrids where LLM delegate algorithmic
parts to appropriate programs can work better and
cheaper.  Current costs seem to consequence of approach
were improvements to quality of LLM output are mainly
due to use of increasingly larger models and of pushing
models to do more computations (beam search, the "chain
of thought" approach, agents etc).  IMO, if LLM provide
sufficient quality there is potential to lower costs.
And I think that for US businesses $300000 a year for
replacing a professional is attractive.  Of course,
for jobs needing low qualifications or entry level
jobs cost must be lower, but I think that for jobs
that can be done now it is lower.

If one wants to look at analogies in the past, then I think
Itanium and internet bubble are relevant.  Internet bubble
caused losses for investors, but internet businnes as a field
survived: weaker players were eliminated, some creasy ideas
dropped, winners were devaluated but businnes still goes
on.  Itanium is an example of technological mistake,
betting that compiler technology will improve.  But it
is also example of betting on monopoly power, without
competiton Itanium had a chance to dominate market.
ATM it is not clear if there is any viable alternative
ot LLM-s.  By viable I mean technology that could
substantially improve productivity.  My understanding
of economic reports is that in developing countries
productivity was stagnant for several years and economic
progress was delivered by outsourcing to countries with
lower labour costs.  But apparently outsourcing is at
its peak.

Like Itanium for Intel, current capital intensive approach
via LLM-s is pretty attractive to big capital.  If it
manages to deliver higher productivity it will fundamentally
shift balance of power between workers and capital in favour
of capital.  That is what big capital wants and due to this
LLM-s are likely to get more funding.  Of course, if LLM-s
fail to deliver they may get defunded at some moment.  But
to deliver what big capital wants LLM-s do not need to be
very good.  Already replacing say 5% of workers at
comparable cost would be big win for capital.  And since
we can expect cost of computations to decrease in the
future main thing is if LLM-s can do acceptable job.

-- 
                              Waldek Hebisch

[toc] | [prev] | [next] | [standalone]

#234704

antispam@fricas.org (Waldek Hebisch) writes:

> However, I think that LLM-s already can do some work like
> first line helpdesk or "paper pushers" work where people
> check that on the input there are appropriate documents
> and push them for furthere processing.  IIUC LLM-based
> machine translation while not perfect made substantial
> progress.  There are potentially very lucrative markets
> like autonomous weapons and mass surveilance.

And it's a *fantastic* idea, too!

https://mashable.com/article/air-canada-forced-to-refund-after-chatbot-misinformation

> [...] there is potential to serve as
> "copyright anihilator".  First, LLM-s probably can create
> non-GPL version of GPL-ed software with good testsuite.

You say that like it's a good thing.

> Raw costs of computations seem to go down.  As Deep Seek
> showed better architecture can lower amout of computations
> needed by a LLM at given quality.

There is *no guarantee* that this will continue to be the case.

-- 
Regards,
Jonathan Lamothe
https://jlamothe.net

[toc] | [prev] | [next] | [standalone]

#234705

On Thu, 16 Apr 2026 00:08:49 -0400, Jonathan Lamothe wrote:

> And it's a *fantastic* idea, too!
>
> https://mashable.com/article/air-canada-forced-to-refund-after-chatbot-misinformation

We had a fun one here in 🇳🇿 few years ago: a recipe advice bot which
would happily accept lists of arbitrary ingredients, including
downright dangerous and poisonous ones, and suggest tasty recipes
using them
<https://www.theregister.com/2023/08/11/supermarket_reins_in_ai_recipebot/>.

[toc] | [prev] | [next] | [standalone]

#234707

Jonathan Lamothe <jonathan@jlamothe.net> wrote:
> antispam@fricas.org (Waldek Hebisch) writes:
> 
>> However, I think that LLM-s already can do some work like
>> first line helpdesk or "paper pushers" work where people
>> check that on the input there are appropriate documents
>> and push them for furthere processing.  IIUC LLM-based
>> machine translation while not perfect made substantial
>> progress.  There are potentially very lucrative markets
>> like autonomous weapons and mass surveilance.
> 
> And it's a *fantastic* idea, too!
> 
> https://mashable.com/article/air-canada-forced-to-refund-after-chatbot-misinformation
> 
>> [...] there is potential to serve as
>> "copyright anihilator".  First, LLM-s probably can create
>> non-GPL version of GPL-ed software with good testsuite.
> 
> You say that like it's a good thing.

