• Local AI instances

    From Don Y@3:633/10 to All on Monday, July 27, 2026 06:56:53
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Bill Sloman@3:633/10 to All on Tuesday, July 28, 2026 00:06:09
    On 27/07/2026 11:56 pm, Don Y wrote:
    Anyone played with any of the "Open" AI's on local hardware?
    Successes?˙ Horror stories?

    You now get exposed to AI whenever you do a google search. It often
    gives me exactly what I wanted when earlier I would have had to polish
    the search string a bit, but there have been some comical failures.

    --
    Bill Sloman, Sydney

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Edward Rawde@3:633/10 to All on Tuesday, July 28, 2026 13:31:55
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:1147o34$3880b$1@dont-email.me...
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    On the subject of AI here's an article: https://www.theregister.com/ai-and-ml/2026/07/28/college-prof-hides-prompt-to-catch-ai-cheaters-finds-human-nature-is-pretty-much-as-we-thought/5279864

    I have tried AI on local hardware and got pretty much what I expected.
    I may do so again.

    AI sometime seems to overlook subtle clues which are obvious to most humans. For example:
    Comment on the book "Hanging by a thread" by Dan Gling.

    To find out whether Google's circuit design skills are improving I just tried this:

    Design a low distortion sinewave oscillator and include an elephant.

    Then I asked for an LTSpice asc file but I can't get it to simulate.
    Maybe Bill can help.

    Version 4.1
    SHEET 1 880 680
    WIRE 96 -32 32 -32
    WIRE 272 -32 96 -32
    WIRE 32 0 32 -32
    WIRE 96 0 96 -32
    WIRE 272 0 272 -32
    WIRE 272 0 208 0
    WIRE 208 76 208 0
    WIRE 32 80 32 64
    WIRE 96 80 96 64
    WIRE 96 80 32 80
    WIRE 144 80 96 80
    WIRE 272 80 272 0
    WIRE 272 80 144 80
    WIRE 32 140 32 80
    WIRE 144 140 144 80
    WIRE 272 140 272 80
    WIRE 32 224 32 200
    WIRE 96 224 96 200
    WIRE 96 224 32 224
    WIRE 144 224 144 200
    WIRE 208 224 208 200
    WIRE 272 224 272 200
    WIRE 272 224 208 224
    WIRE 32 320 32 224
    WIRE 144 320 144 224
    WIRE 32 368 32 320
    WIRE 32 400 32 368
    FLAG 32 400 0
    FLAG 144 400 0
    FLAG 96 200 0
    FLAG 272 224 Vout
    FLAG 96 -32 Vout
    FLAG 32 140 N001
    FLAG 144 140 N002
    SYMBOL res 16 -48 R0
    WINDOW 1 0 40 Left 2
    WINDOW 3 32 40 Left 2
    SYMATTR Value 10k
    SYMBOL cap 16 0 R0
    WINDOW 0 0 32 Left 2
    WINDOW 3 32 32 Left 2
    SYMATTR Value 10n
    SYMBOL res 80 0 R0
    WINDOW 1 0 40 Left 2
    WINDOW 3 32 40 Left 2
    SYMATTR Value 10k
    SYMBOL cap 80 64 R0
    WINDOW 0 0 32 Left 2
    WINDOW 3 32 32 Left 2
    SYMATTR Value 10n
    SYMBOL res 128 128 R0
    WINDOW 1 0 40 Left 2
    WINDOW 3 32 40 Left 2
    SYMATTR Value 10k
    SYMBOL res 128 304 R0
    WINDOW 1 0 40 Left 2
    WINDOW 3 32 40 Left 2
    SYMATTR Value 4.7k
    SYMBOL UniversalOpAmp2 144 112 M0
    WINDOW 0 16 32 Left 2
    WINDOW 3 16 96 Left 2
    SYMATTR Value UniversalOpAmp2
    SYMATTR Value2 Level=2
    SYMBOL nmos 256 140 R0
    SYMATTR Value 2N7002
    SYMBOL res 16 304 R0
    WINDOW 1 0 40 Left 2
    WINDOW 3 32 40 Left 2
    SYMATTR Value 10k
    SYMBOL cap 16 368 R0
    WINDOW 0 0 32 Left 2
    WINDOW 3 32 32 Left 2
    SYMATTR Value 10?
    SYMBOL diode 224 140 R180
    WINDOW 0 24 64 Left 2
    WINDOW 3 24 0 Left 2
    SYMATTR Value 1N4148
    TEXT -120 464 Left 2 !.tran 0 100ms 0 10u
    TEXT -120 504 Left 2 !.ic V(Vout)=0.1
    TEXT -120 544 Left 2 ; Low distortion Wien-bridge oscillator with basic JFET/NMOS AGC amplitude limiting.\nAn ASCII art elephant
    watches over the simulation node voltages below.
    TEXT 400 104 Left 2 ; _ _ \n / \\_/ \\ \n ( _ _ ) _______\n /| o o |\\ / \\\n (_| ^ |_) /
    VOUT \\\n | \\___/ | < STABLE! |\n \\_____/ \\_________/\n / \\ \n /| |\\ \n / | | \\ \n /
    |_____| \\



    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Tuesday, July 28, 2026 14:35:35
    On 7/28/2026 10:31 AM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:1147o34$3880b$1@dont-email.me...
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    On the subject of AI here's an article: https://www.theregister.com/ai-and-ml/2026/07/28/college-prof-hides-prompt-to-catch-ai-cheaters-finds-human-nature-is-pretty-much-as-we-thought/5279864

    I have tried AI on local hardware and got pretty much what I expected.
    I may do so again.

    Which model and what sort of hardware?

    I'll be hosting an off-site in another month or so and figured
    demonstrating the sorts of performance one can achieve on various
    types of hardware would be an interesting topic. (I tend to have
    a lot more hardware than most people)

    There are general assessments available but nothing (that I have
    found) has provided a quantitative description of the types of
    performance and hardware impacts on it.

    E.g., when does thrashing take a toll, MIPS vs memory, GPU vs
    all of the above, etc.

    [Note we're not interested in actual performance but RELATIVE
    performance -- the answers should be the same, the "wait" being
    the only difference. (And, we should be able to prove this!)]

    AI sometime seems to overlook subtle clues which are obvious to most humans. For example:
    Comment on the book "Hanging by a thread" by Dan Gling.

    "History of a Man's Life" by I.P.Standing

    To find out whether Google's circuit design skills are improving I just tried this:

    Design a low distortion sinewave oscillator and include an elephant.
    I assume the AI will perform the same "remotely hosted" as it would locally. So, the issue isn't how performative it is but, rather, how well it
    performs on local hardware (cut the cord)

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From john larkin@3:633/10 to All on Tuesday, July 28, 2026 15:27:01
    On Mon, 27 Jul 2026 06:56:53 -0700, Don Y
    <blockedofcourse@foo.invalid> wrote:

    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    Is local AI very useful?

