Artificial Intelligence

[Ran]

I saw an interesting proposal that the only thing that really needs regulation is what share of energy AIs are allowed to consume in relation to all other non-AI use.

[Mudguard]

1 hour ago, Ran said:

I saw an interesting proposal that the only thing that really needs regulation is what share of energy AIs are allowed to consume in relation to all other non-AI use.

Do you have a link?  What are they trying to accomplish by regulating the energy used?  How do they propose to implement this regulation?  This doesn't seem a practical solution.

[Kalbear]

AI chips and systems are very power hungry - very similar to bitcoin, which makes sense given that they, too, are doing massive amounts of math calculations as fast as they can. 

That said it ain't gonna stop China or anything like that.

[Mudguard]

35 minutes ago, Kalbear said:

AI chips and systems are very power hungry - very similar to bitcoin, which makes sense given that they, too, are doing massive amounts of math calculations as fast as they can. 

That said it ain't gonna stop China or anything like that.

Yes, but I don't see how you would practically regulate the industry based on power consumption.  It's not clear to me what the ultimate goal of the regulation is.  

[Kalbear]

12 minutes ago, Mudguard said:

Yes, but I don't see how you would practically regulate the industry based on power consumption.  It's not clear to me what the ultimate goal of the regulation is.  

Ah, okay. It would potentially regulate the industry in terms of making them actually pick and choose interesting problems instead of just throwing AI at everything, but it certainly isn't going to make it not abusive or dangerous by itself. 

[Mudguard]

10 minutes ago, Kalbear said:

Ah, okay. It would potentially regulate the industry in terms of making them actually pick and choose interesting problems instead of just throwing AI at everything, but it certainly isn't going to make it not abusive or dangerous by itself. 

This doesn't seem workable in practice, unless the entire industry was centrally planned by the government.  If the cap was placed at the industry level, and one company just decided to consume as much as it wanted, how would that affect other companies?

[Kalbear]

25 minutes ago, Mudguard said:

This doesn't seem workable in practice, unless the entire industry was centrally planned by the government.  If the cap was placed at the industry level, and one company just decided to consume as much as it wanted, how would that affect other companies?

You can either tax them significantly past that point in terms of consumption or require them to purchase power allowances (similar to water rights or other things) and have them have civil or criminal penalties for going over. 

As to how it would affect other companies - there is only so much power and resources to go around; this would mean that the biggest fish would not be able to simply take all the resources for computation that exist and not let anyone else work on them. 

[Mudguard]

2 hours ago, Kalbear said:

You can either tax them significantly past that point in terms of consumption or require them to purchase power allowances (similar to water rights or other things) and have them have civil or criminal penalties for going over. 

As to how it would affect other companies - there is only so much power and resources to go around; this would mean that the biggest fish would not be able to simply take all the resources for computation that exist and not let anyone else work on them. 

The only thing this appears to address is the amount of energy used by the industry, which is an odd concern to me at this time.  This has zero chance or reason to be implemented as a regulation right now.  It’s a solution to a nonexistent problem that only serves to handicap domestic AI development.

[fionwe1987]

14 hours ago, Kalbear said:

Another example of how limiting access to specific chips probably isn't going to matter very much: 

https://www.tomshardware.com/pc-components/cpus/phisons-new-software-uses-ssds-and-dram-to-boost-effective-memory-for-ai-training-demos-a-single-workstation-running-a-massive-70-billion-parameter-model-at-gtc-2024

In this case software is allowing for extra use of SSDs and memory to run a model that normally requires supercomputers on a single workstation using off the shelf parts. 

Sigh. You might want to read the articles you link. No, you don't need supercomputers to do what this architecture is allowing:

Quote

The company's demo served as a strong proof point for the tech, showing a single workstation with four Nvidia RTX 6000 Ada A100 GPUs running a 70 billion parameter model. Larger AI models are more accurate and deliver better results, but Phison estimates that a model of this size typically requires about 1.4 TB of VRAM spread across 24 AI GPUs dispersed across six servers in a server rack — and all the supporting networking and hardware required.

In which century does this qualify as a supercomputer?

Additionally, 70B parameter models are great, if you're in late 2022 or early 2023. 

