I'd like to be wrong but I can't help but feel that people predicting a revolution are making the same, understandable mistake as my hypothetical 1950s person.
Productivity at information communication tasks just isn’t the entire economy.
I think we are massively more productive. Some of the biggest new companies are ad companies (Google, Facebook), or spend a ton of their time designing devices that can’t be modified by their users (Apple, Microsoft). Even old fashioned companies like tractor and train companies have time to waste on preventing users from performing maintenance. And then the economy has leftover effort to jailbreak all this stuff.
We’re very productive, we’ve just found room for unlimited zero or negative sum behavior.
For sure - I grew up in the mid-late 70s having to walk to the library to research stuff for homework, parents having to use the yellow-pages to find things, etc.
Maybe smartphones are more of a game changer than desk-bound internet though - a global communication device in your pocket that'll give you driving directions, etc, etc.
BUT ... does the world really FEEL that different now, than pre-internet? Only sort-of - more convenient, more connected, but not massively different in the ways that I imagine other inventions such as industrialization, electricity, cars may have done. The invention of the telephone and radio maybe would have felt a bit like the internet - a convenience that made you feel more connected, and maybe more startling being the first such capability?
The administration of the Tax Service uses 4% of the total tax revenue it generates. This percentage has stayed relatively fixed over time.
If IT really improved productivity, wouldn't you expect that that number would decrease, since Tax Administration is presumably an area that we should expect to see great gains from computerisation?
We should be able to do the same amount of work more efficiently with IT, thus decreasing the percentage. If instead the efficiency frees up time allowing more work to be done (because there are people dodging taxes and we need to discover that), then you should expect the amount of tax to increase relatively which should also cause the percentage to decrease.
Therefore IT has not increased productivity.
Either it doesn't do so directly, or it does do so directly, but all the efficiency gains are immediately consumed by more useless beurocracy.
https://archive.ph/baneA https://archive.ph/TrHYN
“Our central theme is that computers and the Internet do not measure up to the Great Inventions of the late nineteenth and early twentieth century, and in this do not merit the label of Industrial Revolution,”
— Robert Gordon, actual economist
On their current trajectory LLMs are just expert systems that will let certain types of simple job be automated. A potential productivity amplifier similar to having a personal assistant that you can assign tasks too. Handy (more so for people doing desk-bound jobs than others), but not a game changer.
An AGI far beyond human capability could certainly accelerate scientific advance and let us understand the world (e.g. how to combat climate change, how to address international conflicts, how to handle pandemics) so be very beneficial, but what that would feel like to us is hard to guess. We get used to slowly introduced (or even not so slowly) changes very quickly and just accept them, even though today's tech would look like science fiction 100 years ago.
What would certainly be a game changer, and presumably will eventually come (maybe only in hundreds of years?) would be if humans eventually relinquish control of government, industry, etc to AGIs. Maybe our egos will cause us to keep pretending we're in control - we're the ones asking the oracle, we could pull the plug anytime (we'll tell ourselves) etc, but it'll be a different world if all the decisions are nonetheless coming from something WAY more intelligent than ourselves.
Human civilization has accumulated many layers of systems since then and the internet changed all of them to the point that many are barely recognizable. Just ask someone who's been in prison since before the internet was a thing - there are plenty of them! They have extreme difficulty adapting to the outside world after they've been gone for forty or fifty years.
I see "AI safety" brought up as a laughable attempt at stopping the progress of LLMs, when in reality the people talking about "AI safety" are the people trying to say that the majority will not benefit from this technology.
When agents start being more reliable I think we will start seeing applications we couldn’t possibly anticipate today
I'd call the current generation "Social media natives", because that is the biggest difference from the previous generation. 90s kids grew up with games and communication, but they were free from facebook, youtube and instagram.
seems odd. What 'patterns' and 'limitations' do you still see? Because I see so much has changed.
My mom, who is 70 years old, regularly tells me how profoundly transformative the internet has been for society.
Food distribution has improved as have most logistics in general. These efficiencies have somewhat been shared with the general public but in a lot of cases, those gains were captured by private enterprise.