Look at context: Zitron claims that LLM-s "do not work".  Here
you have something that LLM-s should be able to do and big
capital may be willing to pay for.  
 
>> Raw costs of computations seem to go down.  As Deep Seek
>> showed better architecture can lower amout of computations
>> needed by a LLM at given quality.
> 
> There is *no guarantee* that this will continue to be the case.

There is no warranty in real life.  However, current
semiconductor improvemements will continue for some time,
I would expect improvements of factor at least 10 in density.
There are improvements in process when manufacturers gain
more experience and scale effects.  So in short term we
can expect that costs of _large_ scale computations will
go down.  In longer term we know that human brain has
more compute power than top current GPU-s and is more
power efficient.  And once you create appropriate genes
organic matter is rather cheap to replicate.  OTOH
inorganic materials seem to have much better electical
properties.  So 3-D inorganic circuits in priciple could
be much more powerful than human brain.

Currently human brain seem to have architectural advantage.
I have seen neural networks folks claim that their
learning algorithms are more efficient than natural
ones.  But human brain seem to learn using much smaller
training data.  We do not know why, maybe despite current
claims natural algortihms are better?  Maybe genetic
material provides pre-learned structure that is only
fine-tuned later?  There is one obvious difference in
sight: classic LLM are trained on fixed data with only
"positive" examples and no feedback.  People learn on
mistakes and schools provide plenty of negative
feedback.  So leting LLM-s act in controlled world
and providing smaller but higher quality training
material could in principle lead to significant
improvement.  As I wrote, there is no warranty of
success, but there are reasonable expectations.

-- 
                              Waldek Hebisch

[toc] | [prev] | [next] | [standalone]

#234708

antispam@fricas.org (Waldek Hebisch) writes:

> There is no warranty in real life.  However, current
> semiconductor improvemements will continue for some time,

What exactly are you basing this on?  We're already making ICs with
internal transistors that are only a few atoms thick.  Unless we use
some drastically different approach to computing, we're pretty close to
hitting a wall here.

https://en.wikipedia.org/wiki/2_nm_process

-- 
Regards,
Jonathan Lamothe
https://jlamothe.net

[toc] | [prev] | [next] | [standalone]

#234709

Jonathan Lamothe <jonathan@jlamothe.net> wrote:
> antispam@fricas.org (Waldek Hebisch) writes:
> 
>> There is no warranty in real life.  However, current
>> semiconductor improvemements will continue for some time,
> 
> What exactly are you basing this on?  We're already making ICs with
> internal transistors that are only a few atoms thick.  Unless we use
> some drastically different approach to computing, we're pretty close to
> hitting a wall here.

Transistor gates are few atoms thick and that is unlikely to
change.  But other dimensions are much larger and can shrink.

> https://en.wikipedia.org/wiki/2_nm_process

If you looked inside of that article you will note that 2nm
is nominal resolution.  Actual chip features are much (10-20 times)
larger.  And most installed GPU-s were made on older process.
It is expected that improvement eventually will hit a wall.
I would say that we have about 5 years before what is possible
now becomes mainstream.  Beyond that it is hard to make
predicitions, we maybe there will be progress, maybe
difficulties are too hard to overcome.

-- 
                              Waldek Hebisch

[toc] | [prev] | [next] | [standalone]

#234710

On Thu, 16 Apr 2026 12:51:57 -0000 (UTC), Waldek Hebisch wrote:

> But human brain seem to learn using much smaller training data.

Also prone to believing things that aren’t true.

Coincidence? You be the judge.

[toc] | [prev] | [next] | [standalone]

#234712

Lawrence D’Oliveiro <ldo@nz.invalid> writes:

> On Thu, 16 Apr 2026 12:51:57 -0000 (UTC), Waldek Hebisch wrote:
>
>> But human brain seem to learn using much smaller training data.
>
> Also prone to believing things that aren’t true.
>
> Coincidence? You be the judge.

Are you implying LLMs aren't?

-- 
Regards,
Jonathan Lamothe
https://jlamothe.net

[toc] | [prev] | [next] | [standalone]

#234714 — Re: BRAXMAN: The Skynet Moment: How Mythos AI Just Changed Cybersecurity Forever - And Why It Should Scare You

On 17 Apr 2026, Lawrence D´Oliveiro wrote
(in article <10rrqh6$225jt$1@dont-email.me>):

> On Thu, 16 Apr 2026 12:51:57 -0000 (UTC), Waldek Hebisch wrote:
>
> > But human brain seem to learn using much smaller training data.
>
> Also prone to believing things that aren´t true.

