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

Groups > alt.comp.software.thunderbird > #21618 > unrolled thread

"Don't mess with that which you don't understand" ...

Back to article view | Back to alt.comp.software.thunderbird

Contents

  "Don't mess with that which you don't understand" ... "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-07-30 19:38 +0100
    Re: "Don't mess with that which you don't understand" ... T <T@invalid.invalid> - 2026-07-30 16:01 -0700
      Re: "Don't mess with that which you don't understand" ... The Real Bev <bashley101@gmail.com> - 2026-07-30 19:27 -0700
        Re: "Don't mess with that which you don't understand" ... "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-07-31 09:10 +0100
          Re: "Don't mess with that which you don't understand" ... dillinger <dillinger@not.invalid> - 2026-07-31 11:45 +0200
            Re: "Don't mess with that which you don't understand" ... MikeS <MikeS@fred.com> - 2026-07-31 11:56 +0100
              Re: "Don't mess with that which you don't understand" ... T <T@invalid.invalid> - 2026-07-31 11:53 -0700
              Re: "Don't mess with that which you don't understand" ... Paul <nospam@needed.invalid> - 2026-07-31 15:29 -0400
                Re: "Don't mess with that which you don't understand" ... T <T@invalid.invalid> - 2026-08-02 02:05 -0700
            Re: "Don't mess with that which you don't understand" ... Paul <nospam@needed.invalid> - 2026-07-31 10:59 -0400
        Re: "Don't mess with that which you don't understand" ... Paul <nospam@needed.invalid> - 2026-07-31 14:48 -0400
          Re: "Don't mess with that which you don't understand" ... The Real Bev <bashley101@gmail.com> - 2026-08-01 07:19 -0700
            Re: "Don't mess with that which you don't understand" ... Paul <nospam@needed.invalid> - 2026-08-01 14:16 -0400
              Re: "Don't mess with that which you don't understand" ... The Real Bev <bashley101@gmail.com> - 2026-08-01 18:38 -0700
      Re: "Don't mess with that which you don't understand" ... "s|b" <me@privacy.invalid> - 2026-07-31 15:03 +0200
        Re: "Don't mess with that which you don't understand" ... T <T@invalid.invalid> - 2026-07-31 12:37 -0700
    Re: "Don't mess with that which you don't understand" ... T <T@invalid.invalid> - 2026-07-30 16:20 -0700
    Re: "Don't mess with that which you don't understand" ... Retirednoguilt <HapilyRetired@fakeaddress.invalid> - 2026-07-31 10:48 -0400
      Re: "Don't mess with that which you don't understand" ... "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-07-31 21:45 +0100

#21618 — "Don't mess with that which you don't understand" ...

... is often good advice, but IMO a very bad way to give it. It suggests
you have a very low opinion of the person asking, being not far from
insulting - and, how is that person supposed to _get_ such
understanding, other than by asking? That's surely one of the reasons
for the existence of this 'group (and others similar).

Thank goodness for Paul and others willing to _explain_ things. Even if
I don't always understand his explanations, I almost never feel insulted
by them - unlike the responses from another, which weren't even to a
question from me!
-- 
J. P. Gilliver. UMRA: 1960/<1985 MB++G()ALIS-Ch++(p)Ar++T+H+Sh0!:`)DNAf

Look out for #1. Don't step in #2 either.

[toc] | [next] | [standalone]

#21623

On 7/30/26 11:38 AM, J. P. Gilliver wrote:
> ... is often good advice, but IMO a very bad way to give it. It suggests
> you have a very low opinion of the person asking, being not far from
> insulting - and, how is that person supposed to _get_ such
> understanding, other than by asking? That's surely one of the reasons
> for the existence of this 'group (and others similar).
> 
> Thank goodness for Paul and others willing to _explain_ things. Even if
> I don't always understand his explanations, I almost never feel insulted
> by them - unlike the responses from another, which weren't even to a
> question from me!


Both Paul and I learned by "What does this button/knob do?".

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

#21625

On 7/30/26 16:01, T wrote:
> On 7/30/26 11:38 AM, J. P. Gilliver wrote:
>> ... is often good advice, but IMO a very bad way to give it. It suggests
>> you have a very low opinion of the person asking, being not far from
>> insulting - and, how is that person supposed to _get_ such
>> understanding, other than by asking? That's surely one of the reasons
>> for the existence of this 'group (and others similar).
>> 
>> Thank goodness for Paul and others willing to _explain_ things. Even if
>> I don't always understand his explanations, I almost never feel insulted
>> by them - unlike the responses from another, which weren't even to a
>> question from me!
> 
> Both Paul and I learned by "What does this button/knob do?".

