Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > alt.comp.os.windows-11 > #35980 > unrolled thread
| Started by | MegaNZ <MegaNZ@newsgroups.com> |
|---|---|
| First post | 2026-10-09 00:29 +0100 |
| Last post | 2026-10-16 10:08 -0400 |
| Articles | 19 on this page of 59 — 15 participants |
Back to article view | Back to alt.comp.os.windows-11
Adobe is panicking MegaNZ <MegaNZ@newsgroups.com> - 2026-10-09 00:29 +0100
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-09 01:59 -0400
Re: Adobe is panicking Steve Hayes <hayesstw@telkomsa.net> - 2026-10-10 06:22 +0200
Re: Adobe is panicking WolfFan <akwolffan@zoho.com> - 2026-10-11 10:13 -0400
Re: Adobe is panicking "....winston" <winstonmvp@gmail.com> - 2026-10-09 03:21 -0400
Re: Adobe is panicking dillinger <dillinger@not.invalid> - 2026-10-09 20:02 +0200
Re: Adobe is panicking "....winston" <winstonmvp@gmail.com> - 2026-10-09 20:53 -0400
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-09 21:46 -0400
Re: Adobe is panicking "....winston" <winstonmvp@gmail.com> - 2026-10-13 13:49 -0400
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-13 16:38 -0400
Re: Adobe is panicking "....winston" <winstonmvp@gmail.com> - 2026-10-15 11:35 -0400
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-15 20:50 +0000
Re: Adobe is panicking "....winston" <winstonmvp@gmail.com> - 2026-10-16 02:04 -0400
[OT[: "24/7/365" (was: Re: Adobe is panicking) "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-10-16 08:16 +0100
Re: [OT[: "24/7/365" "Carlos E.R." <robin_listas@es.invalid> - 2026-10-16 12:20 +0200
Re: [OT[: "24/7/365" "....winston" <winstonmvp@gmail.com> - 2026-10-16 10:52 -0400
Re: [OT[: "24/7/365" "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-10-16 18:39 +0100
Re: [OT[: "24/7/365" Frank Slootweg <this@ddress.is.invalid> - 2026-10-16 17:51 +0000
Re: [OT[: "24/7/365" "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-10-16 20:44 +0100
Re: [OT[: "24/7/365" Frank Slootweg <this@ddress.is.invalid> - 2026-10-17 09:55 +0000
Re: [OT[: "24/7/365" Daniel70 <daniel47@nomail.afraid.org> - 2026-10-17 22:09 +1100
Re: [OT[: "24/7/365" "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-17 11:18 +0000
Re: [OT[: "24/7/365" Frank Slootweg <this@ddress.is.invalid> - 2026-10-17 11:45 +0000
Re: [OT[: "24/7/365" Daniel70 <daniel47@nomail.afraid.org> - 2026-10-17 23:28 +1100
Re: [OT[: "24/7/365" Frank Slootweg <this@ddress.is.invalid> - 2026-10-17 15:37 +0000
Re: [OT[: "24/7/365" "Carlos E.R." <robin_listas@es.invalid> - 2026-10-17 19:42 +0200
Re: [OT[: "24/7/365" "....winston" <winstonmvp@gmail.com> - 2026-10-16 15:48 -0400
Re: [OT[: "24/7/365" "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-10-17 01:06 +0100
Re: [OT[: "24/7/365" Daniel70 <daniel47@nomail.afraid.org> - 2026-10-17 19:59 +1100
Re: [OT[: "24/7/365" (was: Re: Adobe is panicking) Mark Lloyd <nomail@here.invalid> - 2026-10-16 15:40 +0000
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 07:42 +0000
Re: Adobe is panicking "....winston" <winstonmvp@gmail.com> - 2026-10-16 11:06 -0400
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 19:27 +0000
Re: Adobe is panicking "....winston" <winstonmvp@gmail.com> - 2026-10-16 15:40 -0400
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 19:43 +0000
Re: Adobe is panicking Lars Poulsen <lars@beagle-ears.com> - 2026-10-13 18:52 -0700
Re: Adobe is panicking Brian Gregory <void-invalid-dead-dontuse@email.invalid> - 2026-10-11 19:21 +0100
Re: Adobe is panicking Andy Burns <usenet@andyburns.uk> - 2026-10-11 22:39 +0100
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-12 00:00 -0400
Re: Adobe is panicking Brian Gregory <void-invalid-dead-dontuse@email.invalid> - 2026-10-12 14:06 +0100
