4 ms·

>Furthermore, the fact that LLMs seem to need such a stupendous amount of data to get such mediocre reasoning indicates that they simply are not generalizing. I

>Furthermore, the fact that LLMs seem to need such a stupendous amount of data to get such mediocre reasoning indicates that they simply are not generalizing. If these models can’t get anywhere close to human level performance with the data a human would see in 20,000 years, we should entertain the possibility that 2,000,000,000 years worth of data will be also be insufficient. There’s no amount of jet fuel you can add to an airplane to make it reach the moon. Never thought about it in this sense. Is he wrong?

gchamonlive 3y ago
I don't think he is wrong. I also don't think the goal of LLMs is to reproduce human intelligence. That is, we don't need human-like inteligence in a box for a tool to be useful. So this assertion could be right and still miss the point of this tech in my opinion. Edit: to expand, if the goal is AGI then yes we need all the help we can get. But even so, AGI is in a totally different league compared to human intelligence, they might as well be a different species.
- jpk 3y ago
  The context of the fine article is scaling LLMs into AGI. It's not about whether the tool is useful or not, as usefulness is a threshold well before AGI. Some folks are spooked that LLMs are a few optimizations away from the singularity, and the article just discusses some reasons why that probably isn't the case.
  - gchamonlive 3y ago
    The article is really good! I was responding to "is he wrong" part of the comment, not the article itself.
- tremarley 3y ago
  We don’t need human-like intelligence in a box for a tool to be useful, But human-like intelligence is what many companies are spending billions to try and achieve
- dartos 3y ago
  This. I don’t think LLMs are anywhere near sci-fi AGI (think I, robot) It’s such a vague term anyway, AGI. LLMs provide some really nice text generation, summarization, and outstanding semantic search. It’s drop dead easy to make a natural language interface to anything now. That’s a big deal. That’s what’s going to give this tech it’s longevity, imo.
barrenko 3y ago
he's not wrong, and yet he's not right.
gitfan86 3y ago
Over the past year there have been advances in making models smaller while keeping performance high. So if that continues then he is wrong unless he is defining LLMs in a strict way that does not include new improvement in the future
- viraptor 3y ago
  For an example, the diagrams in the post compare the big gpts, but looking at the number of tokens PHI-2 sits below gpt3. And it still beats it in Humaneval and a few other benchmarks.
- Xelynega 3y ago
  It's not about the size of the models, it's about the size of the training data. Humans are able to begin to generalize with a single persons experiences over less than a year, so the fact that LLMs cannot with billions of person-years of information could be an indicator of their inability to generalize no matter how much training data you throw at it.
az226 3y ago
LLMs are closer to discoveries on the spectrum than inventions. Nobody predicted or planned the many emergent capabilities we’ve seen. Almost like magic. Now is a period of moving along the axis to invention with many intentional design, architecture, and feature development alongside testing and evaluation. We are far from done with LLMs, plenty of room for many more discoveries, lots to explore. It’s definitely a precursor to AGI. They offer a platform to build and scale data sets and test beds. We haven’t had ML models this large before. There’s innovation in architecture but we often come back to the bitter lesson: more data. We’re likely going to see experimentation with language models to learn from few examples. Fine tuning pretrained LLMs shows they have quite a remarkable ability to learn from few examples. Liquid AI has a new learning architecture for dynamic learning and much smaller models. Some people seem mad about the bitter lesson, they want their model based on human features to work better when so far usually more data wins. I think the next evolution here is in increasing the quality of the training data and giving it more structure. I suspect the right setup can seed emergent capabilities.
- gitfan86 3y ago
  The trick is to make many LLMs work together in feedback loops. Some small some big. That will get us to what was previously known as AGI. The definition of AGI will change, but we will have systems that put perform humans in most ways.
  - nnoremap 3y ago
    Isaiah 7:14 (NIV): "Therefore the Lord himself will give you a sign: The virgin will conceive and give birth to a son, and will call him Immanuel."
    - esafak 3y ago
      "The Shoe Is The Sign!" -- Monty Python.
- beardedwizard 3y ago
  > It’s definitely a precursor to AGI. What are you basing this claim on? There is no intelligence in an LLM, only humans fooled by randomness.
auggierose 3y ago
And yet, we reached the moon, and I would say airplanes were a necessary step on the way, even if only for psychological reasons. For airplanes we had at least an example in nature, birds. But I am not aware of any animal that travelled from earth to the moon on its own, except us.
- manojlds 3y ago
  We are talking of LLMs, not whether we will be able to reach AGI or not.
  - ImHereToVote 3y ago
    Airplanes in this analogy are essentially the collection of matrix multiplications that emulate reasoning in a very rough but useful manner in an LLM. It's unclear whether a rocket ship is a multimodal neural net. Or some sort of swarm of LLM's in an adversarial relationship, or something completely novel. Regardless, we might be as far between LLM's to ASI's, as airplanes are to rocket ships. Or not.
- Eddy_Viscosity2 3y ago
  But we didn't use airplanes to get there. It needed a new approach, different propulsion, different fuel, different attitude control, etc. etc. LLM may be a necessary step to get to AGI, but it (probably) won't be the one that achieves that goal.