That's one of the best ways.  There isn't always an expert to ask, 
although Perplexity and Claude can do a pretty good job.  The 
constantly-increasing complexity of everything is daunting at times.

-- 
Cheers, Bev
    If trees could scream, would we be so cavalier about cutting
    them down?  We might, if they screamed all the time, for
    no good reason.                                 - Jack Handy

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

#21626

On 2026/7/31 3:27:57, The Real Bev wrote:
> On 7/30/26 16:01, T wrote:
>> On 7/30/26 11:38 AM, J. P. Gilliver wrote:
>>> ... is often good advice, but IMO a very bad way to give it. It suggests
>>> you have a very low opinion of the person asking, being not far from
>>> insulting - and, how is that person supposed to _get_ such
>>> understanding, other than by asking? That's surely one of the reasons
>>> for the existence of this 'group (and others similar).

My rant was prompted by at least two posts here by one who did indeed
say more or less the subject, to someone asking what seemed to me an
innocent question.>>>
>>> Thank goodness for Paul and others willing to _explain_ things. Even if
>>> I don't always understand his explanations, I almost never feel insulted
>>> by them - unlike the responses from another, which weren't even to a
>>> question from me!
>>
>> Both Paul and I learned by "What does this button/knob do?".

"A small black light labelled in black lit up saying 'do not press this
button again'." THHGTTG :-)>
> That's one of the best ways.  There isn't always an expert to ask, 

Yes, though wise to back up (e. g. your Thunderbird profile) before
trying :-)

> although Perplexity and Claude can do a pretty good job.  The 

They're (AI in general - I don't know those two by name, though I may
use them - the ones I mostly use are whichever ones comes up if you just
type a question into the Google search box, and the one in YouTube)
certainly getting better. Though have a tendency to state things as true
when they're the opposite! (And then get apologetic when you point this
out! For example, "how do I stop it showing me so many cat videos, or
videos I've just seen?" "Try clearing your viewed history." "That'll
_increase_ the chance of it suggesting the same ones again." "I am
sorry, you are quite right ...")

> constantly-increasing complexity of everything is daunting at times.
> 
It was ever thus. The rate seems to increase, but that impression has
probably been received by those getting older since time immemorial!
-- 
J. P. Gilliver. UMRA: 1960/<1985 MB++G()ALIS-Ch++(p)Ar++T+H+Sh0!:`)DNAf

Censorship is telling a man he can't have a steak just because a baby
can't chew it - Anonymous, 2025-7-2

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

#21627

Op 31-07-2026 om 10:10 schreef J. P. Gilliver:
> My rant was prompted by at least two posts here by one who did indeed
> say more or less the subject, to someone asking what seemed to me an
> innocent question.

Basically Keith asked if was alright to erase his email, and if he could
still access it after he did that, not a very innocent question.
I'm sorry if you felt offended but I'd rather save someone's mail by
offending them a bit then the other way around.

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

#21628

On 31/07/2026 10:45, dillinger wrote:
> Op 31-07-2026 om 10:10 schreef J. P. Gilliver:
>> My rant was prompted by at least two posts here by one who did indeed
>> say more or less the subject, to someone asking what seemed to me an
>> innocent question.
> 
> Basically Keith asked if was alright to erase his email, and if he could
> still access it after he did that, not a very innocent question.
> I'm sorry if you felt offended but I'd rather save someone's mail by
> offending them a bit then the other way around.

There is a great problem with technical advice on forums, especially 
when questions are asked by those with little or zero knowledge 
themselves. How do *they* distinguish responders with real knowledge 
from the many avid posters who masquerade as experts?

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

#21634

On 7/31/26 3:56 AM, MikeS wrote:
> How do *they* distinguish responders with real knowledge from the many 
> avid posters who masquerade as experts?

If they are arrogant and/or condescending, they are posers.