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-12 18:24 -0400
Re: Adobe is panicking Brian Gregory <void-invalid-dead-dontuse@email.invalid> - 2026-10-15 23:15 +0100
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 07:53 +0000
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-16 07:29 -0400
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 11:37 +0000
Re: Adobe is panicking Andy Burns <usenet@andyburns.uk> - 2026-10-16 13:05 +0100
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 12:28 +0000
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-16 09:01 -0400
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 13:13 +0000
Re: Adobe is panicking Andy Burns <usenet@andyburns.uk> - 2026-10-16 14:23 +0100
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 13:58 +0000
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-16 15:13 -0400
Re: Adobe is panicking "Jeff Gaines" <jgnewsid@outlook.com> - 2026-10-16 19:29 +0000
Re: Adobe is panicking Andy Burns <usenet@andyburns.uk> - 2026-10-16 21:12 +0100
Re: Adobe is panicking Andy Burns <usenet@andyburns.uk> - 2026-10-16 21:14 +0100
Re: Adobe is panicking "J. P. Gilliver" <G6JPG@255soft.uk> - 2026-10-17 00:33 +0100
Re: Adobe is panicking Andy Burns <usenet@andyburns.uk> - 2026-10-17 09:26 +0100
Re: Adobe is panicking Andy Burns <usenet@andyburns.uk> - 2026-10-16 14:15 +0100
Re: Adobe is panicking Paul <nospam@needed.invalid> - 2026-10-16 10:08 -0400
Page 3 of 3 — ← Prev page 1 2 [3]
| From | Paul <nospam@needed.invalid> |
|---|---|
| Date | 2026-10-12 18:24 -0400 |
| Message-ID | <11ajmno$1n9i$1@dont-email.me> |
| In reply to | #36176 |
On Mon, 10/12/2026 9:06 AM, Brian Gregory wrote: > On 11/10/2026 22:39, Andy Burns wrote: >> Brian Gregory wrote: >> >>> Professionals who compare the two side by side say the clones fall short >>> in ways that are quite likely to be important to most professionals. >> >> This week ... once AI gets its teeth into bugfixing that could change surprisingly quickly. > > It's not so much bugs, it's that Adobe knows great ways to implement > things like object masking. If the AI doesn't know, and if it's not > actually reverse engineering it probably won't know, then it can't > implement it so well. In this example, a large percentage of the code is byte-for byte identical to the original. Only there is now source code, to make those bytes. https://www.tomshardware.com/tech-industry/artificial-intelligence/dev-used-ai-agents-and-more-than-500-billion-tokens-to-rebuild-call-of-duty-modern-warfare-2s-source-code-83-percent-of-the-2009-shooters-functions-now-match-the-original-byte-for-byte Once you have the source code to make the bytes of object, you can have the LLM-AI write a summary of what algorithm(s) are involved. Then work in the forward direction again, to make a FOSS-inspired executable that does not match the source bytes. If I'm Adobe, now I'm worried. It would be pretty expensive to do it this way, but then, the results would also be pretty impressive. Paul
[toc] | [prev] | [next] | [standalone]
| From | Brian Gregory <void-invalid-dead-dontuse@email.invalid> |
|---|---|
| Date | 2026-10-15 23:15 +0100 |
| Message-ID | <njd1kgFktb1U2@mid.individual.net> |
| In reply to | #36196 |
On 12/10/2026 23:24, Paul wrote: > If I'm Adobe, now I'm worried. It would be pretty expensive to do it > this way, but then, the results would also be pretty impressive. Fair point. Maybe laws will be changed and/or code will soon be obfuscated or hidden in secure enclaves. However look how many companies still use Microsoft Office when several very capable compatible alternatives exist. -- Brian Gregory (in England).