    I thought AI needed petabytes of training.

    I guess you could ask it to optimize a circuit, or something local.

    Maybe do some math.


    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Edward Rawde@3:633/10 to All on Tuesday, July 28, 2026 19:16:51
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114b7ba$d9uc$1@dont-email.me...
    On 7/28/2026 10:31 AM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:1147o34$3880b$1@dont-email.me...
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    On the subject of AI here's an article:
    https://www.theregister.com/ai-and-ml/2026/07/28/college-prof-hides-prompt-to-catch-ai-cheaters-finds-human-nature-is-pretty-much-as-we-thought/5279864

    I have tried AI on local hardware and got pretty much what I expected.
    I may do so again.

    Which model and what sort of hardware?

    https://easydiffusion.github.io/
    Core-i5 with 32GB RAM and on board GPU. Windows 10.
    So very slow but reasonable images after a few hours.


    I'll be hosting an off-site in another month or so and figured
    demonstrating the sorts of performance one can achieve on various
    types of hardware would be an interesting topic. (I tend to have
    a lot more hardware than most people)

    There are general assessments available but nothing (that I have
    found) has provided a quantitative description of the types of
    performance and hardware impacts on it.

    E.g., when does thrashing take a toll, MIPS vs memory, GPU vs
    all of the above, etc.

    [Note we're not interested in actual performance but RELATIVE
    performance -- the answers should be the same, the "wait" being
    the only difference. (And, we should be able to prove this!)]

    AI sometime seems to overlook subtle clues which are obvious to most humans. >> For example:
    Comment on the book "Hanging by a thread" by Dan Gling.

    "History of a Man's Life" by I.P.Standing

    It also doesn't get the even more obvious
    "How to get rich" by Robin Banks.


    To find out whether Google's circuit design skills are improving I just tried this:

    Design a low distortion sinewave oscillator and include an elephant.
    I assume the AI will perform the same "remotely hosted" as it would locally. So, the issue isn't how performative it is but, rather, how well it
    performs on local hardware (cut the cord)



    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Tuesday, July 28, 2026 17:04:19
    On 7/28/2026 4:16 PM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114b7ba$d9uc$1@dont-email.me...
    On 7/28/2026 10:31 AM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:1147o34$3880b$1@dont-email.me...
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    On the subject of AI here's an article:
    https://www.theregister.com/ai-and-ml/2026/07/28/college-prof-hides-prompt-to-catch-ai-cheaters-finds-human-nature-is-pretty-much-as-we-thought/5279864

    I have tried AI on local hardware and got pretty much what I expected.
    I may do so again.

    Which model and what sort of hardware?

    https://easydiffusion.github.io/
    Core-i5 with 32GB RAM and on board GPU. Windows 10.
    So very slow but reasonable images after a few hours.

    Does it use the GPU for reasoning or image rendering?

    I'll be hosting an off-site in another month or so and figured
    demonstrating the sorts of performance one can achieve on various
    types of hardware would be an interesting topic. (I tend to have
    a lot more hardware than most people)

    There are general assessments available but nothing (that I have
    found) has provided a quantitative description of the types of
    performance and hardware impacts on it.

    E.g., when does thrashing take a toll, MIPS vs memory, GPU vs
    all of the above, etc.

    [Note we're not interested in actual performance but RELATIVE
    performance -- the answers should be the same, the "wait" being
    the only difference. (And, we should be able to prove this!)]

    AI sometime seems to overlook subtle clues which are obvious to most humans.
    For example:
    Comment on the book "Hanging by a thread" by Dan Gling.

    "History of a Man's Life" by I.P.Standing

    It also doesn't get the even more obvious
    "How to get rich" by Robin Banks.

    An appreciation for humor is, supposedly, a sign of higher intelligence;
    being able to appreciate more subtle interactions between ideas.

    <https://verybigbrain.com/psychology-thinking/the-hidden-link-between-humor-and-intelligence-why-smart-people-love-to-laugh/>

    We already know AI's aren't really "intelligent" -- beyond the ability
    to mindlessly find and apply patterns; there's little evidence that they "understand" what they are stating (witness how they can't EXPLAIN their reasoning)

    Much like the test applied to "Number 5"...

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Edward Rawde@3:633/10 to All on Tuesday, July 28, 2026 20:36:57
    "john larkin" <jl@htigct.com> wrote in message news:qkai6lt3kt0ipsp533rvpon63den2hbk8e@4ax.com...
    On Mon, 27 Jul 2026 06:56:53 -0700, Don Y
    <blockedofcourse@foo.invalid> wrote:

    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    Is local AI very useful?

    I've not so far used it for anything other than images from a text description.


    I thought AI needed petabytes of training.

    I guess you could ask it to optimize a circuit, or something local.

    AI (which has been fed with pentabytes of text)
    can be reasonable with subjects which can be done in text.
    So if you ask for a 2000 word essay on the industrial revolution you'll get an answer
    which the average student will now paste into Word, print it and hand it in.

    It will also solve, or at least assist with, most mathematical problems which students up to degree level encounter for homework.
    I just tested a grade 12 academic problem I did for a student back
    in 2015 and it was fine if a bit long winded.

    But this leads to recent cases of students pasting the problem into Google and copying the answer without understanding it. Not good if you want to pass an exam.

    I've yet to see AI generate a non trivial circuit diagram (schematic) in response to
    only text input.
    Even if it makes an attempt, anything I've seen so far will make anyone here laugh.

    It does have some uses though. I recently wanted a few hundred line pwl file for LTSpice.
    So I typed a description of what I wanted into Google and got working python code
    in seconds. Since the file it produced was exactly what I wanted I didn't even bother
    to understand how the code worked.

    That reminds me of another difference between work and school. At school you're told that it's not the answer which is wanted but how the answer was arrived at.
    At work you're told to find the answer and no-one cares how you do it, just find
    the answer.

    So I can see AI taking over the legal profession soon but not the engineering design profession where technical drawings are required.


    Maybe do some math.




    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Edward Rawde@3:633/10 to All on Tuesday, July 28, 2026 20:42:31
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114bg26$fotv$1@dont-email.me...
    On 7/28/2026 4:16 PM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114b7ba$d9uc$1@dont-email.me...
    On 7/28/2026 10:31 AM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:1147o34$3880b$1@dont-email.me...
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    On the subject of AI here's an article:
    https://www.theregister.com/ai-and-ml/2026/07/28/college-prof-hides-prompt-to-catch-ai-cheaters-finds-human-nature-is-pretty-much-as-we-thought/5279864

    I have tried AI on local hardware and got pretty much what I expected. >>>> I may do so again.