Meanwhile, the new Nvidia GB200 is designed to support 27 trillion parameter models:

Quote

Nvidia is counting on companies to buy large quantities of these GPUs, of course, and is packaging them in larger designs, like the GB200 NVL72, which plugs 36 CPUs and 72 GPUs into a single liquid-cooled rack for a total of 720 petaflops of AI training performance or 1,440 petaflops (aka 1.4 exaflops) of inference. It has nearly two miles of cables inside, with 5,000 individual cables.

Each tray in the rack contains either two GB200 chips or two NVLink switches, with 18 of the former and nine of the latter per rack. In total, Nvidia says one of these racks can support a 27-trillion parameter model. GPT-4 is rumored to be around a 1.7-trillion parameter model.

Source: https://www.theverge.com/2024/3/18/24105157/nvidia-blackwell-gpu-b200-ai

The Phison architecture is primarily a way to save on cost. It's really impressive, and there's a heck of a lot of use cases where 70B parameters is more than enough, and a company wanting to retrain Meta or Mistral's open source models with private data would be well advised to go that route than hope to lay their hands on Nvidia chips.

But we're very far from the state of the art being challenged. It will happen, of course. Neuromorphic chips and other chip architectures are coming which may totally wipe out Nvidia's current lead. But nothing stops regulation from being adjusted for that. 

Well, the complete clusterfuck that is the US government being so hapless does, but nothing conceptually prevents state of the art chips, of whatever architecture, from being regulated, especially if the goal is to keep them from China. Individual companies can still get slower, cheaper systems and do plenty of mischief, but that genie got out of the bottle before we knew enough to even think of regulating it. Now we know better, and the focus should be on conditional access to these chips. 

A 27 trillion parameter model, I think, can start approaching video and audio based training at levels of success that earlier models had with text. And that's truly terrifying, and regulations should focus there to make sure the tech isn't reaching actors who will use it to cause the kind of complete mayhem you can get if you can generate video with the level of quality models can generate text today. 

Edited March 27 by fionwe1987

[Kalbear]

11 hours ago, fionwe1987 said:

Meanwhile, the new Nvidia GB200 is designed to support 27 trillion parameter models:

Source: https://www.theverge.com/2024/3/18/24105157/nvidia-blackwell-gpu-b200-ai

Just to be clear, that's on a rack that has 18 of them. It's not quite the absurd leap from a single chip that implies. 

11 hours ago, fionwe1987 said:

The Phison architecture is primarily a way to save on cost. It's really impressive, and there's a heck of a lot of use cases where 70B parameters is more than enough, and a company wanting to retrain Meta or Mistral's open source models with private data would be well advised to go that route than hope to lay their hands on Nvidia chips.

But we're very far from the state of the art being challenged. It will happen, of course. Neuromorphic chips and other chip architectures are coming which may totally wipe out Nvidia's current lead. But nothing stops regulation from being adjusted for that. 

Again, we don't need to challenge state of the art to cause major problems. AI modeling is already at a stage where the current paradigms of models can be used on training data to be massively problematic. And because it already is built to do scaling and fan-out processing all controlling chips does is increase either the price or time you have to spend. 

I think thinking about controlling these as a security concern is a category error. We typically think about that in terms of either making sure that people don't have access to the underlying technological underpinnings, or they don't have access to the actual hardware because it's so hard to duplicate. The former is not particularly special because NVidia has been so open about how their chips are made and done; the secrets that they have are in manufacturing, not in the technology. And getting hands on the chips matters a hell of a lot less when you don't have to worry about it taking up space; when you're trying to pilot a missile having 30 chips and processing slower isn't going to be a useful workaround, but that isn't at all the case when you're just putting stuff in a datacenter. 

11 hours ago, fionwe1987 said:

A 27 trillion parameter model, I think, can start approaching video and audio based training at levels of success that earlier models had with text. And that's truly terrifying, and regulations should focus there to make sure the tech isn't reaching actors who will use it to cause the kind of complete mayhem you can get if you can generate video with the level of quality models can generate text today. 

Eh. I'm pretty sure that either that's already there in sufficient abilities or has nothing to do with the actual models in question. Throwing a lot of power at video/audio training isn't the primary solution, any more than it is with LLMs, because more power does not mean more contextual understanding. 

