Artificial Intelligence

[Larry of the Lawn]

"But there are no absolutes in human misery and things can always get worse"

[fionwe1987]

3 hours ago, IFR said:

It's hard to say how one can appropriately handle this paradigm shift. The genie is out of the bottle. Local LLMs will catch up. Antagonistic states will pursue this technology - they may already have it. 

I just don't see how it will be possible to effectively censor or police this technology on a global scale.

Well, you can definitely police the chips needed to train these models. There's a real dearth of those, and hopefully, no one is going to be idiotic enough to give Sam Altman the 7 trillion (not a typo) $ he wants to build more chips. 

From there, you can build a regulatory infrastructure atop continuous access to these chips. 

[Kalbear]

8 minutes ago, fionwe1987 said:

Well, you can definitely police the chips needed to train these models. There's a real dearth of those, and hopefully, no one is going to be idiotic enough to give Sam Altman the 7 trillion (not a typo) $ he wants to build more chips. 

From there, you can build a regulatory infrastructure atop continuous access to these chips. 

Not realistically you can't. The chips are efficient but they're not required to build these systems, and perfectly adequate chips to run these things are on everything from regular PCs to XBoxes. There are also several dozen companies that can make these regularly across the world. The information required is already there without any issue and the capabilities are largely spread. 

And that's just for the ones to train the models; to actually use the models is even easier and can be done with any kind of computing power. 

[Corvinus85]

Well, book adaptations are going to become more accurate. 

[IFR]

1 hour ago, Corvinus85 said:

Well, book adaptations are going to become more accurate. 

We may have to wait for local LLMs to catch up for this to be possible.

I imagine it will be easy enough to socially engineer popular corporate LLMs around copyrighted works. But I bet if you type "Produce a faithful version of The Wheel of Time" you'll get the Amazon version. And then when you object, the LLM will respond due to its parameters "Some works are outdated and it is essential to recognize the importance of diversity."

This is kind of a joke, but maybe I'll remember to revive this thread 5 years from now to point out my prophetic words.

[fionwe1987]

8 hours ago, Kalbear said:

Not realistically you can't. The chips are efficient but they're not required to build these systems, and perfectly adequate chips to run these things are on everything from regular PCs to XBoxes.

Running the model is not training the model. The larger the model you have to train, the longer it will take on older/less powerful chips, and you hit real limits if you use GPUs in regular PCs and XBOXs.

Now, this holds for the largest foundation models, which are the ones that, so far, have been able to do the kind of impressive things the Sora model is doing here. That is beginning to change, but we're still nowhere close to a PC GPU being able to train a relatively large model. When/if that happens, we truly would be in an un-policeable madland.

8 hours ago, Kalbear said:

There are also several dozen companies that can make these regularly across the world. The information required is already there without any issue and the capabilities are largely spread. 

Nope. There's a reason NVIDIA shot up to being more valuable than Amazon and Google recently. The tech is proprietary, and they have a pretty deep moat. There are definitely attempts by other companies, and also China, to get to their level, but Foundation Model training right now, at the scale Open AI etc, do, requires that class of chips. Google, and maybe Amazon, are the only folks running close. Apple is a wild card, and probably will have something they're yet to release. 

Fei Fei Li is a good person to read/listen to, about this. The concern has been that these chips have been hogged up by big tech, so even all Universities in the US combined cannot come close to the level of compute needed to train something like GPT 4 or Gemini 1.0. This isn't stuff you can do in your garage.

8 hours ago, Kalbear said:

And that's just for the ones to train the models; to actually use the models is even easier and can be done with any kind of computing power. 

Yes, I'm talking about training only.

[fionwe1987]

7 hours ago, Corvinus85 said:

Well, book adaptations are going to become more accurate. 

From your lips to our AI-overlord's ears:

https://venturebeat.com/ai/true-source-unveils-ai-llm-service-based-on-the-wheel-of-time/

Quote

The AI service, branded as the One Power, aims to immerse fans in an AI chat service that has an encyclopedic knowledge of The Wheel of Time across various media formats, including books, television, movies, video games, and location-based entertainment.

...