[toc] | [prev] | [next] | [standalone]
| From | "Jeff Gaines" <jgnewsid@outlook.com> |
|---|---|
| Date | 2026-10-16 07:53 +0000 |
| Message-ID | <xn0pw2zzkhl3fer01m@news.individual.net> |
| In reply to | #36339 |
On 15/10/2026 in message <njd1kgFktb1U2@mid.individual.net> Brian Gregory wrote: >On 12/10/2026 23:24, Paul wrote: >>If I'm Adobe, now I'm worried. It would be pretty expensive to do it >>this way, but then, the results would also be pretty impressive. > >Fair point. > >Maybe laws will be changed and/or code will soon be obfuscated or hidden >in secure enclaves. > >However look how many companies still use Microsoft Office when several >very capable compatible alternatives exist. I had an interesting discussion with Google's AI recently in connection with programming for Windows. It was about the ListView control included with Visual Studio which is one of the most complained about controls because you can't control the height of the Header or ListViewItems and it flickers every time you breathe. I have written my own based on the Microsoft version and had fixed the row height and flickering but not the Header height. AI suggested a way of doing it which worked once but it reverted if a column width was changed. I followed up the next day and the AI bot came up with the cure which was intercepting one specific Windows Message and manipulating the result sent back to windows every time it was raised, not just the first time. It works perfectly but raises questions. Any programmer who uses this control will know its limitations and the work rounds but the suggestion from AI has never come up in the 21 years that Visual Studio has been around. Has AI ferreted it out from one of the software companies that provides a better ListView control and released proprietary code into the wild? -- Jeff Gaines Dorset UK Here we go it's getting close, now it's just who wants it most.
[toc] | [prev] | [next] | [standalone]
| From | Paul <nospam@needed.invalid> |
|---|---|
| Date | 2026-10-16 07:29 -0400 |
| Message-ID | <11at1rj$38b1f$1@dont-email.me> |
| In reply to | #36357 |
On Fri, 10/16/2026 3:53 AM, Jeff Gaines wrote: > On 15/10/2026 in message <njd1kgFktb1U2@mid.individual.net> Brian Gregory wrote: > >> On 12/10/2026 23:24, Paul wrote: >>> If I'm Adobe, now I'm worried. It would be pretty expensive to do it >>> this way, but then, the results would also be pretty impressive. >> >> Fair point. >> >> Maybe laws will be changed and/or code will soon be obfuscated or hidden in secure enclaves. >> >> However look how many companies still use Microsoft Office when several very capable compatible alternatives exist. > > I had an interesting discussion with Google's AI recently in connection with programming for Windows. It was about the ListView control included with Visual Studio which is one of the most complained about controls because you can't control the height of the Header or ListViewItems and it flickers every time you breathe. > > I have written my own based on the Microsoft version and had fixed the row height and flickering but not the Header height. AI suggested a way of doing it which worked once but it reverted if a column width was changed. I followed up the next day and the AI bot came up with the cure which was intercepting one specific Windows Message and manipulating the result sent back to windows every time it was raised, not just the first time. > > It works perfectly but raises questions. Any programmer who uses this control will know its limitations and the work rounds but the suggestion from AI has never come up in the 21 years that Visual Studio has been around. Has AI ferreted it out from one of the software companies that provides a better ListView control and released proprietary code into the wild? > It just means it was discussed in a place your search engine never happened to visit. The "base" training set, is 420TB of data. And actual data sets today, are multiple times that big. Have you read 420TB of text since you were born ? I suspect not. When the LLM-AI has not seen a named function, it can't write the code. As a test of local LLM-AI, I asked each one to write me a program to calculate PI to one million digits. The LLM-AI could tell me what the "best" algorithm is today, but there is an addenum to that called "binary splitting". It could not write that code, and it made excuses about the level of complexity as the reason it would not write the code. Well, that code had actually been written in the last three months, by a human, and was in Github. I didn't research this at first (I don't really need a PI program, it was just a test case), but a human made an awesome PI calculating program (it's 10x to 20x faster