  - auggierose 3y ago
    I doubt that LLMs will give us AGI. But they have already given us more intelligence from a computer than I would have imagined to see during my lifetime.
  - sbierwagen 3y ago
    I mean, if you go through the Apollo program contractors it's a who's who of aerospace. Boeing, North American Aviation, Grumman Aircraft, McDonnell Aircraft Corporation, Bell Aerosystems, Rocketdyne... Electrical parts ran at aviation-standard 400hz. Aviation gyroscopes and aviation instruments. Structural parts made of aviation aluminum alloys. Astronauts that are all airplane test pilots. I can imagine doing Apollo from complete scratch (using car manufacturers that have to invent aluminum-handling tech starting from nothing) but it would have taken a lot more than the decade Apollo took.
MPSimmons 3y ago
I don't think the data is the weakness. We're using Transformer architecture right now. There's no reason there won't be further discoveries in AI that are as impactful as "Attention is All You Need". We may be due for another "AI Winter" where we don't see dramatic improvement across the board. We may not. Regardless, LLMs using the Transformer architecture may not have human level intelligence, but they _are_ useful, and they'll continue to be useful. In the 90s, even during the AI winter, we were able to use Bayesian classification for such common tasks as email filtering. There's no reason we can't continue to use Transformer architecture LLMs for common purposes too. Content production alone makes it worth while. We don't _need_ AGI, it just seems like the direction we are heading as a species. If we don't get there, it's fine. No need to throw the baby out with the bath water.
Der_Einzige 3y ago
Even the largest LLM has had less "total information" than most humans take in through all of their senses over their lifetime. A single day for a baby is taking in a continuous stream of among other things high quality video and audio and does a large amount of processing on that. Much of that for very young babies is unsupervised learning (clustering), where baby learns that object A and object B are different despite knowing nothing else about their properties. Humans can learn using every ML learning paradigm in ever modality: unsupervised, self-supervised, semi-supervised, supervised, active, reinforcement based, and anything else I might be missing. Current LLMs are stuck with "self-supervised" with the occasional reinforced (RLHF) or supervised (DPO) cherry on top at the end. non multi-modal LLMs operate with one modality. We are hardly scratching the surface on what's possible with multi-modal LLMs today. We are hardly scratching the surface for training data for these models. The overwhelming majority of todays LLMs are vastly undertrained and exhibit behavior of undertrained systems. The claim from the OP about scale not giving us further emergent properties flies in the face of all of what we know about this field. Expect further significant gains despite nay-sayers claiming it's impossible.
- haltist 3y ago
  You are obviously a believer so you should know I know how to build AGI with a patented and trademarked architecture called "panoptic computronium cathedral"™. Tell all your friends about it. I only need $80B to achieve AGI.
lumost 3y ago
The Phi paper and various approaches to distilling from GPT-4 demonstrate that the training data and plausibly order of presentation matter. The challenge is that we both do not understand which set of data is most beneficial for training, or how it could be efficiently ordered without triggering computationally infeasible problems. However we do know how to massively scale up training.
espadrine 3y ago
Demis Hassabis of Deepmind echoes a similar sentiment[0]: > I still think there are missing things with the current systems. […] I regard it a bit like the Industrial Revolution where there was all these amazing new ideas about energy and power and so on, but it was fueled by the fact that there were dead dinosaurs, and coal and oil just lying in the ground. Imagine how much harder the Industrial Revolution would have been without that. We would have had to jump to nuclear or solar somehow in one go. [In AI research,] the equivalent of that oil is just the Internet, this massive human-curated artefact. […] And of course, we can draw on that. And there's just a lot more information there, I think, it turns out than any of us can comprehend, really. […] [T]here's still things I think that are missing. I think we're not good at planning. We need to fix factuality. I also think there's room for memory and episodic memory. [0]: https://cbmm.mit.edu/video/cbmm10-panel-research-intelligence-age-ai https://cbmm.mit.edu/video/cbmm10-panel-research-intelligenc...
- skippyboxedhero 3y ago
  His view of the Industrial Revolution is completely wrong. Societies pre-IR had multiple periods where energy usage increased significantly, some of them based specifically around coal. No IR. Early IR was largely based around the usage of water power, not coal. IR was pure innovation, people being able to imagine and create the impossible, it was going straight to nuclear already. Ironically, someone who is an innovator believes the very anti-innovation narrative of the IR (very roughly, this is the anti-Eurocentric stuff that began appearing in the 2000s...the world has moved on since then as these theories are obviously wrong). Nothing tells you more about how busted modern universities are than this fact.
  - archon1410 3y ago
    Has the narrative moved on? The historian and blogger Bret Devereaux presents a view on a 2022 blog post that seems to back up what the Deepmind CEO is saying. > The specificity matters here because each innovation in the chain required not merely the discovery of the principle, but also the design and an economically viable use-case to all line up in order to have impact. https://acoup.blog/2022/08/26/collections-why-no-roman-industrial-revolution/ https://acoup.blog/2022/08/26/collections-why-no-roman-indus...