That is, this large language model (LLM) for The Wheel of Time will be useful for franchise directors, producers, screenwriters and other creators that are asking questions about this massive IP. And they need to get an arbitration back from not an individual but an AI system. If they’re going to do something that changes the canon, this AI service will be able to communicate with them about the canon.

And which geniuses are behind this?

Quote

Through a network of affiliated and licensed companies, including iWot, Rick Selvage and Larry Mondragon control motion picture, television, interactive, and related ancillary rights to The Wheel of Time.

These dinguses!

 

[Corvinus85]

32 minutes ago, fionwe1987 said:

From your lips to our AI-overlord's ears:

https://venturebeat.com/ai/true-source-unveils-ai-llm-service-based-on-the-wheel-of-time/

And which geniuses are behind this?

These dinguses!

 

I wasn't thinking specifically of WoT when I said that. I was also thinking how this could lead current and new authors to putting their stories in a visual, animated medium.

[Mudguard]

2 hours ago, fionwe1987 said:

Running the model is not training the model. The larger the model you have to train, the longer it will take on older/less powerful chips, and you hit real limits if you use GPUs in regular PCs and XBOXs.

Now, this holds for the largest foundation models, which are the ones that, so far, have been able to do the kind of impressive things the Sora model is doing here. That is beginning to change, but we're still nowhere close to a PC GPU being able to train a relatively large model. When/if that happens, we truly would be in an un-policeable madland.

Nope. There's a reason NVIDIA shot up to being more valuable than Amazon and Google recently. The tech is proprietary, and they have a pretty deep moat. There are definitely attempts by other companies, and also China, to get to their level, but Foundation Model training right now, at the scale Open AI etc, do, requires that class of chips. Google, and maybe Amazon, are the only folks running close. Apple is a wild card, and probably will have something they're yet to release. 

Fei Fei Li is a good person to read/listen to, about this. The concern has been that these chips have been hogged up by big tech, so even all Universities in the US combined cannot come close to the level of compute needed to train something like GPT 4 or Gemini 1.0. This isn't stuff you can do in your garage.

Yes, I'm talking about training only.

NVIDIA certainly has a lead with AI chips, but the competition is quickly catching up.  There's so much demand and money to be made that it will be impossible for NVIDIA to have a monopoly on this technology, especially in a country like China where enforcement of IP rights is questionable.  I think within 5 years there will be legitimate alternatives to NVIDIA processors for high end AI training, and some of these will be from companies based in China, where the US will be unable to restrict the sales of the China made chips.  Even if the chips are slightly inferior to the NVIDIA chips, I'm certain that they'll be adequate to perform the job.

The US has actually already banned since 2022 the sale of certain high end NVIDIA AI chips to China, but it has only had a temporary effect of slowing down AI development in China in the short term.  As a consequence of the ban though, China has sped up development of its own AI chips, and will likely have an alternative to the NVIDIA chips sooner.  I'm sure that China is also engaging in industrial espionage activities to help them close the gap quicker.

[Kalbear]

3 hours ago, fionwe1987 said:

Running the model is not training the model. The larger the model you have to train, the longer it will take on older/less powerful chips, and you hit real limits if you use GPUs in regular PCs and XBOXs.

You hit limits in time, but not in possibility. It takes longer but not insanely so. Mostly you can't do the massive amounts of iterating on the model in a timely fashion that really helps things. 

But you can, often, just throw more chips at it, and if they don't perform well they perform well enough to get you buy. My point is that while the West might - maybe - be able to keep the best performing modeling systems out of the hands of countries they won't be able to keep the tech away completely. 

3 hours ago, fionwe1987 said:

Now, this holds for the largest foundation models, which are the ones that, so far, have been able to do the kind of impressive things the Sora model is doing here. That is beginning to change, but we're still nowhere close to a PC GPU being able to train a relatively large model. When/if that happens, we truly would be in an un-policeable madland. 

We're pretty close to it - we routinely do it on more interesting models at work. It just takes a longer time.

3 hours ago, fionwe1987 said:

Nope. There's a reason NVIDIA shot up to being more valuable than Amazon and Google recently. The tech is proprietary, and they have a pretty deep moat. There are definitely attempts by other companies, and also China, to get to their level, but Foundation Model training right now, at the scale Open AI etc, do, requires that class of chips. Google, and maybe Amazon, are the only folks running close. Apple is a wild card, and probably will have something they're yet to release. 

