News AI’s Dreadful December: Lawsuits, plagiarism and child abuse images show the perils of training on data taken without consent.

[Admin]

Three major news headlines in the past week have soured the generative AI party: a lawsuit from the NY Times, a new image model that copies frames from movies and the discovery of child abuse images in a popular dataset.

AI’s Dreadful December: Lawsuits, plagiarism and child abuse images show the perils of training on data taken without consent. : Read more

 

[setx]

AI articles look to be the only ones where Tom's put effort into.

How about remembering that you are hardware site and put the same effort in hardware reviews? Like, with with argumentation, linking to other sites research etc. what you are doing here?

 

[Jeremy Kaplan]

It's hard to ignore the biggest topic in technology, @setx! But I hear you. In the last 3 months, we've hired half a dozen new news writer and really doubled down on our coverage of the semiconductor space. We've also published for the first time ever our RAM benchmark hierarchy and SSD benchmark hierarchy. What specifically would you like to see more of?

 

[setx]

What specifically would you like to see more of?

I want to see true, authentic articles about hardware like this one: Hacking Your Mouse To Fix The Misclick Of Doom

Not the ads masquerading as articles that Tom is full of now. Like "best (for us) amazon links".

Let's take your RAM benchmark hierarchy and look at DDR5. First of all: why there are separate lists for AMD and Intel? Are some parameters more important for former or later? Why AMD list is only for 2x16GB and tops at DDR5-6200 (while I definitely know there are people running DDR5-8000 there)? Did you have any problems with other capacities or speed? The "score" was measured with which CPU? Is there a difference for high-end/mid/low CPUs? Is there difference with different microcode versions (there is for sure for AMD)? There are way more questions than answers after looking at that list. Seriously, that "article" isn't useful at all.

 

[parkerthon]

I agree to an extent. I enjoy reading about AI which is no doubt why it’s a popular choice for content, but news summarizing news about how the AI industry is fairing is something I don’t come to Toms to read about. Someone needs to cover computer hardware more even if it’s niche. AI is, after all, heavily dependent on hardware even if it is all built for a microsoft/amazon/google data center.

 

[apiltch]

After having worked on the

I agree to an extent. I enjoy reading about AI which is no doubt why it’s a popular choice for content, but news summarizing news about how the AI industry is fairing is something I don’t come to Toms to read about. Someone needs to cover computer hardware more even if it’s niche. AI is, after all, heavily dependent on hardware even if it is all built for a microsoft/amazon/google data center.

We have covered and continue to cover AI hardware. Please see the AI benchmarking story I did on various laptops last week:

https://www.tomshardware.com/laptop...-workloads-and-amds-chips-sometimes-beat-them

I wouldn't call the current article a news summary as much as it's an op-ed on the problems with allowing software vendors to build their platforms on massive theft of copyrighted IP. This has serious implications for the future of computing.

 

[George³]

wouldn't call the current article a news summary as much as it's an op-ed on the problems with allowing software vendors to build their platforms on massive theft of copyrighted IP.

Journalism copyright is under questions. What you saying IP? Unless the journalists and owners of the NY Tymes claim that they organized the events they cover.

 

[coromonadalix]

anything published on the web now serve as a learning database

for sure Ai gurus will act as : now it's online, you have no control of it anymore, like facecrap and others

good luck putting strong regulations over that

 

[tamalero]

They were warned ages ago about this.
And its going to be worse. We're entering the age of POST TRUTH.
Where truth will be almost impossible to distinguish itself from fabricated AI bull.
Everyone can be targeted, specially those with the money and means.
Like the surge of AI CP and deep fake sex videos.

 

[rabbit4me2]

Goes to show you just like typical American behavior you can't have your cake and eat it too.

 

[Alvar "Miles" Udell]

Getting around a paywall is different than reproducing the open web (open web defined as accessible without paying or requiring certain credentials). If I go to a certain article from TomsHardware, for example, I'm using an adblocker so I don't see ads, I use NoScript so dotmetrics, google-analytics, jwplatform, onesignal, parsley, permutive, and scorecardresearch are all blocked. If I ask Bing Copilot to reproduce the same article, I'm still not seeing any ads, I'm not triggering any of the other things, but I lack any images or charts which may be in the article. In both cases the website gets no monetary compensation from my visit. Compare this to a non-open website, even if ads and other content is blocked, they still get the paywall fee vs AI providing it to me for free and the website getting nothing.