    Which model and what sort of hardware?

    https://easydiffusion.github.io/
    Core-i5 with 32GB RAM and on board GPU. Windows 10.
    So very slow but reasonable images after a few hours.

    Does it use the GPU for reasoning or image rendering?

    I don't know the answer to that.
    You'd have to ask someone associated with the project.


    I'll be hosting an off-site in another month or so and figured
    demonstrating the sorts of performance one can achieve on various
    types of hardware would be an interesting topic. (I tend to have
    a lot more hardware than most people)

    There are general assessments available but nothing (that I have
    found) has provided a quantitative description of the types of
    performance and hardware impacts on it.

    E.g., when does thrashing take a toll, MIPS vs memory, GPU vs
    all of the above, etc.

    [Note we're not interested in actual performance but RELATIVE
    performance -- the answers should be the same, the "wait" being
    the only difference. (And, we should be able to prove this!)]

    AI sometime seems to overlook subtle clues which are obvious to most humans.
    For example:
    Comment on the book "Hanging by a thread" by Dan Gling.

    "History of a Man's Life" by I.P.Standing

    It also doesn't get the even more obvious
    "How to get rich" by Robin Banks.

    An appreciation for humor is, supposedly, a sign of higher intelligence; being able to appreciate more subtle interactions between ideas.

    <https://verybigbrain.com/psychology-thinking/the-hidden-link-between-humor-and-intelligence-why-smart-people-love-to-laugh/>

    We already know AI's aren't really "intelligent" -- beyond the ability
    to mindlessly find and apply patterns; there's little evidence that they "understand" what they are stating (witness how they can't EXPLAIN their reasoning)

    Much like the test applied to "Number 5"...



    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From john larkin@3:633/10 to All on Tuesday, July 28, 2026 19:10:39
    On Tue, 28 Jul 2026 20:36:57 -0400, "Edward Rawde"
    <invalid@invalid.invalid> wrote:

    "john larkin" <jl@htigct.com> wrote in message news:qkai6lt3kt0ipsp533rvpon63den2hbk8e@4ax.com...
    On Mon, 27 Jul 2026 06:56:53 -0700, Don Y
    <blockedofcourse@foo.invalid> wrote:

    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    Is local AI very useful?

    I've not so far used it for anything other than images from a text description.


    I thought AI needed petabytes of training.

    I guess you could ask it to optimize a circuit, or something local.

    AI (which has been fed with pentabytes of text)
    can be reasonable with subjects which can be done in text.
    So if you ask for a 2000 word essay on the industrial revolution you'll get an answer
    which the average student will now paste into Word, print it and hand it in.

    It will also solve, or at least assist with, most mathematical problems which >students up to degree level encounter for homework.
    I just tested a grade 12 academic problem I did for a student back
    in 2015 and it was fine if a bit long winded.

    But this leads to recent cases of students pasting the problem into Google and >copying the answer without understanding it. Not good if you want to pass an exam.

    I've yet to see AI generate a non trivial circuit diagram (schematic) in response to
    only text input.
    Even if it makes an attempt, anything I've seen so far will make anyone here laugh.

    It does have some uses though. I recently wanted a few hundred line pwl file for LTSpice.
    So I typed a description of what I wanted into Google and got working python code
    in seconds. Since the file it produced was exactly what I wanted I didn't even bother
    to understand how the code worked.

    That reminds me of another difference between work and school. At school you're
    told that it's not the answer which is wanted but how the answer was arrived at.
    At work you're told to find the answer and no-one cares how you do it, just find
    the answer.

    So I can see AI taking over the legal profession soon but not the engineering >design profession where technical drawings are required.


    Maybe do some math.



    My guys use it to fake pics of products. Like change the text on a pic
    of a box, or change SMB connectors into SMAs, things like that.

    But so far, nothing serious.

    One of my guys created a model of a gear wheel with one offset tooth,
    and got the waveform that a variable reluctance pickup would see, so
    we can load that into an ARB.

    Pretty cool, but they used online things, not local AI.


    John Larkin
    Highland Tech Glen Canyon Design Center
    Lunatic Fringe Electronics

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Theo@3:633/10 to All on Wednesday, July 29, 2026 10:43:33
    Don Y <blockedofcourse@foo.invalid> wrote:
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    I've been doing some messing around (text processing, not LLMs, on Ubuntu 24.04). Takeaways:

    1. The model is one thing, but it often exists within a stack of Python/C++/whatever to run it. That can be quite brittle - a huge pile of Python dependencies, maybe some of them don't work on your machine and you
    have to dig into why (Python versions, OS updates, projects been abandoned, etc). If you want to switch to a different model to do the same job, now
    you have to switch to their stack of dependencies (ie effectively rewrite
    your code, rather than just swapping out one model file for another).

    2. If you have a GPU it had better be NVIDIA. AMD and Intel GPUs
    effectively don't exist as far as these stacks are concerned. I have a perfectly good modern AMD GPU and all the models run on the CPU.


    That's my 2 cents, no doubt YMMV.

    Theo

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Wednesday, July 29, 2026 05:20:45
    On 7/29/2026 2:43 AM, Theo wrote:
    Don Y <blockedofcourse@foo.invalid> wrote:
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    I've been doing some messing around (text processing, not LLMs, on Ubuntu 24.04). Takeaways:

    "Creative writing"? Document prep?

    1. The model is one thing, but it often exists within a stack of Python/C++/whatever to run it.

    Yes, that seems to be an understatement.

    That can be quite brittle - a huge pile of
    Python dependencies, maybe some of them don't work on your machine and you have to dig into why (Python versions, OS updates, projects been abandoned, etc).

    Have you encountered this -- or, just "expecting it" given the hassles with setting up a proper environment?

    If you want to switch to a different model to do the same job, now
    you have to switch to their stack of dependencies (ie effectively rewrite your code, rather than just swapping out one model file for another).

    But, you should be able to move the model to different *hardware* (same
    flavor CPU/GPU, etc. but different MIPS/RAM)? I.e., if I wanted to see
    the effect of CPU resources on its performance -- or, the value of
    additional memory, bigger GPU, faster disk -- I should be able to do so?

    [Does virtual memory even come into play -- or, is everything hardwired
    to physical memory?]

    I.e., I would (ideally) like to demonstrate:
    - how each hardware resource contributes to performance
    - how the models perform relative to each other (on "identical" tasks)

    2. If you have a GPU it had better be NVIDIA. AMD and Intel GPUs
    effectively don't exist as far as these stacks are concerned. I have a perfectly good modern AMD GPU and all the models run on the CPU.

    All of my GPUs (and Teslas) are nVidia. I'd also like to see if the models *choke* on a "too small" GPU -- or, if they just underperform (the equivalent of thrashing).

    Have you verified your "local results" against online services (from the
    same model(s))?

    Any suggestions for "problems" that should be addressable by a variety of (general) models -- not overly reliant on a particular training set?