than SuperPI!). That had the binary splitting function, and it was 23 lines of code and it is recursive. It really wasn't all that complicated. The code writing exercise might have run for around 50 minutes on the Big Machine, before coming back with the refusal to finish that module. They can't really think outside of the box, not all of them. Occasionally Claude can emit a single English sentence, that makes you wonder what capabilities it has. But some of the others are relative poor students, and they have trouble escaping from a wet paper bag. I suppose it is normal for the less capable ones to be released as open weight models. I'm not an expert at this, just a user. And like many people here, I run test cases to "sense" what the future will bring. Are bigger and bigger models the answer ? What is the "utility" versus "risk" ratio for various things ? Can we afford to be carefree and reckless, forever ? The history of computing has already shown us what a rising "risk" ratio can do. Paul
[toc] | [prev] | [next] | [standalone]
| From | "Jeff Gaines" <jgnewsid@outlook.com> |
|---|---|
| Date | 2026-10-16 11:37 +0000 |
| Message-ID | <xn0pw35wnht32ea01o@news.individual.net> |
| In reply to | #36361 |
On 16/10/2026 in message <11at1rj$38b1f$1@dont-email.me> Paul wrote: >>I had an interesting discussion with Google's AI recently in connection >>with programming for Windows. It was about the ListView control included >>with Visual Studio which is one of the most complained about controls >>because you can't control the height of the Header or ListViewItems and it >>flickers every time you breathe. >> >>I have written my own based on the Microsoft version and had fixed the row >>height and flickering but not the Header height. AI suggested a way of >>doing it which worked once but it reverted if a column width was changed. >>I followed up the next day and the AI bot came up with the cure which was >>intercepting one specific Windows Message and manipulating the result sent >>back to windows every time it was raised, not just the first time. >> >>It works perfectly but raises questions. Any programmer who uses this >>control will know its limitations and the work rounds but the suggestion >>from AI has never come up in the 21 years that Visual Studio has been >>around. Has AI ferreted it out from one of the software companies that >>provides a better ListView control and released proprietary code into the >>wild? >> > >It just means it was discussed in a place your search engine never >happened to visit. >The "base" training set, is 420TB of data. And actual data sets today, are >multiple >times that big. Have you read 420TB of text since you were born ? I >suspect not. It's not just me of course, I don't know how many people are writing code and want a decent LsitView but there are thousands of comments and suggestions all over the Internet dating back to 2005. -- Jeff Gaines Dorset UK All things being equal, fat people use more soap
[toc] | [prev] | [next] | [standalone]
| From | Andy Burns <usenet@andyburns.uk> |
|---|---|
| Date | 2026-10-16 13:05 +0100 |
| Message-ID | <njei7rFsdhsU1@mid.individual.net> |
| In reply to | #36362 |
Jeff Gaines wrote: > I don't know how many people are writing code and want a decent > LsitView but there are thousands of comments and suggestions all over > the Internet dating back to 2005. The thousands of comments and suggestions aren't relevant, but they are what search engines will tend to give you; the one correct answer is what counts and AI has absorbed that information ... if you ask it the same question again, does it give you any context to where it found it?
[toc] | [prev] | [next] | [standalone]
| From | "Jeff Gaines" <jgnewsid@outlook.com> |
|---|---|
| Date | 2026-10-16 12:28 +0000 |
| Message-ID | <xn0pw379lhuwq5r01q@news.individual.net> |
| In reply to | #36365 |
On 16/10/2026 in message <njei7rFsdhsU1@mid.individual.net> Andy Burns wrote: > >Jeff Gaines wrote: > >>I don't know how many people are writing code and want a decent >>LsitView but there are thousands of comments and suggestions all over >>the Internet dating back to 2005. > >The thousands of comments and suggestions aren't relevant, but they are >what search engines will tend to give you; the one correct answer is what >counts and AI has absorbed that information ... if you ask it the same >question again, does it give you any context to where it found it? The one correct answer must be among those comments and suggestions then, although neither I nor anybody else has found it or it would be common knowledge. -- Jeff Gaines Dorset UK Though no-one can go back and make a new start, everyone can start from now and make a new ending.