As for art and video, artists and creators for decades have been taking (sometimes not insubstantial amounts of) money from people to produce content, often pornographic, of copyrighted property without permission or a license, and often with the use of paid software, but I haven't seen studios and copyright owners sue the creators for taking $xxx to create a pornographic picture of (copyrighted character). AI is just allowing anyone to create content and not just professionals.

 

[hannibal]

Stealing content is still stealing even if you or original author don´t ask money for it!

And a good article. AI definitely belongs to tech related news...

 

[rluker5]

They were warned ages ago about this.
And its going to be worse. We're entering the age of POST TRUST.
Where truth will be almost impossible to distinguish itself from fabricated AI bull.
Everyone can be targeted, specially those with the money and means.
Like the surge of AI CP and deep fake sex videos.

There needs to be an AI generated halloween costume.
Like some messed up hand gloves and random misshapements.

You would definitely need to label your kid as such though. Or else some might think they are trying to make fun of some of the handicapped.

 

[thestryker]

There are an awful lot of people who don't see any issue with what amounts to as mass theft. It seems awfully clear cut, but who knows how the legality will play out. This seems to be status quo for a lot of modern tech: roll in do whatever you want without consequences and then bully your way out.

As for art and video, artists and creators for decades have been taking (sometimes not insubstantial amounts of) money from people to produce content, often pornographic, of copyrighted property without permission or a license, and often with the use of paid software, but I haven't seen studios and copyright owners sue the creators for taking $xxx to create a pornographic picture of (copyrighted character). AI is just allowing anyone to create content and not just professionals.

None of the creators doing this are mass producing these pieces for sale which is why they don't get targeted.

Let's take your RAM benchmark hierarchy and look at DDR5. First of all: why there are separate lists for AMD and Intel?

Why there are separate lists should be really obvious: they're completely different platforms from different manufacturers.

Why AMD list is only for 2x16GB and tops at DDR5-6200 (while I definitely know there are people running DDR5-8000 there)?

I'm not sure you actually read the piece you're criticizing:

AMD's latest AGESA 1.0.0.7c firmware brought many improvements. It has upped the Ryzen processor's ability to run DDR5-6400 at 1:1, up from the previous DDR5-6000. The updated AGESA code also supports high-frequency DDR5 memory up to DDR5-8000 and potentially beyond. The only downside is that it drops the UCLK and MCLK ratio to 2:1 (UCLK: MCLK / 2), imposing a performance penalty.

I'm not sure that they've reviewed any 6400 or lower kits since that AGESA release (Aug 2023). I couldn't find reviews for some of the higher capacity kits listed and some predated Zen 4's release.

The "score" was measured with which CPU? Is there difference with different microcode versions (there is for sure for AMD)? There are way more questions than answers after looking at that list. Seriously, that "article" isn't useful at all.

All of the system specifications including firmware are listed in the article as well.

Is there a difference for high-end/mid/low CPUs?

This isn't specified because it's a ranking article, but that would make for an interesting piece.

Intel non-K SKUs won't be able to run at the same memory clocks as K SKUs which would be a useful piece of information to have in the hierarchy piece though.

 

[setx]

Why there are separate lists should be really obvious: they're completely different platforms from different manufacturers.

And why does it matter for performance of memory kits? I'd like to have a useful recommendation like "hey, notice that those kits underperform on AMD or Intel, and those perform nicely regardless of platform", not just "here is one number and link to amazon".

I'm not sure you actually read the piece you're criticizing

I'm not sure you understand what I'm criticizing: yes, they copy-pasted the public information about AGESA 1.0.0.7c, but that has zero relation to the charts shown. No even mention that published scores changed significantly/insignificantly or that some version had some problems with their kits that was resolved in another.

All of the system specifications including firmware are listed in the article as well.