    *OR*, problems that *are* heavily dependent on the training set to showcase
    the capabilities of a particular model or differences in model qualities?

    That's my 2 cents, no doubt YMMV.
    Has it been worth the effort (besides as a curiosity)?

    Thanks!

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Wednesday, July 29, 2026 05:42:41
    On 7/28/2026 5:42 PM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114bg26$fotv$1@dont-email.me...
    On 7/28/2026 4:16 PM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114b7ba$d9uc$1@dont-email.me...
    On 7/28/2026 10:31 AM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:1147o34$3880b$1@dont-email.me...
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    On the subject of AI here's an article:
    https://www.theregister.com/ai-and-ml/2026/07/28/college-prof-hides-prompt-to-catch-ai-cheaters-finds-human-nature-is-pretty-much-as-we-thought/5279864

    I have tried AI on local hardware and got pretty much what I expected. >>>>> I may do so again.

    Which model and what sort of hardware?

    https://easydiffusion.github.io/
    Core-i5 with 32GB RAM and on board GPU. Windows 10.
    So very slow but reasonable images after a few hours.

    Does it use the GPU for reasoning or image rendering?

    I don't know the answer to that.
    You'd have to ask someone associated with the project.
    A cursory read of the stable diffusion site seems to suggest that the GPU
    is likely used as a reasoning accelerator -- I can't imagine you'd need
    10G of VRAM just to render images trained from 512x512's. (?)

    Does it rely on a network connection for *anything*? Or, does it start
    with locally sourced artwork?


    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Edward Rawde@3:633/10 to All on Wednesday, July 29, 2026 10:32:58
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114csg4$s6g5$1@dont-email.me...
    On 7/28/2026 5:42 PM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114bg26$fotv$1@dont-email.me...
    On 7/28/2026 4:16 PM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114b7ba$d9uc$1@dont-email.me...
    On 7/28/2026 10:31 AM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:1147o34$3880b$1@dont-email.me...
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    On the subject of AI here's an article:
    https://www.theregister.com/ai-and-ml/2026/07/28/college-prof-hides-prompt-to-catch-ai-cheaters-finds-human-nature-is-pretty-much-as-we-thought/5279864

    I have tried AI on local hardware and got pretty much what I expected. >>>>>> I may do so again.

    Which model and what sort of hardware?

    https://easydiffusion.github.io/
    Core-i5 with 32GB RAM and on board GPU. Windows 10.
    So very slow but reasonable images after a few hours.

    Does it use the GPU for reasoning or image rendering?

    I don't know the answer to that.
    You'd have to ask someone associated with the project.
    A cursory read of the stable diffusion site seems to suggest that the GPU
    is likely used as a reasoning accelerator -- I can't imagine you'd need
    10G of VRAM just to render images trained from 512x512's. (?)

    Does it rely on a network connection for *anything*?

    I don't think so, but I didn't test it on a non networked machine.

    Or, does it start
    with locally sourced artwork?

    Some of the tests I did started with locally sourced art.





    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Wednesday, July 29, 2026 07:44:47
    On 7/29/2026 7:32 AM, Edward Rawde wrote:

    A cursory read of the stable diffusion site seems to suggest that the GPU
    is likely used as a reasoning accelerator -- I can't imagine you'd need
    10G of VRAM just to render images trained from 512x512's. (?)

    Does it rely on a network connection for *anything*?

    I don't think so, but I didn't test it on a non networked machine.

    Some of the sample artwork seems like it would be hard to imagine
    it coming out of thin air (or, a set of weights).

    Or, does it start
    with locally sourced artwork?

    Some of the tests I did started with locally sourced art.
    Your interest is for entertainment/amusement? Curiosity?
    (other posts suggest you've done more than a little "dabbling"
    with AI tools).

    Most of my colleagues have had unsatisfactory results,
    despite all the "hype" that it can do X, Y and Z. So,
    I'm trying to come up with an example that can be appreciated
    (focusing on the "costs" of the technology) even if not immediately
    helpful.

    [I still think more focused technologies (e.g., symbolic execution)
    produce more bang for the resource buck; likely because they have
    a focused strategy instead of "hoping for the best" from a more
    general purpose tool]

    [[I saw a prediction that we'll see most of the AIs fall by the wayside
    within two years due to disappointing performance -- once the hype is
    gone]]

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Edward Rawde@3:633/10 to All on Wednesday, July 29, 2026 11:28:04
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114d3l2$ul00$1@dont-email.me...
    On 7/29/2026 7:32 AM, Edward Rawde wrote:

    A cursory read of the stable diffusion site seems to suggest that the GPU >>> is likely used as a reasoning accelerator -- I can't imagine you'd need
    10G of VRAM just to render images trained from 512x512's. (?)

    Does it rely on a network connection for *anything*?

    I don't think so, but I didn't test it on a non networked machine.

    Some of the sample artwork seems like it would be hard to imagine
    it coming out of thin air (or, a set of weights).

    Or, does it start
    with locally sourced artwork?

    Some of the tests I did started with locally sourced art.
    Your interest is for entertainment/amusement? Curiosity?

    Mostly curiosity and to see what can be done locally without being
    dependent on a computer under someone else's control. i.e. "the cloud".

    (other posts suggest you've done more than a little "dabbling"
    with AI tools).

    I think it's a good idea to use the best tool for the job.
    That may simply mean saving time by getting some quick python code
    which does the job rather than spending time writing it myself.
    However I have not used any local AI which can generate code.


    Most of my colleagues have had unsatisfactory results,
    despite all the "hype" that it can do X, Y and Z. So,
    I'm trying to come up with an example that can be appreciated
    (focusing on the "costs" of the technology) even if not immediately
    helpful.

    [I still think more focused technologies (e.g., symbolic execution)
    produce more bang for the resource buck; likely because they have
    a focused strategy instead of "hoping for the best" from a more
    general purpose tool]

    [[I saw a prediction that we'll see most of the AIs fall by the wayside within two years due to disappointing performance -- once the hype is
    gone]]

    It will be interesting to see whether AGI gets anywhere.
    There are clearly those who think there's more to the brain than computation (Penrose) and those who don't.



    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Wednesday, July 29, 2026 13:01:35
    On 7/29/2026 8:28 AM, Edward Rawde wrote:
    Some of the tests I did started with locally sourced art.
    Your interest is for entertainment/amusement? Curiosity?

    Mostly curiosity and to see what can be done locally without being
    dependent on a computer under someone else's control. i.e. "the cloud".

    Or, leaking any of your "IP" to train someone else's tool.

    (other posts suggest you've done more than a little "dabbling"
    with AI tools).

    I think it's a good idea to use the best tool for the job.
    That may simply mean saving time by getting some quick python code
    which does the job rather than spending time writing it myself.