[toc] | [prev] | [next] | [standalone]
| From | Paul <nospam@needed.invalid> |
|---|---|
| Date | 2026-10-16 09:01 -0400 |
| Message-ID | <11at77b$3abni$1@dont-email.me> |
| In reply to | #36367 |
On Fri, 10/16/2026 8:28 AM, Jeff Gaines wrote: > On 16/10/2026 in message <njei7rFsdhsU1@mid.individual.net> Andy Burns wrote: > >> >> Jeff Gaines wrote: >> >>> I don't know how many people are writing code and want a decent >>> LsitView but there are thousands of comments and suggestions all over >>> the Internet dating back to 2005. >> >> The thousands of comments and suggestions aren't relevant, but they are what search engines will tend to give you; the one correct answer is what counts and AI has absorbed that information ... if you ask it the same question again, does it give you any context to where it found it? > > The one correct answer must be among those comments and suggestions then, although neither I nor anybody else has found it or it would be common knowledge. > Make it your mission to find it. Ask Gemini or similar, to answer the question, plus give some indication what kind of forum contains this sort of knowledge. Since you know the answer already, you can feed in all the keywords from the answer, and get what you want. The thing about "cites" on LLM-AI, is they are damn lazy. Sometimes, they cite their asses off. Other times, they will find one lame/derivative web URL, and when you examine it, you discover it is actually written by an LLM-AI, and only has summary statements (not human paragraphs with reasoning attached to ideas). In some cases, the web link has nothing to do with the topic! That's not actually that unusual. A URL cite is different from the training set, as for example, a chopped up and scanned copy of Chaucer does not have a URL. It won't be particularly easy to get cites, but the reason you would be doing this experiment, is to determine whether you can "see" the material as being possible in the training set. Like, everything that was in Reddit, should be in a training set. Even though Reddit isn't actually a very strong site for technical info. Lots of bogus threads. Also, when you ask these questions, it helps if the hosting interface to the LLM-AI (whether remote or local) has a "High Reasoning" setting. Using High Reasoning, the LLM-AI can tell you "I would be just guessing if..." when previously with a snap answer, it does not give-a-shit where the answer came from. If you've run High Reasoning, the LLM-AI "chatters to itself" and it detects convergence while doing so. If an answer converges in two runs, say, that's a "strong" response and the machine knows that the answer it is about to give, is based on multiple occurrences in the training set. When the session takes a little bit longer, it may come back with, instead of "I don't know", it words it "I would be just guessing if..." as an indication that the info is not really in the training set. When I ran High Reasoning locally, that phase took 8 minutes, while the token output took 42 minutes. Whereas for the most simple query your can make, a flash answer "thinks" for only 2 seconds (on my Big Machine computer) and then the token output takes forever. If an MoE fits within the VRAM on a video card, they tell me the output tokens can be produced by the video card itself. If the model does not happen to fit, some amount of thinking can be done on the video card, but the output tokens are done via matrix multiply using CPU cores. Which is, well, dog slow. That's why the runtime for queries on my machine are so long. It only has a 4GB video card. It really needs a 48GB video card. Intel just came out with a card for Inference, which is 480GB and LPDDR5x (relatively slow) memory, no HBM. Should be $15K-$20K or so, so out of reach really. It would be tons better, but still a joke compared to the machines we get to use on the Internet. Paul
[toc] | [prev] | [next] | [standalone]
| From | "Jeff Gaines" <jgnewsid@outlook.com> |
|---|---|
| Date | 2026-10-16 13:13 +0000 |
| Message-ID | <xn0pw38hnhwjdlx01r@news.individual.net> |
| In reply to | #36368 |
On 16/10/2026 in message <11at77b$3abni$1@dont-email.me> Paul wrote: >On Fri, 10/16/2026 8:28 AM, Jeff Gaines wrote: >>On 16/10/2026 in message <njei7rFsdhsU1@mid.individual.net> Andy Burns >>wrote: >> >>> >>>Jeff Gaines wrote: >>> >>>>I don't know how many people are writing code and want a decent >>>>LsitView but there are thousands of comments and suggestions all over >>>>the Internet dating back to 2005. >>> >>>The thousands of comments and suggestions aren't relevant, but they are >>>what search engines will tend to give you; the one correct answer is >>>what counts and AI has absorbed that information ... if you ask it the >>>same question again, does it give you any context to where it found it? >> >>The one correct answer must be among those comments and suggestions then, >>although neither I nor anybody else has found it or it would be common >>knowledge. >> > >Make it your mission to find it. No chance! I have been looking for it since 2005 and it is one of the most common programming questions that is asked, frankly if it's not on StackOverflow it doesn't exist in the public domain. -- Jeff Gaines Dorset UK The true meaning of life is to plant trees under whose shade you do not expect to sit.