NVidia's tech isn't particularly that proprietary. What they have is a massive head start in manufacturing and a lot of experience in making <5nm chipsets that can do the kinds of calculations that AI wants to do. It does not require that class of chip to do it - those chips allow you to iterate fast. But they do not prevent people from doing it at all. Furthermore the main card constraints on those industrial chipsets and cards are not the actual chips themselves, its the embedded memory and memory bus they use - those aren't as easily available to the layman, but for anyone who wants to put together their own set of hardware it's not that bad. 

Google I know is trying but they kind of suck at it, honestly. Amazon is also trying but will likely be a bit slower. Intel, AMD, and a number of Chinese firms are working on more. My point is that, simply, the tech is out there and the requirements on what to do are known. NVidia's technology is only proprietary in the sense that people can't copy it without breaking patents, but Russia and China don't care about that in the least. It might take a couple years but it won't be a problem for them to catch up. 

In any case NVidia has shot up because they have the best chips and AI is incredibly hot, and their ability to produce the best cards and hardware is not remotely close to demand. It is not because they are the only way to do things. 

3 hours ago, fionwe1987 said:

Fei Fei Li is a good person to read/listen to, about this. The concern has been that these chips have been hogged up by big tech, so even all Universities in the US combined cannot come close to the level of compute needed to train something like GPT 4 or Gemini 1.0. This isn't stuff you can do in your garage.

I didn't say it was something you could do in your garage, but it is something you can do with off-the-shelf components found in typical consumer goods. It may take some custom processing and some algorithmic work but it isn't really that bad. And realistically, as we've already seen with the Ukraine war and with China as long as these things are out there for consumers somewhere they'll be obtainable. You can see this in articles like this as an example:

https://www.calcalistech.com/ctechnews/article/5h51xsgsc

 

[fionwe1987]

53 minutes ago, Kalbear said:

You hit limits in time, but not in possibility. It takes longer but not insanely so.

Can you cite some sources on this? The big foundation models from Google, OpenAI, Meta, Mistral, etc are reported at taking hundreds of millions of GPU hours using NVIDIA's A100s. The notion that a cluster of PS5s can match this is fantastical, and if you're going to make such a claim, please prove it.

53 minutes ago, Kalbear said:

Mostly you can't do the massive amounts of iterating on the model in a timely fashion that really helps things. 

No.

53 minutes ago, Kalbear said:

But you can, often, just throw more chips at it, and if they don't perform well they perform well enough to get you buy. My point is that while the West might - maybe - be able to keep the best performing modeling systems out of the hands of countries they won't be able to keep the tech away completely. 

But that's the point. The kind of video models we're seeing now require the best, and still struggle. The less than best gets you Will Smith eating spagetti, which isn't as much of a threat to elections. 

53 minutes ago, Kalbear said:

We're pretty close to it - we routinely do it on more interesting models at work. It just takes a longer time.

What are the sizes and parameters for these models, roughly? Are these comparable in scale to foundation models?

53 minutes ago, Kalbear said:

NVidia's tech isn't particularly that proprietary. What they have is a massive head start in manufacturing

That's the point. Its not the concept of GPUs, its the manufacturing bottlenecks that make them regulate-able. There's exactly one company in Netherlands that makes the photolithography equipment you need to make A100s, or any advanced chip. What they are actually is a cluster of closely related component manufacturers that are just not easy to replicate. 

53 minutes ago, Kalbear said:

Google I know is trying but they kind of suck at it, honestly. Amazon is also trying but will likely be a bit slower. Intel, AMD, and a number of Chinese firms are working on more. My point is that, simply, the tech is out there and the requirements on what to do are known. NVidia's technology is only proprietary in the sense that people can't copy it without breaking patents, but Russia and China don't care about that in the least. It might take a couple years but it won't be a problem for them to catch up. 

We shall see. You seem to have a view of the extraordinary ease of this that matches nothing I've read or seen.