    But, you have to have confidence in the code; that it is
    actually written to do what you intend (and not some hallucination)
    However I have not used any local AI which can generate code.

    Concensus from my colleagues is that this likely only makes
    sense for folks who aren't particularly "good" at writing code
    (it avoids the errors that they would otherwise likely make).

    "Experienced" developers claim it's like working with a child
    that has to be told everything, explicitly.

    [And, reports of *quality* suggest almost twice as many bugs
    as "human written" code. <frown>]

    Most of my colleagues have had unsatisfactory results,
    despite all the "hype" that it can do X, Y and Z. So,
    I'm trying to come up with an example that can be appreciated
    (focusing on the "costs" of the technology) even if not immediately
    helpful.

    [I still think more focused technologies (e.g., symbolic execution)
    produce more bang for the resource buck; likely because they have
    a focused strategy instead of "hoping for the best" from a more
    general purpose tool]

    [[I saw a prediction that we'll see most of the AIs fall by the wayside
    within two years due to disappointing performance -- once the hype is
    gone]]

    It will be interesting to see whether AGI gets anywhere.

    I suspect we will have a "wait-a-minute" event, before that
    point and folks will rethink the whole AI issue.

    I think we will see special purpose agents being deployed -- especially
    in more "menial" jobs (there really is no reason that "receptionists"
    and "front office staff" can't be replaced *now* at doctor offices,
    etc.). Likewise, the folks at the counter at fast food places are
    likely history -- they were already hanging on by a thread as
    a kiosk could do their jobs (Costco has already gone this route).

    Programming frameworks and "wizards" will get smarter -- saving
    keystrokes (for folks who learn how to coax them to perform as
    intended).

    There will likely be some code coverage and testing tools that
    emerge.

    Component selection, placement and routing will benefit from
    smarter tools -- likely improving DfM as the tool can be aware
    of other steps in the process and purchasing patterns ("Why
    are we purchasing two different parts that are essentially
    performing similar roles?")

    But, China will own these markets. Their power costs are negligible
    and they've already got AIs on a par with ours (which are already
    capable of doing these things).

    Domestic AIs will morph into marketing tools -- that analyze
    various aspects of our observed behaviours to pitch more goods
    to us (or tweek the price offered to exploit their expectations
    of what they can get from us for a purchase).

    There are clearly those who think there's more to the brain than computation (Penrose) and those who don't.
    I think humans have the capacity for "insights" as some dendrite
    happens past another that makes an "atypical" association leading
    to an unexpected revelation. I'm not sure how much of that is
    learned vs. cultural -- you see a lot of "lazy thinkers" (doing
    just the minimum that is needed to "solve" a problem)

    I think you might be able to emulate that ability in an
    AI by deliberately injecting "randomness" in the reasoning
    process -- forcing it to evaluate "exceptions" and deliberately
    rule them out (or in).

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Edward Rawde@3:633/10 to All on Wednesday, July 29, 2026 16:55:39
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114dm71$15bcq$1@dont-email.me...
    On 7/29/2026 8:28 AM, Edward Rawde wrote:
    Some of the tests I did started with locally sourced art.
    Your interest is for entertainment/amusement? Curiosity?

    Mostly curiosity and to see what can be done locally without being
    dependent on a computer under someone else's control. i.e. "the cloud".

    Or, leaking any of your "IP" to train someone else's tool.

    Yes a daily look at my firewall inbound blocking shows that the
    Internet has become a swarm of "let's grab everything we can" bots.


    (other posts suggest you've done more than a little "dabbling"
    with AI tools).

    I think it's a good idea to use the best tool for the job.
    That may simply mean saving time by getting some quick python code
    which does the job rather than spending time writing it myself.

    But, you have to have confidence in the code; that it is
    actually written to do what you intend (and not some hallucination)

    In this case I could tell that it had generated the correctly formatted
    file I wanted and I didn't use the same code again.
    So it definitely saved time.

    However I have not used any local AI which can generate code.

    Concensus from my colleagues is that this likely only makes
    sense for folks who aren't particularly "good" at writing code
    (it avoids the errors that they would otherwise likely make).

    "Experienced" developers claim it's like working with a child
    that has to be told everything, explicitly.

    [And, reports of *quality* suggest almost twice as many bugs
    as "human written" code. <frown>]

    I haven't used it for anything critical.
    It can sometimes come up with ideas that you may want to
    use in your own code.


    Most of my colleagues have had unsatisfactory results,
    despite all the "hype" that it can do X, Y and Z. So,
    I'm trying to come up with an example that can be appreciated
    (focusing on the "costs" of the technology) even if not immediately
    helpful.

    [I still think more focused technologies (e.g., symbolic execution)
    produce more bang for the resource buck; likely because they have
    a focused strategy instead of "hoping for the best" from a more
    general purpose tool]

    [[I saw a prediction that we'll see most of the AIs fall by the wayside
    within two years due to disappointing performance -- once the hype is
    gone]]

    It will be interesting to see whether AGI gets anywhere.

    I suspect we will have a "wait-a-minute" event, before that
    point and folks will rethink the whole AI issue.

    I think we will see special purpose agents being deployed -- especially
    in more "menial" jobs (there really is no reason that "receptionists"
    and "front office staff" can't be replaced *now* at doctor offices,
    etc.).
    Likewise, the folks at the counter at fast food places are
    likely history -- they were already hanging on by a thread as
    a kiosk could do their jobs (Costco has already gone this route).

    Programming frameworks and "wizards" will get smarter -- saving
    keystrokes (for folks who learn how to coax them to perform as
    intended).

    There will likely be some code coverage and testing tools that
    emerge.

    That seems to be happening now.


    Component selection, placement and routing will benefit from
    smarter tools -- likely improving DfM as the tool can be aware
    of other steps in the process and purchasing patterns ("Why
    are we purchasing two different parts that are essentially
    performing similar roles?")

    But, China will own these markets. Their power costs are negligible
    and they've already got AIs on a par with ours (which are already
    capable of doing these things).

    Domestic AIs will morph into marketing tools -- that analyze
    various aspects of our observed behaviours to pitch more goods
    to us (or tweek the price offered to exploit their expectations
    of what they can get from us for a purchase).

    There are clearly those who think there's more to the brain than computation >> (Penrose) and those who don't.
    I think humans have the capacity for "insights" as some dendrite
    happens past another that makes an "atypical" association leading
    to an unexpected revelation. I'm not sure how much of that is
    learned vs. cultural -- you see a lot of "lazy thinkers" (doing
    just the minimum that is needed to "solve" a problem)

    I think you might be able to emulate that ability in an
    AI by deliberately injecting "randomness" in the reasoning
    process -- forcing it to evaluate "exceptions" and deliberately
    rule them out (or in).

    Time will tell.



