8 ms·

The same was said 10 years ago. It's astonishing what can be done, but you can already see fundamental limits. I think it will raise productivity for some tasks

The same was said 10 years ago. It's astonishing what can be done, but you can already see fundamental limits. I think it will raise productivity for some tasks, but not fundamentally change society.

danielbln 4y ago
Can you elaborate on fundamental limits?
- maxlamb 4y ago
  I’m guessing one is data. The limit would be once you’ve trained a LLM on all public (or even private) data. Sure you can still make some improvements or try to find some additional private data but still, a fundamental limit has been reached.
  - macrolime 4y ago
    Is it even feasible any time soon to train an LLM on all of YouTube?
    - danielbln 4y ago
      Napkin math, assuming around 156 million hours of video on all of Youtube: 156 million hours of YouTube videos 9,000 words/hour 6 characters/word (including space) First, let's find out the total number of characters: 9,000 words/hour \* 6 characters/word = 54,000 characters/hour Now, let's calculate the total number of characters for 156 million hours of YouTube videos: 54,000 characters/hour \* 156,000,000 hours = 8,424,000,000,000 characters Since 1 character is typically 1 byte, we can convert this to gigabytes: 8,424,000,000,000 bytes / (1024 \* 1024 \* 1024) ≈ 7,842.11 GB So, 8TB of text? Seems doable.
      - macrolime 4y ago
        I mean the actual video, that's much bigger With a vision transformer each token may be around 16x16 pixels. I found an example where they use images of resolution 224x224 for training a vision transformer so if we go with that that 256 pixels per token and 50176 pixels per image, so 196 tokens per frame, 24 frames per second, that's 4704 tokens per second or 16934400 token / hour. In total we're at 2.6x10^15 tokens. GPT-3 was trained on 5x10^11 tokens, so YouTube done this way would be around four orders of magnitude more tokens that GPT-3 was trained on. GPT-3 was undertrained by 1-2 orders of magnitude, so the compute required to trained a model on YouTube would then be around 6 orders of magnitude higher than what was used to train GPT-3, so about one million times more. I did a linear regression on the training costs from cerebras(1) and came up with the formula (1901.67366*X)-197902.72715 where X is number of tokens in billions. Plugging in 5x10^15 tokens we get a training cost of 5 billion dollars. I guess a lot of optimizations could be done that would decrease the cost, so maybe its doable in a few years. 1. https://cirrascale.com/cerebras.php https://cirrascale.com/cerebras.php
  - FrojoS 4y ago
    Good point. But isn't the next logical step to allow these systems to collect real world data on their own? And also, potentially even more dangerous, act in the real world and try out things, and fail, to further its learning.
    - pixl97 4y ago
      This is likely exactly what we will do... which is very questionable when you may have an unaligned paperclip maximizer hidden in there.
  - Espressosaurus 4y ago
    I'm wondering what happens once LLMs are generating large portions of the internet. What then? It's poisoning its own well at that point.
  - great_psy 4y ago
    I think that’s actually really not a limit. We don’t teach babies by throwing lots of data at them, instead we teach them by giving them useful data. The loop I see is: - train on a lot of existing data - run out of useful data - people use ai and give feedback (we’re here) - perform reinforcement learning on the data collected Loop over the last two steps. There is already more than enough data available, it’s just not nicely labeled to say if it’s high quality or should be discarded. Those last two steps will implicitly do the labeling.
    - pixl97 4y ago
      Reality throws a lot of unfiltered data at a baby. Now I guess you can consider things like gravity and pain useful data because of the consequences of violating them. But it's this data that grounds the baby in the world it exists in.