I've missed that. But actually that brings even more questions:
Why did they use for AMD firmware version that is not publicly available? Why test Intel on top CPU and AMD on low-mid (and, especially, one-CCD what can change results for memory)?

 

[thestryker]

And why does it matter for performance of memory kits? I'd like to have a useful recommendation like "hey, notice that those kits underperform on AMD or Intel, and those perform nicely regardless of platform", not just "here is one number and link to amazon".

It's not a matter of performance, but rather whether or not they work. AMD can only run up to 6400 (and it was 6000 until Aug 2023) in a 1:1 configuration. Every kit they've tested since Zen 4 came out that was 6000 or less should be showing up for both AMD and Intel.

I'm not sure you understand what I'm criticizing: yes, they copy-pasted the public information about AGESA 1.0.0.7c, but that has zero relation to the charts shown. No even mention that published scores changed significantly/insignificantly or that some version had some problems with their kits that was resolved in another.

I snipped the part where you were asking about DRAM speed and pasted the part from the article that explains it. Until Aug 2023 you couldn't really run past 6000 without manually tweaking settings and it was extremely inconsistent. After Aug 2023 you can, but inflict penalties on performance which can negate the bandwidth benefit and then some.

Now do I think it'd be worthwhile to do a Zen 4 memory scaling article? Absolutely, because then people could reference that as to whether or not they wanted to deal with the penalties. I haven't seen any publication do one since AMD unlocked the capability so it would be doubly useful.

Do I think this should be in a roundup style piece which gives an overview with limited information? No.

A good addition to the hierarchy page would be test dates to provide clarity.

I've missed that. But actually that brings even more questions:
Why did they use for AMD firmware version that is not publicly available?

No clue on this one, but that was likely the latest available at the time of testing which would likely be a pre-1.0.0.7c version.

Why test Intel on top CPU and AMD on low-mid (and, especially, one-CCD what can change results for memory)?

There could be any number of reasons behind the choices and could be as simple as that's what they had available when setting up the memory benches. Of course picking from the CPUs they have with the best memory controller could also be the case.

Zen 4 IMC is located in the IO die so there isn't any significant difference between any SKUs with regards to memory (G SKUs this will be different). Intel locks out SA voltage on any non-K SKU parts which directly impacts memory clocks so they are limited in choice here.

 

[setx]

It's not a matter of performance, but rather whether or not they work. AMD can only run up to 6400 (and it was 6000 until Aug 2023) in a 1:1 configuration.

Does it matter that kit is run in 1:1 or not if there is performance benefit? Do those high-freq kits run in "1:1" on Intel?

Every kit they've tested since Zen 4 came out that was 6000 or less should be showing up for both AMD and Intel.

Uhm, no? There is 2x32GB DDR5-6000 kit for Intel but not for AMD.

Zen 4 IMC is located in the IO die so there isn't any significant difference between any SKUs with regards to memory

There is very significant difference between 1-CCD and 2-CCD models. There is information that non-server 1-CCD doesn't have enough inter-CCD bandwidth to fully saturate memory controller. Can't say is that significant or not (and that's one of the reasons I would like to see proper article).

 

[thestryker]

Does it matter that kit is run in 1:1 or not if there is performance benefit?

Yes it does matter when you lose performance as well so now you have something better, equal or worse depending on application.

Do those high-freq kits run in "1:1" on Intel?

No they don't run 1:1 on Intel because it's a completely different IMC architecture and most of them can only clock 1:1 to about 3600-3800. I'm not sure why you think Intel and AMD are somehow interchangeable, because they're not and have not been in an extremely long time.

Uhm, no? There is 2x32GB DDR5-6000 kit for Intel but not for AMD.

There's no review for that kit up on the website so without it being dated who knows when it was reviewed.

There is very significant difference between 1-CCD and 2-CCD models. There is information that non-server 1-CCD doesn't have enough inter-CCD bandwidth to fully saturate memory controller. Can't say is that significant or not (and that's one of the reasons I would like to see proper article).

CPU bandwidth requirements have nothing to do with IMC stability and there's no way to directly compare them because the IMC in every CPU is different. You could very likely take a random 7950X and be able to clock memory higher than a 7700X and vice versa. This is why someone could have a 14900K that can only run 7200 DRAM and a 14600K that can run 8200 DRAM.