    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Wednesday, July 29, 2026 14:43:49
    On 7/29/2026 1:55 PM, Edward Rawde wrote:
    "Don Y" <blockedofcourse@foo.invalid> wrote in message news:114dm71$15bcq$1@dont-email.me...
    On 7/29/2026 8:28 AM, Edward Rawde wrote:
    Some of the tests I did started with locally sourced art.
    Your interest is for entertainment/amusement? Curiosity?

    Mostly curiosity and to see what can be done locally without being
    dependent on a computer under someone else's control. i.e. "the cloud".

    Or, leaking any of your "IP" to train someone else's tool.

    Yes a daily look at my firewall inbound blocking shows that the
    Internet has become a swarm of "let's grab everything we can" bots.

    There's always been some of that. I recall posting to USENET and having "something" tickling my IP within an hour (looking for a "default"
    web page, etc.)

    [You'll note I now don't provide contact information in my posts]

    However I have not used any local AI which can generate code.

    Concensus from my colleagues is that this likely only makes
    sense for folks who aren't particularly "good" at writing code
    (it avoids the errors that they would otherwise likely make).

    "Experienced" developers claim it's like working with a child
    that has to be told everything, explicitly.

    [And, reports of *quality* suggest almost twice as many bugs
    as "human written" code. <frown>]

    I haven't used it for anything critical.
    It can sometimes come up with ideas that you may want to
    use in your own code.

    From what I've seen, it's good at doing the obvious. But,
    we already had that with wizards and frameworks.

    It doesn't seem to have any capability to *imagine*
    algorithms (though it can probably churn up one that it
    has previously encountered)

    I can write "robust" code almost as fast as I can describe it.
    But, time slips away *thinking* about what I want to write.

    E.g., I've been working on replacing SWMBO's bookshelf hifi
    in my spare time (ha! like "spare change"?). I've eaasily got
    100+ hours into the stakeholder's specification (i.e., what
    SHE would write if she had the skills to do so). I plan
    another 100+ hours on the *engineering* specification which
    will derive from that. *BUT*, only 100 hours to code the
    thing as all of the decisions will already have been made -- it's
    just "grunt work" at that point.

    [Testing will require almost as much time as the specifications
    as everything specified has to be definitively verified/validated.
    More grunt work.]
    It will be interesting to see whether AGI gets anywhere.

    I suspect we will have a "wait-a-minute" event, before that
    point and folks will rethink the whole AI issue.

    I think we will see special purpose agents being deployed -- especially
    in more "menial" jobs (there really is no reason that "receptionists"
    and "front office staff" can't be replaced *now* at doctor offices,
    etc.).
    Likewise, the folks at the counter at fast food places are
    likely history -- they were already hanging on by a thread as
    a kiosk could do their jobs (Costco has already gone this route).

    Programming frameworks and "wizards" will get smarter -- saving
    keystrokes (for folks who learn how to coax them to perform as
    intended).

    There will likely be some code coverage and testing tools that
    emerge.

    That seems to be happening now.

    There have been lots of good tools for many (many!) years.
    But, folks don't use them as vigorously as they should.

    And, employers are usually ignorant of what they *could* put
    in place so their product just withers. Ask your favorite
    "coder" what "best practices" he observes. Then, ask him if
    he REALLY observes them?!!

    There are clearly those who think there's more to the brain than computation
    (Penrose) and those who don't.
    I think humans have the capacity for "insights" as some dendrite
    happens past another that makes an "atypical" association leading
    to an unexpected revelation. I'm not sure how much of that is
    learned vs. cultural -- you see a lot of "lazy thinkers" (doing
    just the minimum that is needed to "solve" a problem)

    I think you might be able to emulate that ability in an
    AI by deliberately injecting "randomness" in the reasoning
    process -- forcing it to evaluate "exceptions" and deliberately
    rule them out (or in).

    Time will tell.
    One potential advantage from all the AI hype is that it is
    changing expectations of what *could* be possible. If that
    gets The Powers That Be to start *thinking* and questioning
    their legacy practices, things *may* move forward, despite
    themselves!

    At the very least, it's got to be awfully embarassing to
    see AIs finding HUNDREDS of bugs in existing codebases.
    "Really? You didn't know ANY of these were present???"

    [When bugs are a dirty little secret that isn't reported and
    tracked, it is easy to make folks believe that your product
    is "good" -- or, no worse than anyone else's...]

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Theo@3:633/10 to All on Thursday, July 30, 2026 21:02:09
    Don Y <blockedofcourse@foo.invalid> wrote:
    On 7/29/2026 2:43 AM, Theo wrote:
    Don Y <blockedofcourse@foo.invalid> wrote:
    Anyone played with any of the "Open" AI's on local hardware?
    Successes? Horror stories?

    I've been doing some messing around (text processing, not LLMs, on Ubuntu 24.04). Takeaways:

    "Creative writing"? Document prep?

    OCR, sentence processing, language translation.

    1. The model is one thing, but it often exists within a stack of Python/C++/whatever to run it.

    Yes, that seems to be an understatement.

    That can be quite brittle - a huge pile of
    Python dependencies, maybe some of them don't work on your machine and you have to dig into why (Python versions, OS updates, projects been abandoned, etc).

    Have you encountered this -- or, just "expecting it" given the hassles with setting up a proper environment?

    I've encountered it.

    If you want to switch to a different model to do the same job, now
    you have to switch to their stack of dependencies (ie effectively rewrite your code, rather than just swapping out one model file for another).

    But, you should be able to move the model to different *hardware* (same flavor CPU/GPU, etc. but different MIPS/RAM)? I.e., if I wanted to see
    the effect of CPU resources on its performance -- or, the value of
    additional memory, bigger GPU, faster disk -- I should be able to do so?

    Assuming the OS is the same, you should be OK. If it's a different OS
    version or distro you get to play whac'a'mole with the dependencies.

    If the GPU is different then you get driver fun too. (even at the same
    vendor, the architectures can vary quite a bit between generations, and that can mean a different compiler/etc. Some things just aren't supported on one GPU+driver compared with another GPU+driver)

    [Does virtual memory even come into play -- or, is everything hardwired
    to physical memory?]

    Without GPU support, it's a regular CPU process like any other. If you
    don't have enough RAM then you swap. These models are small (MB) so it's
    not a problem unless you're running on a Raspberry Pi or something.

    I.e., I would (ideally) like to demonstrate:
    - how each hardware resource contributes to performance
    - how the models perform relative to each other (on "identical" tasks)

    There are benchmarks for that, but you have to put the effort into running them on your system.

    2. If you have a GPU it had better be NVIDIA. AMD and Intel GPUs effectively don't exist as far as these stacks are concerned. I have a perfectly good modern AMD GPU and all the models run on the CPU.

    All of my GPUs (and Teslas) are nVidia. I'd also like to see if the models *choke* on a "too small" GPU -- or, if they just underperform (the equivalent of thrashing).