- lamp987 4y ago
  there is no channel for uncertainty. LLMs of this type will just start making up shit when they dont know something. because they simply generate the most probable next token based on previous x tokens. this is not fixable. this alone makes these LLMs practically unusable in vast majority of real-world applications where you would otherwise imagine this tech to be used.
  - greyman 4y ago
    What if LLM knowledge will expand over time to be sufficient for certain real-world application?
    - Jeff_Brown 4y ago
      It already is: translation.
  - jazzyjackson 4y ago
    yea its a simulator of human text on the internet for instance, your comment confidently states this is unfixable - presumably based on the frequency you've seen similar text on the internet. why should anyone believe the veracity of your statement? These things didn't have any of these emergent capabilities one year ago, why are you so sure you understand their nature one year from now?
    - lamp987 4y ago
      "your comment confidently states this is unfixable - presumably based on the frequency you've seen similar text on the internet. why should anyone believe the veracity of your statement? " no its because GPT is based on transformers.
      - coldtea 4y ago
        and you aren't? Aren't you just a function of your input and memories (stuff you've read, sensory input) as run through/managed by some neural network? What makes you think the rest isn't just emergent properties? And what makes you think you can't hook up the LLM with some algorithms or layers that handle some of the rest behavior of what your brain does?
- datpiff 4y ago
  > According to NVidia, LLM sizes have been increasing 10X per year for the last few years. Clearly this cannot continue, as the training costs will exceed all the compute capacity in existence. The other limit is training data, eventually you run out of cheap sources.
  - Ambolia 4y ago
    They may also run out of data if they already consumed most of the internet. Or start producing so much of the internet's content that LLMs start consuming what they write in a closed loop.
    - warkdarrior 4y ago
      Sure, but they do not need to grow infinitely more, they just need to grow sufficiently to perform better than 90% of humans.
  - saiojd 4y ago
    I would not be so sure about compute capacity? Neural network architectures are still in their infancy, it is very likely that more efficient approaches exist.
    - datpiff 4y ago
      Sure, but the small & efficient networks so far have been pretty poor. Edge AI was hugely hyped a few years ago and it petered out. All of the big tech companies seems to be chasing the biggest networks with the largest training sets, that seems to be the direction for now. There could be huge efficiency savings within the implementations but when training costs are already this high it seems naive to think the low-hanging fruit is still there.
      - saiojd 4y ago
        I wouldn't be so confident. Flash attention is a recent, significant improvement to training times and sure looks low-hanging now. I was not familiar with Edge AI, interesting concept. I feel like improving the efficiency of very large models is much more likely. The recent successes will lead to an even larger influx of $ in the short term --- this is an existential threat to Google, after all. We will see where things go!
sebzim4500 4y ago
The number of people saying it now is many orders of magnitude more than the number of people saying it 10 years ago. Not saying that means it will happen, but it isn't the same situation.
- lxgr 4y ago
  > The number of people saying it Popular sentiment is a pretty meaningless metric for predicting the future.
  - AgentME 4y ago
    Then we shouldn't use the 10 year-old (lesser) popular sentiment as a reason to discount AI either.
    - lxgr 4y ago
      That still doesn’t work. A reliably wrong predictor is still valuable (just negate the output). I‘m arguing public sentiment will be right 50% of the time (in difficult binary predictions), not 0%.