gchamonlive 3y ago

I don't think he is wrong. I also don't think the goal of LLMs is to reproduce human intelligence. That is, we don't need human-like inteligence in a box for a tool to be useful. So this assertion could be right and still miss the point of this tech in my opinion. Edit: to expand, if the goal is AGI then yes we need all the help we can get. But even so, AGI is in a totally different league compared to human intelligence, they might as well be a different species.

jpk 3y ago
The context of the fine article is scaling LLMs into AGI. It's not about whether the tool is useful or not, as usefulness is a threshold well before AGI. Some folks are spooked that LLMs are a few optimizations away from the singularity, and the article just discusses some reasons why that probably isn't the case.
- gchamonlive 3y ago
  The article is really good! I was responding to "is he wrong" part of the comment, not the article itself.
tremarley 3y ago
We don’t need human-like intelligence in a box for a tool to be useful, But human-like intelligence is what many companies are spending billions to try and achieve
dartos 3y ago
This. I don’t think LLMs are anywhere near sci-fi AGI (think I, robot) It’s such a vague term anyway, AGI. LLMs provide some really nice text generation, summarization, and outstanding semantic search. It’s drop dead easy to make a natural language interface to anything now. That’s a big deal. That’s what’s going to give this tech it’s longevity, imo.

barrenko 3y ago

he's not wrong, and yet he's not right.