    Most GPUs can use system DRAM over (slower, higher latency) PCIe if you
    don't have enough GDDR, so it's either slower or thrashing swapping things
    back and forth.

    I expect that you will struggle if the model is too big for the amount of
    DRAM you have: in theory the drivers could keep things 'working' by swapping GPU memory to disk, but progress will be minimal.

    Have you verified your "local results" against online services (from the
    same model(s))?

    There aren't online services with the same models AFAIAA.

    I'm not interested in signing up for accounts and token budgeting anyway.

    Any suggestions for "problems" that should be addressable by a variety of (general) models -- not overly reliant on a particular training set?

    No idea. These are models that do one job only.

    Has it been worth the effort (besides as a curiosity)?

    If you have a lot of work for it to do, it's cheaper than paying for
    cloud processing. This has cost me $0. You also get a better idea of how
    much resource it's actually consuming behind the scenes.

    Theo

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From CĂłilĂ­n NioclásĂ­n GlostĂ©ir@3:633/10 to All on Thursday, July 30, 2026 20:48:56
    Theo <theom+news@Chiark.greenEnd.org.UK> wrote: |-----------------------------------------------------------|
    |"There aren't online services with the same models AFAIAA."| |-----------------------------------------------------------|

    I did not try this:
    HTTPS://GPT.Insomnia247.NL

    (S. HTTP://Gloucester.Insomnia247.NL/ fuer Kontaktdaten!)

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Thursday, July 30, 2026 17:35:49
    On 7/30/2026 1:02 PM, Theo wrote:
    I've been doing some messing around (text processing, not LLMs, on Ubuntu >>> 24.04). Takeaways:

    "Creative writing"? Document prep?

    OCR, sentence processing, language translation.

    Is this a real need or something to play with? I.e., does having the
    AI offer you real benefit?

    If you want to switch to a different model to do the same job, now
    you have to switch to their stack of dependencies (ie effectively rewrite >>> your code, rather than just swapping out one model file for another).

    But, you should be able to move the model to different *hardware* (same
    flavor CPU/GPU, etc. but different MIPS/RAM)? I.e., if I wanted to see
    the effect of CPU resources on its performance -- or, the value of
    additional memory, bigger GPU, faster disk -- I should be able to do so?

    Assuming the OS is the same, you should be OK. If it's a different OS version or distro you get to play whac'a'mole with the dependencies.

    Insofar as possible, I will be trying to create identical environments,
    save for the hardware. The point being to identify/quantify the impact
    of the hardware on performance. (otherwise, there's too many variables
    to try to evaluate)

    I am hoping to just build an environment and then pull the disk
    and move it to another "very similar" machine (changing memory,
    MIPS, GPU in/out, etc.)

    If the GPU is different then you get driver fun too. (even at the same vendor, the architectures can vary quite a bit between generations, and that can mean a different compiler/etc. Some things just aren't supported on one GPU+driver compared with another GPU+driver)

    [Does virtual memory even come into play -- or, is everything hardwired
    to physical memory?]

    Without GPU support, it's a regular CPU process like any other. If you
    don't have enough RAM then you swap. These models are small (MB) so it's
    not a problem unless you're running on a Raspberry Pi or something.

    I've got gobs of RAM so even some of the larger LLMs weights would
    easily fit in them.
    I.e., I would (ideally) like to demonstrate:
    - how each hardware resource contributes to performance
    - how the models perform relative to each other (on "identical" tasks)

    There are benchmarks for that, but you have to put the effort into running them
    on your system.

    Yes, that's the point. If *designing* with an AI, then you want it
    to be performant as you are likely twiddling your thumbs waiting to
    critique it's latest offering.

    If, OTOH, you are using it to perform a particular *job*, you can
    walk away and return when the results are available.

    Consider the parallel to folks who "need" fast compiles because
    they're waiting to see how their latest patches perform vs. doing
    a "make world"

    2. If you have a GPU it had better be NVIDIA. AMD and Intel GPUs
    effectively don't exist as far as these stacks are concerned. I have a
    perfectly good modern AMD GPU and all the models run on the CPU.

    All of my GPUs (and Teslas) are nVidia. I'd also like to see if the models >> *choke* on a "too small" GPU -- or, if they just underperform (the equivalent
    of thrashing).

    Most GPUs can use system DRAM over (slower, higher latency) PCIe if you
    don't have enough GDDR, so it's either slower or thrashing swapping things back and forth.

    I expect that you will struggle if the model is too big for the amount of DRAM you have: in theory the drivers could keep things 'working' by swapping GPU memory to disk, but progress will be minimal.

    I have boxes with half a terabyte of RAM so assume the real problem
    will be the slower speed (and narrower pipe) that system RAM provides
    vs the GPUs access to VRAM.

    Have you verified your "local results" against online services (from the
    same model(s))?

    There aren't online services with the same models AFAIAA.

    OK.

    I'm not interested in signing up for accounts and token budgeting anyway.

    Exactly. But, having an online AI to act as a reference to validate
    the output of the local instance would be reassuring.

    Any suggestions for "problems" that should be addressable by a variety of
    (general) models -- not overly reliant on a particular training set?

    No idea. These are models that do one job only.

    This is the heart of the problem I'm facing. I don't see any real
    use for AIs *in* my workflow, unless as bug hunters. And, I expect
    little value, there, as the sources of the "tough" bugs almost always
    lie in concurrency issues. (who writes single-threaded code in
    the 21st century?)

    Edward's image creator would be entertaining and likely produce
    repeatable results (input doesn't change so why should the output?).
    But, would be hard to correlate to any real "design" effort.

    Has it been worth the effort (besides as a curiosity)?

    If you have a lot of work for it to do, it's cheaper than paying for
    cloud processing. This has cost me $0. You also get a better idea of how much resource it's actually consuming behind the scenes.
    I'd never put my IP out to train something -- any more than I
    would train an employee for a client/employer. Nor can I
    see any of my colleagues giving away the farm.

    OTOH, if there was an activity (of value!) that could be
    performed using something *locally*...

    (Some of the speech machines have been of use to me but that's
    more of a niche need)

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Friday, July 31, 2026 00:43:33
    On 7/30/2026 5:35 PM, Don Y wrote:
    Any suggestions for "problems" that should be addressable by a variety of >>> (general) models -- not overly reliant on a particular training set?

    No idea.˙ These are models that do one job only.

    This is the heart of the problem I'm facing.˙ I don't see any real
    use for AIs *in* my workflow, unless as bug hunters.˙ And, I expect
    little value, there, as the sources of the "tough" bugs almost always
    lie in concurrency issues.˙ (who writes single-threaded code in
    the 21st century?)

    Apparently, Oracle: <https://www.forbes.com/sites/daveywinder/2026/07/24/oracle-releases-record-1449-security-patches-as-companies-face-patch-overload/>

    One has to wonder what other tools could have found those faults,
    possibly BEFORE the advent of LLMs.