abudabi123 4y ago
The A.I. Monte Carlo Ray Tracing Hallucination Engine can change society by showing as-is and to-be next state. Two use-cases: new infrastructure installation or upgrade and time interval inspection tracking ideal-case vs real world condition. Suppose a lazy contractor skips steps and cuts corners, or a pathological contractor builds the thing and pulls the thing apart over and over again when all that was needed was a one and done, or the change is for the worse. A civil engineer can walk around with an iPad and iPhone to check-in the master plan.
- Jeff_Brown 4y ago
  You mean AI can detect when a plumber is bulshitting me? That sounds great, could you elaborate?
  - coldtea 4y ago
    Given a few sensors that could be installed on the house's pipes, one can imagine several ways an AI can check that all is good. Same for checking blueprints for quality/validity and purchases the plumber says are "needed"...
- majormajor 4y ago
  You need to replace the civil engineer there (in today's world this is an inspector, I don't know for sure if they are professional engineers or not) for it to be useful, but you still need to have someone not working for the contractor who's incentivized to fake it. The trouble with many of those construction examples is that they're point in time. Installer fucks up how the windows are installed re: flashing and water-proofing, but then puts the siding over it... the error is now completely hidden. You could automated the inspection by AI photo analysis of every single window install on the project, say - but we could already do that for the inspection vs sending someone out, and send the photos to the expert instead, and we don't. Whether that's for lack of incentive to improve, or for ability to go deeper out of distrust for builders? I don't know.
HervalFreire 4y ago
What they said 10 years ago was correct. It did hit society like a sledge hammer. Machine learning basically took over the AI space and penetrated the consumer space with applications that were all but impossible in the previous decade. There's AI chips in smart phones now. What you're seeing here with LLMs is sledge hammer number 2. It's understandable how most people don't notice the sledge hammer. The decade prior to 2010 there was another sledge hammer had no smart phones. We were hit with a smart phone hammer AND an AI sledge hammer and the integration was so seamless we didn't even notice. Much of the same will happen with LLMs. In 5 years it's so normal, nobody cares and likely we will forget what life was like before LLMs.
- anon7725 4y ago
  Sledge hammer #1 (voice assistants, AI chips in phones) didn’t cause unemployment. It was at the level of new features and capabilities. Sledge hammer #2 is aimed squarely at “white collar” work without much in the way of bounds to its capabilities.
  - Jeff_Brown 4y ago
    We can't trust it. That's a pretty hard bound.
    - HervalFreire 4y ago
      For article writing verification is quick, I don't need to worry about a "bug" in the generated article as I do with code. For art, verification is instant, a bad generation is rejected. Trust is only a bound for certain areas, and this bound is eroding.
    - anon7725 4y ago
      Consider that a lot of useful work involves summarization (search++). “What work has gone on for supply chain improvement in our European operations this year?” - this is the kind of question that is easy to ask in natural language but might take someone a week of searching, messaging, etc to assemble. An LLM with access to all of the code, documents, chats, etc could just give an answer with citations. We are not betting $1B on the answer that it gives, but it has saved us 1 week of work and allows us to move on to the next step in the project. There are plenty of tasks like this which are highly valuable yet don’t require high trust. The one-shot “what is the airspeed of a coconut-laden swallow” type questions are actually fairly rare.