[toc] | [prev] | [next] | [standalone]
| From | Andy Burns <usenet@andyburns.uk> |
|---|---|
| Date | 2026-10-16 14:23 +0100 |
| Message-ID | <njemqbFt5osU1@mid.individual.net> |
| In reply to | #36370 |
Jeff Gaines wrote: > Paul wrote: > >> Make it your mission to find it. > > No chance! I have been looking for it since 2005 and it is one of the > most common programming questions that is asked, frankly if it's not on > StackOverflow it doesn't exist in the public domain. Yes, but you've been asking search engines before, now you've turned to AI, I thought you liked being nice to them :-P -- 23467009
[toc] | [prev] | [next] | [standalone]
| From | "Jeff Gaines" <jgnewsid@outlook.com> |
|---|---|
| Date | 2026-10-16 13:58 +0000 |
| Message-ID | <xn0pw39lehy4uhy01s@news.individual.net> |
| In reply to | #36373 |
On 16/10/2026 in message <njemqbFt5osU1@mid.individual.net> Andy Burns wrote: >Jeff Gaines wrote: > >>Paul wrote: >> >>>Make it your mission to find it. >> >>No chance! I have been looking for it since 2005 and it is one of the >>most common programming questions that is asked, frankly if it's not on >>StackOverflow it doesn't exist in the public domain. > > >Yes, but you've been asking search engines before, now you've turned to >AI, I thought you liked being nice to them :-P I am, Skynet is real! -- Jeff Gaines Dorset UK The only thing Flat Earthers fear is sphere itself.
[toc] | [prev] | [next] | [standalone]
| From | Paul <nospam@needed.invalid> |
|---|---|
| Date | 2026-10-16 15:13 -0400 |
| Message-ID | <11att13$3j2i7$1@dont-email.me> |
| In reply to | #36375 |
On Fri, 10/16/2026 9:58 AM, Jeff Gaines wrote: > On 16/10/2026 in message <njemqbFt5osU1@mid.individual.net> Andy Burns wrote: > >> Jeff Gaines wrote: >> >>> Paul wrote: >>> >>>> Make it your mission to find it. >>> >>> No chance! I have been looking for it since 2005 and it is one of the most common programming questions that is asked, frankly if it's not on StackOverflow it doesn't exist in the public domain. >> >> >> Yes, but you've been asking search engines before, now you've turned to AI, I thought you liked being nice to them :-P > > I am, Skynet is real! > OK, can you do this ? Paste the answer in Harmony format (which is the Markdown it uses) into a posting, for the answer you got where it was revealed what the important/mystery bits were given to you. Place a <=== next to the bits that arrived from the unknown-type source. I keep some of my runs, question and answer, in separate text files, for reference later. Not all the runs though, usually just the ones with programs in them. Our best learning, comes from "audience challenges", where you get to see a query made from a different perspective and how the machine behaved. That's how we poke the corners and edges of the machine, individual human queries are great for that, because people will try the strangest things. Like Micky asking a question about a Time Life book (short essays), where the machine kept giving different answers for what characters were being written about. That was my first chance to use High Reasoning to get the machine to admit the info was not in the training material and it was just making shit up. It's hard to understand how it scraped the title of the book and the concept the book had short essays in it, yet it did not manage to record what the topic of the essays was (Albert Einstein or Soupy Sales). The first shared challenge here, was the "list the US presidents", where Arlen asked the LLM-AI to sort the answer a particular way. And the machine kept dropping one of the Presidents names. Not a controversial figure, more of a relatively obscure one, and the machine did not managed to answer the question properly until in frustration I added "And don't lose any Presidents!" to the query. It doesn't make a lot of sense, why that caused the correct answer to come out of the machine. Other people had also been doing that test on the Internet, and observed the same behavior, one individual kept being dropped from the list. That's really about the worst case scenario you can imagine for an LLM-AI, where the query is not challenging, it's definitely in the training set, and it got fumbled. It means we can't trust them to do anything right. Paul