 

[setx]

No they don't run 1:1 on Intel because it's a completely different IMC architecture and most of them can only clock 1:1 to about 3600-3800. I'm not sure why you think Intel and AMD are somehow interchangeable, because they're not and have not been in an extremely long time.

I'm not sure why you have double standards for AMD/Intel. Do you "lose performance" going not 1:1 in Intel? Yes. But somehow it's totally normal, but for AMD it's something not only worth special mention but also rejecting testing kits that can't be run 1:1.

There's no review for that kit up on the website so without it being dated who knows when it was reviewed.

i9-13900K is newer than Ryzen 7 7700X, so either they tested on different Intel platform or I see no reasons to specially exclude AMD.

CPU bandwidth requirements have nothing to do with IMC stability and there's no way to directly compare them because the IMC in every CPU is different. You could very likely take a random 7950X and be able to clock memory higher than a 7700X and vice versa.

I have no idea what you are talking about. I'm not talking about overclocking – only about normal operation where 1-CCD CPU probably has lower ceiling of usable memory bandwidth so it's poor choice for memory chart to characterize the whole AMD platform.

 

[thestryker]

I'm not sure why you have double standards for AMD/Intel. Do you "lose performance" going not 1:1 in Intel? Yes. But somehow it's totally normal, but for AMD it's something not only worth special mention but also rejecting testing kits that can't be run 1:1.

AMD's Zen 4 architecture is designed around running 1:1 and this is in all of their recommendations while Intel is not. They are not the same and shouldn't be considered the same thing.

i9-13900K is newer than Ryzen 7 7700X, so either they tested on different Intel platform or I see no reasons to specially exclude AMD.

If it was indeed tested on the 13900K I certainly agree, but since there's no corresponding review or dates it's impossible to tell anything beyond its a kit that was tested at some point which is my point.

I have no idea what you are talking about. I'm not talking about overclocking – only about normal operation where 1-CCD CPU probably has lower ceiling of usable memory bandwidth so it's poor choice for memory chart to characterize the whole AMD platform.

You're claiming that potential CPU memory bandwidth usage somehow impacts usable memory clocks or their test and it doesn't. A slower kit is always slower and a faster kit is always faster it's about the platform/IMC.

 

[setx]

AMD's Zen 4 architecture is designed around running 1:1 and this is in all of their recommendations while Intel is not. They are not the same and shouldn't be considered the same thing.

What AMD can run some useful DDR5 frequency in 1:1 while Intel can't doesn't mean that they are restricted to only do so. There are no AMD recommendations for running DDR5-8000 in 1:1, it was "designed" to be run in 1:2.
Controllers are obviously different but conclusions shouldn't be based on double standards.

If it was indeed tested on the 13900K I certainly agree, but since there's no corresponding review or dates it's impossible to tell anything beyond its a kit that was tested at some point which is my point.

Basically, you are saying they might be lying about testing hardware?

You're claiming that potential CPU memory bandwidth usage somehow impacts usable memory clocks or their test and it doesn't.

Memory clocks? Where have I said that?
Affects program speed? Yes, by definition.

A slower kit is always slower and a faster kit is always faster it's about the platform/IMC.

No they aren't. If both can saturate the available bandwidth they will be the same speed, again, by definition.


I don't understand why you want to go to such length to defend that pathetic "article". You've already spent more effort here than it's author. Compare that to actual article that used to be published here that I've linked. No amount of small details that we are talking here about can shrink the huge gap between them.

 

[tamalero]

There needs to be an AI generated halloween costume.
Like some messed up hand gloves and random misshapements.

You would definitely need to label your kid as such though. Or else some might think they are trying to make fun of some of the handicapped.

I dont think handicapped kids are usually with 7 fingers and 6 legs with ghost patches of skin appearing randomly outside their clothes.

 

[tamalero]

Stealing content is still stealing even if you or original author don´t ask money for it!

And a good article. AI definitely belongs to tech related news...

You should see the latest video of Kyle Hill, where he mentions the rampant abuse and outright stealing from fake pseudo science websites and many more, trying to just suck as many money as possible.