The same was said 10 years ago. It's astonishing what can be done, but you can already see fundamental limits. I think it will raise productivity for some tasks, but not fundamentally change society.

danielbln 4y ago

Can you elaborate on fundamental limits?

maxlamb 4y ago
I’m guessing one is data. The limit would be once you’ve trained a LLM on all public (or even private) data. Sure you can still make some improvements or try to find some additional private data but still, a fundamental limit has been reached.
- macrolime 4y ago
  Is it even feasible any time soon to train an LLM on all of YouTube?
  - danielbln 4y ago
    Napkin math, assuming around 156 million hours of video on all of Youtube: 156 million hours of YouTube videos 9,000 words/hour 6 characters/word (including space) First, let's find out the total number of characters: 9,000 words/hour \* 6 characters/word = 54,000 characters/hour Now, let's calculate the total number of characters for 156 million hours of YouTube videos: 54,000 characters/hour \* 156,000,000 hours = 8,424,000,000,000 characters Since 1 character is typically 1 byte, we can convert this to gigabytes: 8,424,000,000,000 bytes / (1024 \* 1024 \* 1024) ≈ 7,842.11 GB So, 8TB of text? Seems doable.
    - macrolime 4y ago
      I mean the actual video, that's much bigger With a vision transformer each token may be around 16x16 pixels. I found an example where they use images of resolution 224x224 for training a vision transformer so if we go with that that 256 pixels per token and 50176 pixels per image, so 196 tokens per frame, 24 frames per second, that's 4704 tokens per second or 16934400 token / hour. In total we're at 2.6x10^15 tokens. GPT-3 was trained on 5x10^11 tokens, so YouTube done this way would be around four orders of magnitude more tokens that GPT-3 was trained on. GPT-3 was undertrained by 1-2 orders of magnitude, so the compute required to trained a model on YouTube would then be around 6 orders of magnitude higher than what was used to train GPT-3, so about one million times more. I did a linear regression on the training costs from cerebras(1) and came up with the formula (1901.67366*X)-197902.72715 where X is number of tokens in billions. Plugging in 5x10^15 tokens we get a training cost of 5 billion dollars. I guess a lot of optimizations could be done that would decrease the cost, so maybe its doable in a few years. 1. https://cirrascale.com/cerebras.php https://cirrascale.com/cerebras.php
- FrojoS 4y ago
  Good point. But isn't the next logical step to allow these systems to collect real world data on their own? And also, potentially even more dangerous, act in the real world and try out things, and fail, to further its learning.
  - pixl97 4y ago
    This is likely exactly what we will do... which is very questionable when you may have an unaligned paperclip maximizer hidden in there.
- Espressosaurus 4y ago
  I'm wondering what happens once LLMs are generating large portions of the internet. What then? It's poisoning its own well at that point.
- great_psy 4y ago
  I think that’s actually really not a limit. We don’t teach babies by throwing lots of data at them, instead we teach them by giving them useful data. The loop I see is: - train on a lot of existing data - run out of useful data - people use ai and give feedback (we’re here) - perform reinforcement learning on the data collected Loop over the last two steps. There is already more than enough data available, it’s just not nicely labeled to say if it’s high quality or should be discarded. Those last two steps will implicitly do the labeling.
  - pixl97 4y ago
    Reality throws a lot of unfiltered data at a baby. Now I guess you can consider things like gravity and pain useful data because of the consequences of violating them. But it's this data that grounds the baby in the world it exists in.
lamp987 4y ago
there is no channel for uncertainty. LLMs of this type will just start making up shit when they dont know something. because they simply generate the most probable next token based on previous x tokens. this is not fixable. this alone makes these LLMs practically unusable in vast majority of real-world applications where you would otherwise imagine this tech to be used.
- greyman 4y ago
  What if LLM knowledge will expand over time to be sufficient for certain real-world application?
  - Jeff_Brown 4y ago
    It already is: translation.
- jazzyjackson 4y ago
  yea its a simulator of human text on the internet for instance, your comment confidently states this is unfixable - presumably based on the frequency you've seen similar text on the internet. why should anyone believe the veracity of your statement? These things didn't have any of these emergent capabilities one year ago, why are you so sure you understand their nature one year from now?
  - lamp987 4y ago
    "your comment confidently states this is unfixable - presumably based on the frequency you've seen similar text on the internet. why should anyone believe the veracity of your statement? " no its because GPT is based on transformers.
    - coldtea 4y ago
      and you aren't? Aren't you just a function of your input and memories (stuff you've read, sensory input) as run through/managed by some neural network? What makes you think the rest isn't just emergent properties? And what makes you think you can't hook up the LLM with some algorithms or layers that handle some of the rest behavior of what your brain does?
datpiff 4y ago
> According to NVidia, LLM sizes have been increasing 10X per year for the last few years. Clearly this cannot continue, as the training costs will exceed all the compute capacity in existence. The other limit is training data, eventually you run out of cheap sources.
- Ambolia 4y ago
  They may also run out of data if they already consumed most of the internet. Or start producing so much of the internet's content that LLMs start consuming what they write in a closed loop.
  - warkdarrior 4y ago
    Sure, but they do not need to grow infinitely more, they just need to grow sufficiently to perform better than 90% of humans.
- saiojd 4y ago
  I would not be so sure about compute capacity? Neural network architectures are still in their infancy, it is very likely that more efficient approaches exist.
  - datpiff 4y ago
    Sure, but the small & efficient networks so far have been pretty poor. Edge AI was hugely hyped a few years ago and it petered out. All of the big tech companies seems to be chasing the biggest networks with the largest training sets, that seems to be the direction for now. There could be huge efficiency savings within the implementations but when training costs are already this high it seems naive to think the low-hanging fruit is still there.
    - saiojd 4y ago
      I wouldn't be so confident. Flash attention is a recent, significant improvement to training times and sure looks low-hanging now. I was not familiar with Edge AI, interesting concept. I feel like improving the efficiency of very large models is much more likely. The recent successes will lead to an even larger influx of $ in the short term --- this is an existential threat to Google, after all. We will see where things go!