gitfan86 3y ago

Over the past year there have been advances in making models smaller while keeping performance high. So if that continues then he is wrong unless he is defining LLMs in a strict way that does not include new improvement in the future

viraptor 3y ago
For an example, the diagrams in the post compare the big gpts, but looking at the number of tokens PHI-2 sits below gpt3. And it still beats it in Humaneval and a few other benchmarks.
Xelynega 3y ago
It's not about the size of the models, it's about the size of the training data. Humans are able to begin to generalize with a single persons experiences over less than a year, so the fact that LLMs cannot with billions of person-years of information could be an indicator of their inability to generalize no matter how much training data you throw at it.

az226 3y ago

LLMs are closer to discoveries on the spectrum than inventions. Nobody predicted or planned the many emergent capabilities we’ve seen. Almost like magic. Now is a period of moving along the axis to invention with many intentional design, architecture, and feature development alongside testing and evaluation. We are far from done with LLMs, plenty of room for many more discoveries, lots to explore. It’s definitely a precursor to AGI. They offer a platform to build and scale data sets and test beds. We haven’t had ML models this large before. There’s innovation in architecture but we often come back to the bitter lesson: more data. We’re likely going to see experimentation with language models to learn from few examples. Fine tuning pretrained LLMs shows they have quite a remarkable ability to learn from few examples. Liquid AI has a new learning architecture for dynamic learning and much smaller models. Some people seem mad about the bitter lesson, they want their model based on human features to work better when so far usually more data wins. I think the next evolution here is in increasing the quality of the training data and giving it more structure. I suspect the right setup can seed emergent capabilities.

gitfan86 3y ago
The trick is to make many LLMs work together in feedback loops. Some small some big. That will get us to what was previously known as AGI. The definition of AGI will change, but we will have systems that put perform humans in most ways.
- nnoremap 3y ago
  Isaiah 7:14 (NIV): "Therefore the Lord himself will give you a sign: The virgin will conceive and give birth to a son, and will call him Immanuel."
  - esafak 3y ago
    "The Shoe Is The Sign!" -- Monty Python.
beardedwizard 3y ago
> It’s definitely a precursor to AGI. What are you basing this claim on? There is no intelligence in an LLM, only humans fooled by randomness.

auggierose 3y ago

And yet, we reached the moon, and I would say airplanes were a necessary step on the way, even if only for psychological reasons. For airplanes we had at least an example in nature, birds. But I am not aware of any animal that travelled from earth to the moon on its own, except us.

manojlds 3y ago
We are talking of LLMs, not whether we will be able to reach AGI or not.
- ImHereToVote 3y ago
  Airplanes in this analogy are essentially the collection of matrix multiplications that emulate reasoning in a very rough but useful manner in an LLM. It's unclear whether a rocket ship is a multimodal neural net. Or some sort of swarm of LLM's in an adversarial relationship, or something completely novel. Regardless, we might be as far between LLM's to ASI's, as airplanes are to rocket ships. Or not.
Eddy_Viscosity2 3y ago
But we didn't use airplanes to get there. It needed a new approach, different propulsion, different fuel, different attitude control, etc. etc. LLM may be a necessary step to get to AGI, but it (probably) won't be the one that achieves that goal.
- auggierose 3y ago
  I doubt that LLMs will give us AGI. But they have already given us more intelligence from a computer than I would have imagined to see during my lifetime.
- sbierwagen 3y ago
  I mean, if you go through the Apollo program contractors it's a who's who of aerospace. Boeing, North American Aviation, Grumman Aircraft, McDonnell Aircraft Corporation, Bell Aerosystems, Rocketdyne... Electrical parts ran at aviation-standard 400hz. Aviation gyroscopes and aviation instruments. Structural parts made of aviation aluminum alloys. Astronauts that are all airplane test pilots. I can imagine doing Apollo from complete scratch (using car manufacturers that have to invent aluminum-handling tech starting from nothing) but it would have taken a lot more than the decade Apollo took.