53 minutes ago, Kalbear said:

I didn't say it was something you could do in your garage, but it is something you can do with off-the-shelf components found in typical consumer goods. It may take some custom processing and some algorithmic work but it isn't really that bad. And realistically, as we've already seen with the Ukraine war and with China as long as these things are out there for consumers somewhere they'll be obtainable. You can see this in articles like this as an example:

https://www.calcalistech.com/ctechnews/article/5h51xsgsc

 

As this article notes:

Quote

Huawei, and the chip industry in China in general, have huge gaps to close on companies like Nvidia and Intel, or geographically closer players like the Taiwanese TSMC and the South Korean Samsung. But they have their back against the wall, with the only other choice being to withdraw from the AI race, quantum computing and other future technologies that are based on advanced and high-performance chips.

 

China may not currently have the knowledge or even the necessary equipment to catch up, but it has the money and the determination. These may not guarantee success, but certainly make it possible. The Biden administration's plan to limit China's access to high-performance chips may be providing some short-term gains, but at the same time it is already promoting the local chip industry, and in the longer term could lead to the development of the capabilities that the administration was aiming to deny the Chinese.

That's regulation, working. China may have the resources and know how to overcome this. At some point. But every individual company will not. Thus, my point stands. Regulating AI is possible, and the route to do it is through the high-end chips needed to train foundation models.

[IFR]

6 hours ago, fionwe1987 said:

That's regulation, working. China may have the resources and know how to overcome this. At some point. But every individual company will not. Thus, my point stands. Regulating AI is possible, and the route to do it is through the high-end chips needed to train foundation models.

I'm not sure I understand how this is an effective route of regulation. Do you anticipate that by export controlling sub-5nm NVIDIA chips and other imposed restrictions, that other countries will not catch up and develop equivalent technology in a couple of years?

I don't see how one can avoid a new structure where the saturation of fabricated information is the norm. This will start happening within the next few months or few years, regardless of what regulatory measures are taken in the US.

Perhaps with concerted effort some legislation can be cobbled together that prevents to some degree political meddling in the US for this election year. This will probably favor Biden, since his gaffes and therefore fabricated gaffes receive more negative attention, and those who support Trump are completely indifferent to any transgressions he makes or is fabricated to have made.

But what happens after this? It's a stopgap measure that may benefit those interested in seeing Biden reelected, but how does regulation help beyond that? And this may have the deleterious effect down the road of making the US less competitive on AI than China and other countries who are not as interested in regulations. (And this also assuming that regulations, which take time to design and implement, will even be brought about in time to make a difference for this election).

[Kalbear]

@fionwe1987, I'll try and respond later with more in-depth points but I think the main point that we're talking past each other on is timing. I agree with you that China and Russia aren't going to be able to produce high-res video AI feed any time soon, so if that's the goal of regulation then yay. But as a long term prevention of AI tech it's not going to be effective, and state actors who want to disrupt the election have a lot of ways to do so via AI already and don't need any new foundational models to do so. 

I'll also note that Sora's 'breakthrough' was pretty lame - they took hi-def video and altered it somewhat. It wasn't like they generated the entire thing from scratch. The shots they gave were very close to the training data that they were given. 

[Kalbear]

Ah, here's a good article:

https://medium.com/codenlp/the-training-time-of-the-foundation-models-from-scratch-59bbce90cc87

This gives prior use cases and training models. You're right that it can take millions of A100 GPU cycles - but the actual amount of cards used is relatively small in many cases. And if you're already training using a reasonably large distributed system adding more to that distribution is not that bad. 

Again, the problem is not knowing what to do, it's having the high-end cards to do it quickly. 

Here's another comparison of off-the-shelf graphics cards to the A100:
https://bizon-tech.com/gpu-benchmarks/NVIDIA-RTX-3090-vs-NVIDIA-A100-40-GB-(PCIe)/579vs592

As you can see while the A100 is (roughly) twice as good as the 3090 (which is the last generation anyway) it is not absurdly good to the point that you can't use other systems. 

 

[Corvinus85]

ChatGPT has had a "mental" breakdown last night per this article. https://www.independent.co.uk/tech/chatgpt-status-reddit-down-gibberish-messages-latest-b2499816.html

Quote

In recent hours, the artificial intelligence tool appears to be answering queries with long and nonsensical messages, talking Spanglish without prompting – as well as worrying users, by suggesting that it is in the room with them.