Ipse dixit

-- 
Bill Findlay

[toc] | [prev] | [next] | [standalone]

#234744

On 2026-04-16, Waldek Hebisch <antispam@fricas.org> wrote:
> In alt.folklore.computers John Ames <commodorejohn@gmail.com> wrote:
>> On Tue, 14 Apr 2026 16:52:06 -0400
>> Anonymous User <noreply@dirge.harmsk.com> wrote:
>> 
>>> > Singularity point has been reached. Once code begins writing code
>>> > we are there.
>>> >
>>> > https://www.youtube.com/watch?v=X_4rKVXev8k  
>>> 
>>> This PC guy is kind of a dumbshit.
>>> 
>>> Code has been "writing code" since the early 80's.  AIX, VAX/VMS, AT&T
>>> Unix, all "wrote code" based on operational state and runtime
>>> requirements on a regular basis every day.
>> 
>> Also, to the surprise of precisely nobody who's been paying any damn
>> attention to the "AI" playbook, it turns out the whole "found a million
>> zillion super-complicated bugs, but they live in Canada, you wouldn't
>> know them, also we totally *do* have an everything-proof-shield-proof-
>> sword and a real actual wizard staff that does magic, but they're in
>> the treehouse and you're not allowed up there" report is, to put it
>> politely, massively overstated PR hoopla:
>> 
>> https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-claude-mythos-isnt-a-sentient-super-hacker-its-a-sales-pitch-claims-of-thousands-of-severe-zero-days-rely-on-just-198-manual-reviews
>> 
>> ...which will doubtless come as a *total* surprise to everyone who's
>> spent the last three years parroting everything Dario Amodei says un-
>> questioned, immediately forgetting about any given claim when the next
>> one farts out of his mouth, and never bothering to go back and check
>> whether any of his apocalyptic Real Soon Now predictions turned out to
>> be ludicrous bullshit (spoiler: the answer is "very much yes.")
>> 
>> Meanwhile, the money's already drying up, datacenters are behind
>> schedule or not even started, the big players have already started on
>> the "service gets worse" stage of enshittification, OpenAI just killed
>> its most cost-intensive service mere months after announcing a billion-
>> dollar deal with Disney for it, and its CFO was making uncomfortable
>> noises about their prospects for an IPO (which would involve opening
>> the books for public inspection) before getting the vaudeville hook.
>> 
>> All very healthy and normal!
>> 
>> https://www.wheresyoured.at/the-ai-industry-is-lying-to-you/
>> https://www.wheresyoured.at/the-subprime-ai-crisis-is-here/
>> https://www.wheresyoured.at/openai-cfo-news/
>
> Zitron has some valid points.  But he keeps saying that LLM-s
> will not work and that there should be crash.  He wrote that
> in 2024, we have 2026 and no crash yet.  I agree that
> financial part looks like a bubble which eventually is
> likely to burst.
>
> However, I think that LLM-s already can do some work like
> first line helpdesk or "paper pushers" work where people
> check that on the input there are appropriate documents
> and push them for furthere processing.  IIUC LLM-based
> machine translation while not perfect made substantial
> progress.  There are potentially very lucrative markets
> like autonomous weapons and mass surveilance.
>
> Recent news about LLM-generated compiler IMO shows that
> there is some potential for code generation.  ATM there
> seem to be serious limitations, basically generated code
> seem to be as good as available testsuite.  That may
> limit applicability, but there is potential to serve as
> "copyright anihilator".  First, LLM-s probably can create
> non-GPL version of GPL-ed software with good testsuite.
> Fuzzing and scraping of user forums may be able to create
> testsuites for closed-source packages.
>
> In slightly different spirit, classical copyright
> legal theory says that information is _not_ subject
> to copyright and only form is portected.  But LLM-s
> have huge capability to change form.  One can imagine
> LLM-based "clean room" pipeline, with one LLM extracting
> information/specification from a copyrighted artifact(s)
> and other LLM producing new form.
>
> Considering cost, claimed cost of generated compiler was
> $20000.  Even if real cost is higher (due to subsidized
> price of LLM use) it may look attractive for businesses:
> 100 kloc size is likely to require more than 1 year
> of developement, so it probably could cost $300000 to
> create in USA.  From point of view of managers at
> equal cost LLM which can be quickly put to use on demand
> has advantage compared to hiring human developers which
> needs time (and when dissatisfied humans may go away).
> Also, according to report LLM took 2 weeks to create code.
> It would be hard for humans to do similar work in this
> time: too much for single developer and coordinating
> large team could easily take more time than coding.
>
> Raw costs of computations seem to go down.  As Deep Seek
> showed better architecture can lower amout of computations
> needed by a LLM at given quality.  Apparently LLM-s
> struggle with some tasks that are easy for classic
> algorithms, so hybrids where LLM delegate algorithmic
> parts to appropriate programs can work better and
> cheaper.  Current costs seem to consequence of approach
> were improvements to quality of LLM output are mainly
> due to use of increasingly larger models and of pushing
> models to do more computations (beam search, the "chain
> of thought" approach, agents etc).  IMO, if LLM provide
> sufficient quality there is potential to lower costs.
> And I think that for US businesses $300000 a year for
> replacing a professional is attractive.  Of course,
> for jobs needing low qualifications or entry level
> jobs cost must be lower, but I think that for jobs
> that can be done now it is lower.
>
> If one wants to look at analogies in the past, then I think
> Itanium and internet bubble are relevant.  Internet bubble
> caused losses for investors, but internet businnes as a field
> survived: weaker players were eliminated, some creasy ideas
> dropped, winners were devaluated but businnes still goes
> on.  Itanium is an example of technological mistake,
> betting that compiler technology will improve.  But it
> is also example of betting on monopoly power, without
> competiton Itanium had a chance to dominate market.
> ATM it is not clear if there is any viable alternative
> ot LLM-s.  By viable I mean technology that could
> substantially improve productivity.  My understanding
> of economic reports is that in developing countries
> productivity was stagnant for several years and economic
> progress was delivered by outsourcing to countries with
> lower labour costs.  But apparently outsourcing is at
> its peak.
>
> Like Itanium for Intel, current capital intensive approach
> via LLM-s is pretty attractive to big capital.  If it
> manages to deliver higher productivity it will fundamentally
> shift balance of power between workers and capital in favour
> of capital.  That is what big capital wants and due to this
> LLM-s are likely to get more funding.  Of course, if LLM-s
> fail to deliver they may get defunded at some moment.  But
> to deliver what big capital wants LLM-s do not need to be
> very good.  Already replacing say 5% of workers at
> comparable cost would be big win for capital.  And since
> we can expect cost of computations to decrease in the
> future main thing is if LLM-s can do acceptable job.
>