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

#21635

On Fri, 7/31/2026 6:56 AM, MikeS wrote:
> On 31/07/2026 10:45, dillinger wrote:
>> Op 31-07-2026 om 10:10 schreef J. P. Gilliver:
>>> My rant was prompted by at least two posts here by one who did indeed
>>> say more or less the subject, to someone asking what seemed to me an
>>> innocent question.
>>
>> Basically Keith asked if was alright to erase his email, and if he could
>> still access it after he did that, not a very innocent question.
>> I'm sorry if you felt offended but I'd rather save someone's mail by
>> offending them a bit then the other way around.
> 
> There is a great problem with technical advice on forums, especially when 
> questions are asked by those with little or zero knowledge themselves. 
> How do *they* distinguish responders with real knowledge from the many avid
> posters who masquerade as experts?

The more a poster talks, the more you know about them.

You are the sum total, of all the previous answers you gave.

When you receive an answer, you have the option of
amplifying the answer on your own.

When an answer includes tested responses,
that too counts for something, especially if the
response you're seeing... deviates on the very first
command invocation.

   Paul

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

#21648

On 7/31/26 12:29 PM, Paul wrote:
> The more a poster talks, the more you know about them.
> 
> You are the sum total, of all the previous answers you gave.
> 
> When you receive an answer, you have the option of
> amplifying the answer on your own.
> 
> When an answer includes tested responses,
> that too counts for something, especially if the
> response you're seeing... deviates on the very first
> command invocation.
> 
>     Paul

I have also noticed a certain minority of responders that
are not actually helping.  They pepper you with endless
sideways questions that are meant to reveal the poster's
ignorance whilst often times betraying their own ignorance
of the subject while trying to set you up to condescend.

Now it is normal to ask questions to further
clarify a problem.  But when those questions rear off to
other things only remotely related to the problem and
they pepper you with them, I start to get specious.

Also if I play their game, they never give you anything
that is actually helpful (just a lot of snot).  So I
kill file them when I realize what they are up to.  They
are not there to help.  They are there to condescend.
They are stunning examples of the "Dunning–Kruger effect"

The kill file exists to a reason.

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

#21632

On Fri, 7/31/2026 5:45 AM, dillinger wrote:
> Op 31-07-2026 om 10:10 schreef J. P. Gilliver:
>> My rant was prompted by at least two posts here by one who did indeed
>> say more or less the subject, to someone asking what seemed to me an
>> innocent question.
> 
> Basically Keith asked if was alright to erase his email, and if he could
> still access it after he did that, not a very innocent question.
> I'm sorry if you felt offended but I'd rather save someone's mail by
> offending them a bit then the other way around.
> 

I was unaware of the types he was addressing, never seen them
here, but they appear to be "derived types", and made for
the worst of reasons (supporting features that might not work
as they require the OS side of things to work properly and that's
a long distance to go for a pail of water).

I believe GLODA is also supposed to solve those sorts of
problems (integrated search solution), and that's a more self contained
solution that does not leave "that kind of garbage behind".
GLODA makes some kind of inverted Index, which supports email search
later. Since it's self contained, when you find the email
you want, there isn't a fuss trying to get the email open.

For things that are derived, you can erase them. But
as you point out, you have to be pretty sure of the
details of these folders, to make deletion decisions.
It's better to turn off the feature, if there is
slim to none chance of you ever figuring out how
to make the OS side work properly (Federated Search).

Since the design of this program, requires of the user
that they become "somewhat familiar" with the contents
of the profile, that's an inevitable part of usage.
And if more file extensions start appearing, well,
in for a penny, in for a pound, you'll have to learn
whether those are "primary" items, or "derived" items
and safe for deletion. I don't have a problem deleting
things I find in "cache2" on FF or TB, but doing that
sort of cleaning is pointless, as ten minutes later
there are files in there. For that one, it's better
to flip the tool to "RAM caching mode" instead, just
because of the improved hygiene.

Back in Outlook or Outlook Express days, you had
files like .pst and .ost, and like keeping a pet,
you had to learn about the care and feeding of those.
And there were innumerable $39.95 programs for "fixing"
the files. It's not like operating email in the past
was somehow completely free of Profile maintenance.
But the number of things you need to know now, has
definitely gone up. It's like the maintenance
manual for the 747 now.