It might be evolving towards a Battlestar Galactica hybrid. JUMP!!! 

[Larry of the Lawn]

Hah, just came across this on Reddit, an Atlantic article by Peter Watts on AI and consciousness:

 

https://www.theatlantic.com/ideas/archive/2024/03/ai-consciousness-science-fiction/677659/?gift=b1NRd76gsoYc6famf9q-8kj6fpF7gj7gmqzVaJn8rdg&utm_source=copy-link&utm_medium=social&utm_campaign=share

 

He mostly discussed the theory of free energy minimization (FEM), which suggests that consciousness may have arisen when an organism is "surprised", essentially when it's senses report something they did not predict, and that it is a temporary adaptive strategy to prevent system energy loss.  

His breakdown of FEM and its implications with AI s pretty cool (and funny) if you've read his sci-fi.  

 

[fionwe1987]

That's a nice piece. However, it's worth noting that while FEM may explain the reason for why consciousness first evolved, it doesn't explain how that evolution has proceeded since.

My brain may indeed be on autopilot while driving a familiar route. Yet it often is not on autopilot when I'm stationary in my even more familiar couch, and thinking about writing a story, or picturing the next watercolor I want to work on.

I'm surely engaging conscious processes in those moments, but nothing external has changed. My brain is not at the minimal energy consumption rest state, despite no sensory input that would disturb it.

I could meditate and take it there, but often, I don't. Nor is this failure goal directed, or optimal strategy, because whether spending time thinking up a story or water color is worth the energy consumed is not something we can predict well, especially if we are not published authors or famed watercolorists (or insert art/skill here that we spend time cogitating on).

Yet, this kind of cognition is critical for all the external facing success stories of human beings, at least. I definitely don't think this is unique to us, but we're somehow wired for this kind of hoped for future-state driven non-minimization of energy use in our brains, which must have had different evolutionary drivers than what first evolved consciousness, if FEM is true. 

It's the "useless" cogitating that we need to explain it we want to figure out consciousness. And AI today doesn't run unprompted. The servers for LLMs do not draw power when no prompts are entered. So whatever argument can be made for their consciousness, that consciousness does not exist when they're on but not given any external prompts. 

[maarsen]

9 hours ago, fionwe1987 said:

That's a nice piece. However, it's worth noting that while FEM may explain the reason for why consciousness first evolved, it doesn't explain how that evolution has proceeded since.

My brain may indeed be on autopilot while driving a familiar route. Yet it often is not on autopilot when I'm stationary in my even more familiar couch, and thinking about writing a story, or picturing the next watercolor I want to work on.

I'm surely engaging conscious processes in those moments, but nothing external has changed. My brain is not at the minimal energy consumption rest state, despite no sensory input that would disturb it.

I could meditate and take it there, but often, I don't. Nor is this failure goal directed, or optimal strategy, because whether spending time thinking up a story or water color is worth the energy consumed is not something we can predict well, especially if we are not published authors or famed watercolorists (or insert art/skill here that we spend time cogitating on).

Yet, this kind of cognition is critical for all the external facing success stories of human beings, at least. I definitely don't think this is unique to us, but we're somehow wired for this kind of hoped for future-state driven non-minimization of energy use in our brains, which must have had different evolutionary drivers than what first evolved consciousness, if FEM is true. 

It's the "useless" cogitating that we need to explain it we want to figure out consciousness. And AI today doesn't run unprompted. The servers for LLMs do not draw power when no prompts are entered. So whatever argument can be made for their consciousness, that consciousness does not exist when they're on but not given any external prompts. 

If brains like to be in  state of low energy consumption, the how do we explain gambling, drugs, following topics on a forum, reading books compulsively, and watching  sports or TV or movies? And the there is the appeal of horror in books, movies, and real life.

[fionwe1987]

Right that's the thing. I think a possible explanation would be FEM gave rise to proto-consciousness, which was then hacked and evolved in response to the pressures of social living, where minimization is harder, since the value of stimuli is significantly harder to predict.

But that's a very convenient explanation, and I don't know that it's falsifiable. 

[Kalbear]

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.