[toc] | [prev] | [next] | [standalone]
| From | "Jeff Gaines" <jgnewsid@outlook.com> |
|---|---|
| Date | 2026-10-16 19:29 +0000 |
| Message-ID | <xn0pw3icwi9yfr201v@news.individual.net> |
| In reply to | #36392 |
On 16/10/2026 in message <11att13$3j2i7$1@dont-email.me> Paul wrote: >On Fri, 10/16/2026 9:58 AM, Jeff Gaines wrote: >>On 16/10/2026 in message <njemqbFt5osU1@mid.individual.net> Andy Burns >>wrote: >> >>>Jeff Gaines wrote: >>> >>>>Paul wrote: >>>> >>>>>Make it your mission to find it. >>>> >>>>No chance! I have been looking for it since 2005 and it is one of the >>>>most common programming questions that is asked, frankly if it's not on >>>>StackOverflow it doesn't exist in the public domain. >>> >>> >>>Yes, but you've been asking search engines before, now you've turned to >>>AI, I thought you liked being nice to them :-P >> >>I am, Skynet is real! >> > >OK, can you do this ? > >Paste the answer in Harmony format (which is the Markdown it uses) >into a posting, for the answer you got where it was revealed what >the important/mystery bits were given to you. Place a <=== >next to the bits that arrived from the unknown-type source. > >I keep some of my runs, question and answer, in separate text files, for >reference later. Not all the runs though, usually just the >ones with programs in them. > >Our best learning, comes from "audience challenges", where you >get to see a query made from a different perspective and how >the machine behaved. That's how we poke the corners and edges >of the machine, individual human queries are great for that, >because people will try the strangest things. Like Micky asking >a question about a Time Life book (short essays), where the machine >kept giving different answers for what characters were being >written about. That was my first chance to use High Reasoning >to get the machine to admit the info was not in the training >material and it was just making shit up. It's hard to understand >how it scraped the title of the book and the concept the book had >short essays in it, yet it did not manage to record what >the topic of the essays was (Albert Einstein or Soupy Sales). > >The first shared challenge here, was the "list the US presidents", >where Arlen asked the LLM-AI to sort the answer a particular way. >And the machine kept dropping one of the Presidents names. Not >a controversial figure, more of a relatively obscure one, and the >machine did not managed to answer the question properly until >in frustration I added "And don't lose any Presidents!" to the >query. It doesn't make a lot of sense, why that caused the correct >answer to come out of the machine. Other people had also been doing >that test on the Internet, and observed the same behavior, one >individual kept being dropped from the list. That's really about >the worst case scenario you can imagine for an LLM-AI, where the >query is not challenging, it's definitely in the training set, >and it got fumbled. It means we can't trust them to do anything right. > > Paul No, I can't. I copied and pasted it into my notepad then into my code. My point is if that answer was in the public domain then during the last 21 years of seeking the information all over the Internet it would have come to light and it didn't. -- Jeff Gaines Dorset UK It may be that your sole purpose in life is to serve as a warning to others.
[toc] | [prev] | [next] | [standalone]
| From | Andy Burns <usenet@andyburns.uk> |
|---|---|
| Date | 2026-10-16 21:12 +0100 |
| Message-ID | <njfeq6F2fh3U1@mid.individual.net> |
| In reply to | #36394 |
Jeff Gaines wrote: > I copied and pasted it into my notepad then into my code. My point is if > that answer was in the public domain then during the last 21 years of > seeking the information all over the Internet it would have come to > light and it didn't. But you're apparently unwilling to paste it here (just as plain-text) for us to look at?
[toc] | [prev] | [next] | [standalone]
| From | Andy Burns <usenet@andyburns.uk> |
|---|---|
| Date | 2026-10-16 21:14 +0100 |
| Message-ID | <njfethF2fh3U2@mid.individual.net> |
| In reply to | #36400 |
Andy Burns wrote: > But you're apparently unwilling to paste it here (just as plain-text) > for us to look at? Sorry - missed your other reply.