Compilers are deterministic, at least in tons of languages.
LLM's arent.

Even the worst compiler/interpreter from https://t3x.org
it's far more useful than any LLM *over time* because
you can be sure the output will be 100% the same no
matter what you are trying to implement.

[toc] | [prev] | [next] | [standalone]

#234745

Anthk <anthk@disroot.org> writes:

> Compilers are deterministic, at least in tons of languages.
> LLM's arent.
>
> Even the worst compiler/interpreter from https://t3x.org
> it's far more useful than any LLM *over time* because
> you can be sure the output will be 100% the same no
> matter what you are trying to implement.

Also, compilers don't "hallucinate".  This is not a property of LLMs
that anyone has any idea how to correct.

-- 
Regards,
Jonathan Lamothe
https://jlamothe.net - PGP: 9CF2CE03EBF08E8C8B66C3660198463E3CF3FFD1
I � Unicode

[toc] | [prev] | [next] | [standalone]

#234746

Verily, in article <87ldeazyj0.fsf@posteo.de>, did jonathan@jlamothe.net 
deliver unto us this message:
> Anthk <anthk@disroot.org> writes:
> 
> > [quoted text muted]
> >
> > Even the worst compiler/interpreter from https://t3x.org
> > it's far more useful than any LLM *over time* because
> > you can be sure the output will be 100% the same no
> > matter what you are trying to implement.
> 
> Also, compilers don't "hallucinate".  This is not a property of LLMs
> that anyone has any idea how to correct.
> 

The problem is that it's always "hallucinating." It has no idea what 
it's actually saying, at least at this point. It's a tribute to math and 
ingenuity that the result usually makes sense and is often even 
accurate. 

-- 
The True Melissa - Canal Winchester - Ohio
United States of America - North America - Earth
Solar System - Milky Way - Local Group
Virgo Cluster - Laniakea Supercluster - Cosmos

[toc] | [prev] | [next] | [standalone]

Page 1 of 2  [1] 2  Next page →

Back to top | Article view | alt.folklore.computers

csiph-web