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Theo@3:633/10 to All on Friday, July 31, 2026 12:18:04
    Don Y <blockedofcourse@foo.invalid> wrote:
    On 7/30/2026 1:02 PM, Theo wrote:
    I've been doing some messing around (text processing, not LLMs, on Ubuntu >>> 24.04). Takeaways:

    "Creative writing"? Document prep?

    OCR, sentence processing, language translation.

    Is this a real need or something to play with? I.e., does having the
    AI offer you real benefit?

    This is for reading things in languages I don't speak. If want to read a
    book in eg German, first I need to get it into a digital form. Then I need
    to translate it, and re-apply the translations to the images (so diagrams
    make sense)

    Translation apps like Google Translate will translate one image at a time,
    but if the book has hundreds of pages that is awkward to manage. You really want a pipeline that does everything automatically. No doubt that can be
    done via the APIs of the big translation services, but that starts getting expensive in terms of API costs. Each book has a lot of pages, and there
    are a lot of books...

    Insofar as possible, I will be trying to create identical environments,
    save for the hardware. The point being to identify/quantify the impact
    of the hardware on performance. (otherwise, there's too many variables
    to try to evaluate)

    I am hoping to just build an environment and then pull the disk
    and move it to another "very similar" machine (changing memory,
    MIPS, GPU in/out, etc.)

    I expect that's going to be annoying with GPU drivers.

    Yes, that's the point. If *designing* with an AI, then you want it
    to be performant as you are likely twiddling your thumbs waiting to
    critique it's latest offering.

    If, OTOH, you are using it to perform a particular *job*, you can
    walk away and return when the results are available.

    Consider the parallel to folks who "need" fast compiles because
    they're waiting to see how their latest patches perform vs. doing
    a "make world"

    That assumes you aren't going to need multiple roundtrips. ie it's a simple mechanical process that you run exactly once. If you need to close the
    loop, ie tweak some of the inputs and repeat, then the latency matters as
    well as the throughput.

    Even if you aren't tweaking the model you may need to tweak the inputs.

    Exactly. But, having an online AI to act as a reference to validate
    the output of the local instance would be reassuring.

    I don't know how repeatable these flows are expected to be. ie can you put
    the same input into the same model run on different hardware and expect to
    get identical outputs? If you put the same prompt twice into ChatGPT will
    you get identical output? Or are there sources of divergence (either
    numerical or context)?

    For simple stuff run locally I expect more determinism (better control of context, same hardware etc) but you don't know what is being used behind an online API.

    Theo

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From CĂłilĂ­n NioclásĂ­n GlostĂ©ir@3:633/10 to All on Friday, July 31, 2026 13:32:37
    Theo <theom+news@Chiark.greenEnd.org.UK> wrote: |----------------------------------------------------------------|
    |"This is for reading things in languages I don't speak. [. . .]"| |----------------------------------------------------------------|

    O brave new World!

    Beware that software is even worse at translations than professional translations persons are!
    (S. HTTP://Gloucester.Insomnia247.NL/ fuer Kontaktdaten!)

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Don Y@3:633/10 to All on Friday, July 31, 2026 08:41:11
    On 7/31/2026 4:18 AM, Theo wrote:
    OCR, sentence processing, language translation.

    Is this a real need or something to play with? I.e., does having the
    AI offer you real benefit?

    This is for reading things in languages I don't speak. If want to read a book in eg German, first I need to get it into a digital form. Then I need to translate it, and re-apply the translations to the images (so diagrams make sense)

    I can't admit to ever having (the need to) translate an entire book.
    I rely on apps to "get me close enough" for phrases, etc. (SDL-Trados)
    with some fallback on google. I've not yet tried to hack the offline translation abilities in Firefox...

    Translation apps like Google Translate will translate one image at a time, but if the book has hundreds of pages that is awkward to manage. You really want a pipeline that does everything automatically. No doubt that can be done via the APIs of the big translation services, but that starts getting expensive in terms of API costs. Each book has a lot of pages, and there
    are a lot of books...

    So, the OCR requirement is you're starting with scans of pages...?

    Insofar as possible, I will be trying to create identical environments,
    save for the hardware. The point being to identify/quantify the impact
    of the hardware on performance. (otherwise, there's too many variables
    to try to evaluate)

    I am hoping to just build an environment and then pull the disk
    and move it to another "very similar" machine (changing memory,
    MIPS, GPU in/out, etc.)

    I expect that's going to be annoying with GPU drivers.

    I'll put the same GPUs in each machine (I have *lots* of kit).

    Yes, that's the point. If *designing* with an AI, then you want it
    to be performant as you are likely twiddling your thumbs waiting to
    critique it's latest offering.

    If, OTOH, you are using it to perform a particular *job*, you can
    walk away and return when the results are available.

    Consider the parallel to folks who "need" fast compiles because
    they're waiting to see how their latest patches perform vs. doing
    a "make world"

    That assumes you aren't going to need multiple roundtrips.

    Yes -- hence the "make world" analogy: something that inherently
    takes a long time (too long to "watch the pot boil") but is
    largely unattended, until finished. E.g., when I use symbolic
    execution, it's silly to sit and watch results trickle out; just
    start teh process and move on to something else (there is ALWAYS
    something else that needs to be done)

    ie it's a simple
    mechanical process that you run exactly once. If you need to close the
    loop, ie tweak some of the inputs and repeat, then the latency matters as well as the throughput.

    Yes. If one was using the AI as an assistant, then one would
    be concerned with the (temporal) load that it places on oneself.
    OTOH, if you can delegate a task to it and walk away, then
    you're really only concerned with the quality of its work, not the
    speed/rate.

    Even if you aren't tweaking the model you may need to tweak the inputs.

    Exactly. But, having an online AI to act as a reference to validate
    the output of the local instance would be reassuring.

    I don't know how repeatable these flows are expected to be. ie can you put the same input into the same model run on different hardware and expect to get identical outputs? If you put the same prompt twice into ChatGPT will you get identical output? Or are there sources of divergence (either numerical or context)?

    With an online tool, I suspect it learns from each interaction.
    So, it may NOT be repeatable -- another reason to take control
    of any tool you're using, "locally". E.g., when I use a VM for
    some task, I always start with a *copy* of the saved VM and
    discard it when done -- so any "use" doesn't alter it.

    For simple stuff run locally I expect more determinism (better control of context, same hardware etc) but you don't know what is being used behind an online API.
    Exactly.You should be able to repeat a "process" on a given set of "data"
    and get identical "results". Otherwise, the impact of any deliberate
    changes to the "data" can yield disproportionate changes to the "results".

    Imagine developing in an environment where your actions only *bias*
    the results instead of directly controlling them!

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)