   Paul

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

#21633

On Thu, 7/30/2026 10:27 PM, The Real Bev wrote:
> On 7/30/26 16:01, T wrote:
>> On 7/30/26 11:38 AM, J. P. Gilliver wrote:
>>> ... is often good advice, but IMO a very bad way to give it. It suggests
>>> you have a very low opinion of the person asking, being not far from
>>> insulting - and, how is that person supposed to _get_ such
>>> understanding, other than by asking? That's surely one of the reasons
>>> for the existence of this 'group (and others similar).
>>>
>>> Thank goodness for Paul and others willing to _explain_ things. Even if
>>> I don't always understand his explanations, I almost never feel insulted
>>> by them - unlike the responses from another, which weren't even to a
>>> question from me!
>>
>> Both Paul and I learned by "What does this button/knob do?".
> 
> That's one of the best ways.  There isn't always an expert to ask, 
> although Perplexity and Claude can do a pretty good job.  The 
> constantly-increasing complexity of everything is daunting at times.

I recommend finding one that runs locally on one of your computers.
If for any reason, the supply of "free ones" dries up, then
having one in a tin will be "better than nothing".

    # Program to run directly, on Linux, as an AppImage

     https://installers.lmstudio.ai/linux/x64/0.4.12-1/LM-Studio-0.4.12-1-x64.AppImage

    # Asking it to write a program, as a demo.

    [Picture]  LLM-AI at work... a demo of asking a question, with High Reasoning turned on

     https://i.postimg.cc/qqXCTQ7V/LMStudio-memory-usage.gif
     ( https://postimg.cc/tnJgd3P2 )

     https://imgur.com/a/1s2JZcv

Not all of them will fit. They range from, say, 4GB, to 1600GB in size,
and the 4GB one will run on a lot of machines. The one I've got is
58GB or so. And with overhead, when it is loaded, it uses 80GB of RAM.
Size is not guaranteed to equate to goodness. A 60GB one is not 15X
better than a 4GB one. It might be 10% better. I picked a setup
that should "comfortably fit" the machine. If there were no free ones
online some day, I can fall back on that one, even if its supply of
current affairs info is limited.

Mine runs on the CPU, as the GPU is too small and gutless for
this work.

Compared to CNN/DNN projects (where getting the acceleration to work
is a bear), if you put a hosting software on your computer, it is
relatively simple to get a query running. When you see a list of
"open weight models", you still have to do a little research
to see what characteristics they have, before you select one.
The hosting softwares aren't always fancy (you won't be doing
Talky Toaster with it), but at least you have a command line
style query box to use.

*******

This is a different OS, and I just loaded a small model on it,
but haven't asked it a question yet.

     # Installer for Windows, a different version than above.  627,261,376 bytes
     # Whereas the above, is just a (portable) runtime.

     https://bionic-installers.lmstudio.ai/win32/x64/1.0.4-9/Bionic-1.0.4-9-x64.exe

     # Start of a Windows LMStudio session, for Big Machine.
     # CPU must have AVX2, according to the web page (for CPU matrix math engine)
     # This ruled out using my 4th Gen machine. I don't have a good supply
     # of "novel instruction" machines, like no AVX512 in the house. AMD now
     # has a "working" AVX512 implementation.

     [Picture]  You can see the listing of models, and one model is smaller than this.
                The Downloader in the program, just downloaded 6GB of files for E4B.

      https://imgur.com/a/b4ChkBn

      https://postimg.cc/dLXjRVJK
      https://i.postimg.cc/rwyZkDbz/W11-gemma-E4B.gif

     [Picture]  This is a test run. It pretends writing a program is too hard for it,
                so it doesn't even know how to cheat. There is no "Reasoning Level"
                setting at the bottom, so I can't crank it up. But this has happened
                before, where the control settings appear out of nowhere, on a particular run.

      https://imgur.com/a/48LlDPp

      https://postimg.cc/XrN3tY8M
      https://i.postimg.cc/T3gwLy5P/LMStudio-Gemma-E4B.gif

I was hoping for a slightly better first session than that.
Now, I'll have to download the 31B model. I don't know
if there's enough disk space over there...

So they all cheat. Nobody just puts their nose to the grindstone
and writes Chudnovsky using nothing but native C. Where is the
work ethic <snicker> ?

That machine can't use more than 300 watts or so, and the GPU
activity does not overlap with the CPU activity. I was kinda
shocked, that it could use a 4GB video card at all. But it seemed to.
Card only has shader code as a resource, no Tensor, NPU, or other
gadgets to aid the Thinking process. The time consuming part of
any run, is outputting Tokens at 10 Tokens per second using
roughly 11 hyperthreaded CPU cores. It does not use all the cores,
on purpose (pretends to not starve the OS).