MPSimmons 3y ago

I don't think the data is the weakness. We're using Transformer architecture right now. There's no reason there won't be further discoveries in AI that are as impactful as "Attention is All You Need". We may be due for another "AI Winter" where we don't see dramatic improvement across the board. We may not. Regardless, LLMs using the Transformer architecture may not have human level intelligence, but they _are_ useful, and they'll continue to be useful. In the 90s, even during the AI winter, we were able to use Bayesian classification for such common tasks as email filtering. There's no reason we can't continue to use Transformer architecture LLMs for common purposes too. Content production alone makes it worth while. We don't _need_ AGI, it just seems like the direction we are heading as a species. If we don't get there, it's fine. No need to throw the baby out with the bath water.

Der_Einzige 3y ago

Even the largest LLM has had less "total information" than most humans take in through all of their senses over their lifetime. A single day for a baby is taking in a continuous stream of among other things high quality video and audio and does a large amount of processing on that. Much of that for very young babies is unsupervised learning (clustering), where baby learns that object A and object B are different despite knowing nothing else about their properties. Humans can learn using every ML learning paradigm in ever modality: unsupervised, self-supervised, semi-supervised, supervised, active, reinforcement based, and anything else I might be missing. Current LLMs are stuck with "self-supervised" with the occasional reinforced (RLHF) or supervised (DPO) cherry on top at the end. non multi-modal LLMs operate with one modality. We are hardly scratching the surface on what's possible with multi-modal LLMs today. We are hardly scratching the surface for training data for these models. The overwhelming majority of todays LLMs are vastly undertrained and exhibit behavior of undertrained systems. The claim from the OP about scale not giving us further emergent properties flies in the face of all of what we know about this field. Expect further significant gains despite nay-sayers claiming it's impossible.

haltist 3y ago
You are obviously a believer so you should know I know how to build AGI with a patented and trademarked architecture called "panoptic computronium cathedral"™. Tell all your friends about it. I only need $80B to achieve AGI.

lumost 3y ago

The Phi paper and various approaches to distilling from GPT-4 demonstrate that the training data and plausibly order of presentation matter. The challenge is that we both do not understand which set of data is most beneficial for training, or how it could be efficiently ordered without triggering computationally infeasible problems. However we do know how to massively scale up training.

espadrine 3y ago

Demis Hassabis of Deepmind echoes a similar sentiment[0]: > I still think there are missing things with the current systems. […] I regard it a bit like the Industrial Revolution where there was all these amazing new ideas about energy and power and so on, but it was fueled by the fact that there were dead dinosaurs, and coal and oil just lying in the ground. Imagine how much harder the Industrial Revolution would have been without that. We would have had to jump to nuclear or solar somehow in one go. [In AI research,] the equivalent of that oil is just the Internet, this massive human-curated artefact. […] And of course, we can draw on that. And there's just a lot more information there, I think, it turns out than any of us can comprehend, really. […] [T]here's still things I think that are missing. I think we're not good at planning. We need to fix factuality. I also think there's room for memory and episodic memory. [0]: https://cbmm.mit.edu/video/cbmm10-panel-research-intelligence-age-ai https://cbmm.mit.edu/video/cbmm10-panel-research-intelligenc...

skippyboxedhero 3y ago
His view of the Industrial Revolution is completely wrong. Societies pre-IR had multiple periods where energy usage increased significantly, some of them based specifically around coal. No IR. Early IR was largely based around the usage of water power, not coal. IR was pure innovation, people being able to imagine and create the impossible, it was going straight to nuclear already. Ironically, someone who is an innovator believes the very anti-innovation narrative of the IR (very roughly, this is the anti-Eurocentric stuff that began appearing in the 2000s...the world has moved on since then as these theories are obviously wrong). Nothing tells you more about how busted modern universities are than this fact.
- archon1410 3y ago
  Has the narrative moved on? The historian and blogger Bret Devereaux presents a view on a 2022 blog post that seems to back up what the Deepmind CEO is saying. > The specificity matters here because each innovation in the chain required not merely the discovery of the principle, but also the design and an economically viable use-case to all line up in order to have impact. https://acoup.blog/2022/08/26/collections-why-no-roman-industrial-revolution/ https://acoup.blog/2022/08/26/collections-why-no-roman-indus...