sebzim4500 4y ago

The number of people saying it now is many orders of magnitude more than the number of people saying it 10 years ago. Not saying that means it will happen, but it isn't the same situation.

lxgr 4y ago
> The number of people saying it Popular sentiment is a pretty meaningless metric for predicting the future.
- AgentME 4y ago
  Then we shouldn't use the 10 year-old (lesser) popular sentiment as a reason to discount AI either.
  - lxgr 4y ago
    That still doesn’t work. A reliably wrong predictor is still valuable (just negate the output). I‘m arguing public sentiment will be right 50% of the time (in difficult binary predictions), not 0%.

abudabi123 4y ago

The A.I. Monte Carlo Ray Tracing Hallucination Engine can change society by showing as-is and to-be next state. Two use-cases: new infrastructure installation or upgrade and time interval inspection tracking ideal-case vs real world condition. Suppose a lazy contractor skips steps and cuts corners, or a pathological contractor builds the thing and pulls the thing apart over and over again when all that was needed was a one and done, or the change is for the worse. A civil engineer can walk around with an iPad and iPhone to check-in the master plan.

Jeff_Brown 4y ago
You mean AI can detect when a plumber is bulshitting me? That sounds great, could you elaborate?
- coldtea 4y ago
  Given a few sensors that could be installed on the house's pipes, one can imagine several ways an AI can check that all is good. Same for checking blueprints for quality/validity and purchases the plumber says are "needed"...
majormajor 4y ago
You need to replace the civil engineer there (in today's world this is an inspector, I don't know for sure if they are professional engineers or not) for it to be useful, but you still need to have someone not working for the contractor who's incentivized to fake it. The trouble with many of those construction examples is that they're point in time. Installer fucks up how the windows are installed re: flashing and water-proofing, but then puts the siding over it... the error is now completely hidden. You could automated the inspection by AI photo analysis of every single window install on the project, say - but we could already do that for the inspection vs sending someone out, and send the photos to the expert instead, and we don't. Whether that's for lack of incentive to improve, or for ability to go deeper out of distrust for builders? I don't know.

HervalFreire 4y ago

What they said 10 years ago was correct. It did hit society like a sledge hammer. Machine learning basically took over the AI space and penetrated the consumer space with applications that were all but impossible in the previous decade. There's AI chips in smart phones now. What you're seeing here with LLMs is sledge hammer number 2. It's understandable how most people don't notice the sledge hammer. The decade prior to 2010 there was another sledge hammer had no smart phones. We were hit with a smart phone hammer AND an AI sledge hammer and the integration was so seamless we didn't even notice. Much of the same will happen with LLMs. In 5 years it's so normal, nobody cares and likely we will forget what life was like before LLMs.

anon7725 4y ago
Sledge hammer #1 (voice assistants, AI chips in phones) didn’t cause unemployment. It was at the level of new features and capabilities. Sledge hammer #2 is aimed squarely at “white collar” work without much in the way of bounds to its capabilities.
- Jeff_Brown 4y ago
  We can't trust it. That's a pretty hard bound.
  - HervalFreire 4y ago
    For article writing verification is quick, I don't need to worry about a "bug" in the generated article as I do with code. For art, verification is instant, a bad generation is rejected. Trust is only a bound for certain areas, and this bound is eroding.
  - anon7725 4y ago
    Consider that a lot of useful work involves summarization (search++). “What work has gone on for supply chain improvement in our European operations this year?” - this is the kind of question that is easy to ask in natural language but might take someone a week of searching, messaging, etc to assemble. An LLM with access to all of the code, documents, chats, etc could just give an answer with citations. We are not betting $1B on the answer that it gives, but it has saved us 1 week of work and allows us to move on to the next step in the project. There are plenty of tasks like this which are highly valuable yet don’t require high trust. The one-shot “what is the airspeed of a coconut-laden swallow” type questions are actually fairly rare.