    Paul

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

#21641

This is so info-dense that I have saved it for later.  I have ~400GB 
empty on my main drive, but mental effort is problematical :-(

The idea of having my own AI is novel.  I am ashamed of having become so 
dependent on the AIs for information that I formerly googled, but I am 
unable to resist the temptation.  Being able to emulate Captain Picard 
(typing, not voice;  my computer has no eyes or ears) is seductive.


On 7/31/26 11:48, Paul wrote:
> On Thu, 7/30/2026 10:27 PM, The Real Bev wrote:
>> On 7/30/26 16:01, T wrote:
>>> On 7/30/26 11:38 AM, J. P. Gilliver wrote:
>>>> ... is often good advice, but IMO a very bad way to give it. It suggests
>>>> you have a very low opinion of the person asking, being not far from
>>>> insulting - and, how is that person supposed to _get_ such
>>>> understanding, other than by asking? That's surely one of the reasons
>>>> for the existence of this 'group (and others similar).
>>>>
>>>> Thank goodness for Paul and others willing to _explain_ things. Even if
>>>> I don't always understand his explanations, I almost never feel insulted
>>>> by them - unlike the responses from another, which weren't even to a
>>>> question from me!
>>>
>>> Both Paul and I learned by "What does this button/knob do?".
>> 
>> That's one of the best ways.  There isn't always an expert to ask, 
>> although Perplexity and Claude can do a pretty good job.  The 
>> constantly-increasing complexity of everything is daunting at times.
> 
> I recommend finding one that runs locally on one of your computers.
> If for any reason, the supply of "free ones" dries up, then
> having one in a tin will be "better than nothing".
> 
>      # Program to run directly, on Linux, as an AppImage
> 
>       https://installers.lmstudio.ai/linux/x64/0.4.12-1/LM-Studio-0.4.12-1-x64.AppImage
> 
>      # Asking it to write a program, as a demo.
> 
>      [Picture]  LLM-AI at work... a demo of asking a question, with High Reasoning turned on
> 
>       https://i.postimg.cc/qqXCTQ7V/LMStudio-memory-usage.gif
>       ( https://postimg.cc/tnJgd3P2 )
> 
>       https://imgur.com/a/1s2JZcv
> 
> Not all of them will fit. They range from, say, 4GB, to 1600GB in size,
> and the 4GB one will run on a lot of machines. The one I've got is
> 58GB or so. And with overhead, when it is loaded, it uses 80GB of RAM.
> Size is not guaranteed to equate to goodness. A 60GB one is not 15X
> better than a 4GB one. It might be 10% better. I picked a setup
> that should "comfortably fit" the machine. If there were no free ones
> online some day, I can fall back on that one, even if its supply of
> current affairs info is limited.
> 
> Mine runs on the CPU, as the GPU is too small and gutless for
> this work.
> 
> Compared to CNN/DNN projects (where getting the acceleration to work
> is a bear), if you put a hosting software on your computer, it is
> relatively simple to get a query running. When you see a list of
> "open weight models", you still have to do a little research
> to see what characteristics they have, before you select one.
> The hosting softwares aren't always fancy (you won't be doing
> Talky Toaster with it), but at least you have a command line
> style query box to use.
> 
> *******
> 
> This is a different OS, and I just loaded a small model on it,
> but haven't asked it a question yet.
> 
>       # Installer for Windows, a different version than above.  627,261,376 bytes
>       # Whereas the above, is just a (portable) runtime.
> 
>       https://bionic-installers.lmstudio.ai/win32/x64/1.0.4-9/Bionic-1.0.4-9-x64.exe
> 
>       # Start of a Windows LMStudio session, for Big Machine.
>       # CPU must have AVX2, according to the web page (for CPU matrix math engine)
>       # This ruled out using my 4th Gen machine. I don't have a good supply
>       # of "novel instruction" machines, like no AVX512 in the house. AMD now
>       # has a "working" AVX512 implementation.
> 
>       [Picture]  You can see the listing of models, and one model is smaller than this.
>                  The Downloader in the program, just downloaded 6GB of files for E4B.
> 
>        https://imgur.com/a/b4ChkBn
> 
>        https://postimg.cc/dLXjRVJK
>        https://i.postimg.cc/rwyZkDbz/W11-gemma-E4B.gif
> 
>       [Picture]  This is a test run. It pretends writing a program is too hard for it,
>                  so it doesn't even know how to cheat. There is no "Reasoning Level"
>                  setting at the bottom, so I can't crank it up. But this has happened
>                  before, where the control settings appear out of nowhere, on a particular run.
> 
>        https://imgur.com/a/48LlDPp
> 
>        https://postimg.cc/XrN3tY8M
>        https://i.postimg.cc/T3gwLy5P/LMStudio-Gemma-E4B.gif
> 
> I was hoping for a slightly better first session than that.
> Now, I'll have to download the 31B model. I don't know
> if there's enough disk space over there...
> 
> So they all cheat. Nobody just puts their nose to the grindstone
> and writes Chudnovsky using nothing but native C. Where is the
> work ethic <snicker> ?
> 
> That machine can't use more than 300 watts or so, and the GPU
> activity does not overlap with the CPU activity. I was kinda
> shocked, that it could use a 4GB video card at all. But it seemed to.
> Card only has shader code as a resource, no Tensor, NPU, or other
> gadgets to aid the Thinking process. The time consuming part of
> any run, is outputting Tokens at 10 Tokens per second using
> roughly 11 hyperthreaded CPU cores. It does not use all the cores,
> on purpose (pretends to not starve the OS).
> 
>      Paul