[toc] | [prev] | [next] | [standalone]
| From | "J. P. Gilliver" <G6JPG@255soft.uk> |
|---|---|
| Date | 2026-10-17 00:33 +0100 |
| Message-ID | <11auc7n$3oa9q$1@dont-email.me> |
| In reply to | #36373 |
On 2026/10/16 14:23:23, Andy Burns wrote: > Jeff Gaines wrote: > >> Paul wrote: >> >>> Make it your mission to find it. >> >> No chance! I have been looking for it since 2005 and it is one of the >> most common programming questions that is asked, frankly if it's not on >> StackOverflow it doesn't exist in the public domain. > > > Yes, but you've been asking search engines before, now you've turned to > AI, I thought you liked being nice to them :-P > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > 23467009 Sorry, couldn't help wondering - did you lean on the enter key before posting? (And that "--" line didn't have a space - was it supposed to? If so, what's with the number signature?) -- 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]
| From | Andy Burns <usenet@andyburns.uk> |
|---|---|
| Date | 2026-10-17 09:26 +0100 |
| Message-ID | <njgpqeF917fU1@mid.individual.net> |
| In reply to | #36406 |
"J. P. Gilliver" wrote: > Sorry, couldn't help wondering - did you lean on the enter key before > posting? (And that "--" line didn't have a space - was it supposed to? > If so, what's with the number signature?) It was a prospective way of saying "I told you so" that didn't work out.
[toc] | [prev] | [next] | [standalone]
| From | Andy Burns <usenet@andyburns.uk> |
|---|---|
| Date | 2026-10-16 14:15 +0100 |
| Message-ID | <njemc0Ft2g7U1@mid.individual.net> |
| In reply to | #36368 |
Paul wrote: > If an MoE fits within the VRAM on a video card, they tell me the output > tokens can be produced by the video card itself. I've understood that, but also that MoE (mixture of experts) don't need the entire model in VRAM all the time, they can page in different sections one after the other, but then the bottleneck isn't amount of memory, but memory bandwidth (i.e number of PCIe lanes and PCI generation). > If the model does not happen to fit, some amount of thinking can be > done on the video card, but the output tokens are done via matrix > multiply using CPU cores. Which is, well, dog slow. That's why the > runtime for queries on my machine are so long. It only has a 4GB > video card. I almost got sucked into buying a used nVidia data centre card (P40 w/24GB) the other week, they're kind-of trailing edge for what's still semi-useful, about £250 plus a new PSU and 3D print your own fan shroud and put up with the noise and electricity use while it's in use. > It really needs a 48GB video card. Intel just came out with a card > for Inference, which is 480GB and LPDDR5x (relatively slow) memory, > no HBM. Should be $15K-$20K or so, so out of reach really. It would > be tons better, but still a joke compared to the machines we get to > use on the Internet. Yep, it all zooms out of reach over the horizon really quickly ...
[toc] | [prev] | [next] | [standalone]
| From | Paul <nospam@needed.invalid> |
|---|---|
| Date | 2026-10-16 10:08 -0400 |
| Message-ID | <11atb5p$3bsk4$1@dont-email.me> |
| In reply to | #36367 |
On Fri, 10/16/2026 8:28 AM, Jeff Gaines wrote: > On 16/10/2026 in message <njei7rFsdhsU1@mid.individual.net> Andy Burns wrote: > >> >> Jeff Gaines wrote: >> >>> I don't know how many people are writing code and want a decent >>> LsitView but there are thousands of comments and suggestions all over >>> the Internet dating back to 2005. >> >> The thousands of comments and suggestions aren't relevant, but they are what search engines will tend to give you; the one correct answer is what counts and AI has absorbed that information ... if you ask it the same question again, does it give you any context to where it found it? > > The one correct answer must be among those comments and suggestions then, although neither I nor anybody else has found it or it would be common knowledge. > You know the terminology you used to get the answer in the first place. That also positions you well, to do the experiments to figure out where the answer came from. I don't know enough about ListView here, to shadow your project. I'm a copy/paste programmer :-) Paul
[toc] | [prev] | [standalone]
Page 3 of 3 — ← Prev page 1 2 [3]
Back to top | Article view | alt.comp.os.windows-11
csiph-web