[Kalbear]

In more relevant news, AI is being used to pick targets for Israel in Gaza - and it has a 10% failure rate:

https://www.cnn.com/2024/04/03/middleeast/israel-gaza-artificial-intelligence-bombing-intl/index.html

I guess that 10% is acceptable as long as you're not particularly punished for it. 

[Mudguard]

3 hours ago, Kalbear said:

In more relevant news, AI is being used to pick targets for Israel in Gaza - and it has a 10% failure rate:

https://www.cnn.com/2024/04/03/middleeast/israel-gaza-artificial-intelligence-bombing-intl/index.html

I guess that 10% is acceptable as long as you're not particularly punished for it. 

I’m extremely skeptical that the error rate is that low.  Unless they share their validation methodology and allow an independent review of the data, I don’t buy it.  It’s almost certainly a shit model trained with shit data, and now deployed in the field using more shit data.  

[Zorral]

This is how it works:

https://www.theverge.com/2024/4/4/24120352/israel-lavender-artificial-intelligence-gaza-ai

also this:

https://www.bbc.com/news/world-middle-east-68713015

Edited April 4 by Zorral

[fionwe1987]

Are we even allowed to discuss this topic?

On the assumption that we are... 

I agree, the model has shit training data, and very little guarantee that the right features have been selected for training, or that the model is correctly getting updated based on results, with the "fog of war" preventing a lot of useful inputs from being gathered.

This is very much a use of "AI" where it serves more as an excuse to allow humans to do what they want, and point their finger at a model that can be a convenient scapegoat.

[straits]

It will be an interesting study in motivated reasoning, but so far it looks like the main issues are treating the AI output as a human decision, and then allowing disturbingly lax parameters on top of that.

 

Quote

According to six Israeli intelligence officers, who have all served in the army during the current war on the Gaza Strip and had first-hand involvement with the use of AI to generate targets for assassination, Lavender has played a central role in the unprecedented bombing of Palestinians, especially during the early stages of the war. In fact, according to the sources, its influence on the military’s operations was such that they essentially treated the outputs of the AI machine “as if it were a human decision.”

...

One source stated that human personnel often served only as a “rubber stamp” for the machine’s decisions, adding that, normally, they would personally devote only about “20 seconds” to each target before authorizing a bombing — just to make sure the Lavender-marked target is male.

The choice of target and acceptable conditions for attack really seem to be the key issue, along with the generous tolerance for civilian casualties:

Quote

Moreover, the Israeli army systematically attacked the targeted individuals while they were in their homes — usually at night while their whole families were present — rather than during the course of military activity. According to the sources, this was because, from what they regarded as an intelligence standpoint, it was easier to locate the individuals in their private houses.

...

In an unprecedented move, according to two of the sources, the army also decided during the first weeks of the war that, for every junior Hamas operative that Lavender marked, it was permissible to kill up to 15 or 20 civilians;

 

We can put this in the "manmade horrors beyond my comprehension" category.

Edited April 5 by straits
readability

[Zorral]

Israel offers a glimpse into the terrifying world of military AI

https://www.washingtonpost.com/world/2024/04/05/israel-idf-lavender-ai-militarytarget/

Extraordinary chilling reading because this can happen anywhere.

[Corvinus85]

Sigh

https://www.theverge.com/2024/4/18/24133870/us-air-force-ai-dogfight-test-x-62a

Quote

The US Air Force is putting AI in the pilot’s seat. In an update on Thursday, the Defense Advanced Research Projects Agency (DARPA) revealed that an AI-controlled jet successfully faced a human pilot during an in-air dogfight test carried out last year.

 

[Ran]

 

 

[Phylum of Alexandria]

I am disconcerted by the recent emergence of mandatory face scanning at airports. Right now it's a convenience, but should our government fall into the clutches of a paranoid and vindictive autocrat (which is basically a coin flip in the US at this point), you know that tech will be abused. "Oh, you want to flee the country, Mr. Resistance? Oh no, it looks like there's an issue with your passport..."

[Kalbear]

9 minutes ago, Phylum of Alexandria said:

I am disconcerted by the recent emergence of mandatory face scanning at airports. Right now it's a convenience, but should our government fall into the clutches of a paranoid and vindictive autocrat (which is basically a coin flip in the US at this point), you know that tech will be abused. "Oh, you want to flee the country, Mr. Resistance? Oh no, it looks like there's an issue with your passport..."

This is already the case, just not for people like you.