-- 
Cheers, Bev
    "The primary purpose of any government entity
     is to employ the unemployable."

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

#21643

On Sat, 8/1/2026 10:19 AM, The Real Bev wrote:
> This is so info-dense that I have saved it for later.  I have ~400GB empty on my main drive, but mental effort is problematical :-(
> 
> The idea of having my own AI is novel.  I am ashamed of having become so dependent on the AIs for information that I formerly googled, but I am unable to resist the temptation.  Being able to emulate Captain Picard (typing, not voice;  my computer has no eyes or ears) is seductive.
> 

I'm not good at this stuff, but the latest attempt
suggests the smallest models may be too weak. Since
something like three million people had downloaded
the small one (E4B), I figured there must be some evidence
that one was worthwhile. The bigger one seems braver
about how it runs its business, and does not run
away from a job quite as easily.

And then it becomes a question of how much system RAM
have you got. Gemma 31B wrote a program as an output,
whereas Gemma E4B merely sketched an outline with
"comment statements". And an even larger model found
a cheat for itself, to "save itself from writing 200 lines of code".
How heroic.

 16GB of RAM would handle the  7GB or so Gemma E4B needs (a bit weak model).
 64GB of RAM would handle the 35GB or so Gemma 31B needed.
128GB of RAM would handle the 80GB the OpenAi 120B needed.

Look at both how much RAM and how much disk space. The
OS installation with Gemma 31B on it, there is only 7GB left
on the C: drive, as it wasn't sized for this sort of thing.

And there is a model format which is more efficient and
still works. But they have only released one model that
way, and the industry knows now, how to make a smaller one.
They release FP4 and FP16 models, because video cards accelerate
that, but the smallest model is the TRIT [-1,0,1 ] which is
harder for them but easier for us. TRIT works best if trained
as a TRIT, not quantized from a larger format. Whereas other
models in the list in LMStudio, are quantized from larger
models. The offerings of Kimi could be like that (the largest
model being impossibly big, have to buy a $100K computer, good
for Bill Gates home computer).

   Paul

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

#21646

On 8/1/26 11:16, Paul wrote:
> On Sat, 8/1/2026 10:19 AM, The Real Bev wrote:
>> This is so info-dense that I have saved it for later.  I have ~400GB empty on my main drive, but mental effort is problematical :-(
>> 
>> The idea of having my own AI is novel.  I am ashamed of having become so dependent on the AIs for information that I formerly googled, but I am unable to resist the temptation.  Being able to emulate Captain Picard (typing, not voice;  my computer has no eyes or ears) is seductive.
>> 
> 
> I'm not good at this stuff, but the latest attempt
> suggests the smallest models may be too weak. Since
> something like three million people had downloaded
> the small one (E4B), I figured there must be some evidence
> that one was worthwhile. The bigger one seems braver
> about how it runs its business, and does not run
> away from a job quite as easily.
> 
> And then it becomes a question of how much system RAM
> have you got. Gemma 31B wrote a program as an output,
> whereas Gemma E4B merely sketched an outline with
> "comment statements". And an even larger model found
> a cheat for itself, to "save itself from writing 200 lines of code".
> How heroic.
> 
>   16GB of RAM would handle the  7GB or so Gemma E4B needs (a bit weak model).

16GB.  Oh well.

>   64GB of RAM would handle the 35GB or so Gemma 31B needed.
> 128GB of RAM would handle the 80GB the OpenAi 120B needed.
> 
> Look at both how much RAM and how much disk space. The
> OS installation with Gemma 31B on it, there is only 7GB left
> on the C: drive, as it wasn't sized for this sort of thing.
> 
> And there is a model format which is more efficient and
> still works. But they have only released one model that
> way, and the industry knows now, how to make a smaller one.
> They release FP4 and FP16 models, because video cards accelerate
> that, but the smallest model is the TRIT [-1,0,1 ] which is
> harder for them but easier for us. TRIT works best if trained
> as a TRIT, not quantized from a larger format. Whereas other
> models in the list in LMStudio, are quantized from larger
> models. The offerings of Kimi could be like that (the largest
> model being impossibly big, have to buy a $100K computer, good
> for Bill Gates home computer).
> 
>     Paul


-- 
Cheers, Bev
      Salesmen welcome -- dog food is expensive

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

#21630

On Thu, 30 Jul 2026 16:01:29 -0700, T wrote:

> Both Paul and I learned by "What does this button/knob do?".

I learned a lot that way.

-- 
s|b

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

#21636

On 7/31/26 6:03 AM, s|b wrote:
> On Thu, 30 Jul 2026 16:01:29 -0700, T wrote:
> 
>> Both Paul and I learned by "What does this button/knob do?".
> 
> I learned a lot that way.
> 

When I was a sergeant in the military, the test equipment we
brought on site often would get the knobs turned during
transportation.  My guys were unable to troubleshoot in
the field until someone (like me) got there and reset the
knobs for them.  I was a pain in the neck.

So part of what I trained my guys on was how to operate
the test equipment.  I would set test equipment up on
our shop mock up machines, then have them turn away and
I would goof up all the knobs and they would have
to put them back where they needed to be.

I forced them to learn what the knobs did.  And yes, they
did not appreciate and did not care.   But I forced
the issue.  It did eventually help.

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

#21624

On 7/30/26 11:38 AM, J. P. Gilliver wrote:
> Thank goodness for Paul and others willing to_explain_ things. Even if
> I don't always understand his explanations, I almost never feel insulted
> by them - unlike the responses from another, which weren't even to a
> question from me!

You are tripping across the "Dunning–Kruger effect".
Those that know what they are doing realize they do
not know everything.  And you can't shut them up when
they finally get an audience!

Paul is a force of nature.

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

#21631

On 7/30/2026 2:38 PM, J. P. Gilliver wrote:
> ... is often good advice, but IMO a very bad way to give it. It suggests
> you have a very low opinion of the person asking, being not far from
> insulting - and, how is that person supposed to _get_ such
> understanding, other than by asking? That's surely one of the reasons
> for the existence of this 'group (and others similar).
> 
> Thank goodness for Paul and others willing to _explain_ things. Even if
> I don't always understand his explanations, I almost never feel insulted
> by them - unlike the responses from another, which weren't even to a
> question from me!

I learned that lesson quickly in the Navy onboard one of the submarines.
 The rule of thumb is that if you don't already know, you don't need to
know.  And in many cases, the answer was classified at a higher level
than I had clearance for.

Of course, it also went the other way.  My wife soon learned to stop
asking me when I got home "so what's new?".  Or, "what did you do
today?"  Or, "where did you go today?"  She learned quickly that I would
voluntarily tell her what I could each day and not to press me for more
details when I would return her follow-up question with only a smile.

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

#21637

On 2026/7/31 15:48:6, Retirednoguilt wrote:
[]
> I learned that lesson quickly in the Navy onboard one of the submarines.
>  The rule of thumb is that if you don't already know, you don't need to
> know.  And in many cases, the answer was classified at a higher level
> than I had clearance for.
> 
> Of course, it also went the other way.  My wife soon learned to stop
> asking me when I got home "so what's new?".  Or, "what did you do
> today?"  Or, "where did you go today?"  She learned quickly that I would
> voluntarily tell her what I could each day and not to press me for more
> details when I would return her follow-up question with only a smile.

That's the difference - you used a smile.
-- 
J. P. Gilliver. UMRA: 1960/<1985 MB++G()ALIS-Ch++(p)Ar++T+H+Sh0!:`)DNAf

Viagra won't make you James Bond , but it will make you Roger Moore ..
(@robskiing on YouTube, 2023)

[toc] | [prev] | [standalone]

Back to top | Article view | alt.comp.software.thunderbird

csiph-web