Healthcare has improved a little bit, iterative progress can be made more quickly, shared, and moved into translational medicine as practice. Drug discovery has improved quite a bit, as have logistics around getting said drugs in the hands of people who need them and doing so affordably. This improved lives and longevity.
Socially we can communicate far easier. It remains to be seen to me if thise is always an improvement. Humans seem to be designed for much smaller social circles and don't seem to be capable of taking much advantage in their daily lives of increases frequency, scale, and reach of socialization.
The list goes on. It's not exactly linearly correlated with technology growth because ultimately it boils down to actionable information. Just because we have more information or more processing capability around information doesn't mean we get direct returns from that or that we don't reach limits where we simply don't have use for the additional gains. Information has to be actionable in some way, otherwise it's just intermediate data products that may or may not benefit us. I know can ready daily news from some small town in Southern Japan if I wanted to. That doesn't improve my life mostly, but it's there.
We have piles and piles of scientific literature we could share and iterate on towards new discoveries for humanity. That doesn't mean in my daily need for survival and balance with recreation I have time to contribute to things I find interesting or necessary, after all I am to some degree a slave of my needs within the economic system I'm entrenched in. I have bills, I have to earn money, and I have to work.
Even if that wasn't the case maybe or maybe not would I be able to contribute more back to society than I do now at my paid profession. Currently I'd say I do pretty well in this department in terms of reach. Without that I might struggle.
by what criteria do you see the world as the same today vs 70 years ago?
Things "behind the scenes" have perhaps changed a lot -- e.g. financialization, more competitive markets, explosion of communication options, which are the driving force behind those visible changes.
- people eat plants and animals
- people pay money for goods and services
- there are countries, sometimes they fight, sometimes they work together
- men and women come together to create children, and often raise those children together
etc, etc, etc
The “bones” of what make up a capital-S Society are pretty much the same. None of these things had to stay the same, but they have so far.
With AI, the way I see it, it is just virtual other people. Of course, a bit stranger but more simillar than you think.
I've done many language exchanges with people using Google Translate and the lack of improvement/memory of past conversations is a real motivation killer; I'm concerned this will move on to general discourse on the internet with the proliferation of LLMs.
I'm sure many people have already gone around in circles with rules-based customer support. AI can make this worse.
We had designed many tools that beat us in various aspects. This is an invalid analogy.
If it’s closer to a midpoint between GPT-4 and true human intelligence, then sure, I agree with you, it’s a significant change to society but not an overhaul. But if it’s actually a human level (or better) general intelligence, it’ll be the biggest change to human society maybe ever.
I bet their reaction would be a facepalm.
Nothing I’ve seen or learned about LLMs leads me to believe that LLMs are in fact a pathway to AGI.
LLMs trained on more data with more efficient algorithms will make for more interesting tools built with LLMs, but I don’t see this technology as a foundation for AGI.
LLMs don’t “reason” in any sense of the word that I understand and I think the ability to reason is table stakes for AGI.
I'm not sure you realize this, but that is literally what this article was written to explore!
I feel like you just autocompleted what you believe about large language models in this thread, rather than engaging with the article. Engagement might look like "I hold the skeptic position because of X, Y, and Z, but I see that the other position has some really good, hard-to-answer points."
Instead, we just got the first thing that came to your mind talking about AI.
In fact, am I talking to a person?
Yeah but it's "exploration" answers all the reasonable objections by just extrapolating vague "smartness" (EDITED [1]). "LLMs seem smart, more data will make 'em smarter..."
If apparent intelligence were the only measure of where things are going, we could be certain GPT-5 or whatever would reach AGI. But I don't many people think that's the case.
The various critics of LLMs like Gary Marcus make the point that while LLMs increase in ability each iteration, they continue to be weak in particular areas.
My favorite measure is "query intelligence" versus "task accomplishment intelligence". Current "AI" (deep learning/transformers/etc) systems are great at query intelligence but don't seem to scale in their "task accomplishment intelligence" at the same rate. (Notice "baby AGI", ChatGPT+self-talk, fail to produce actual task intelligence).
[1] Edited, original "seemed remarkably unenlightening. Lots of generalities, on-the-one-hand-on-the-other descriptions". Actually, reading more closely the article does raise good objections - but still doesn't answer them well imo.
The point I was trying to make is that I think better LLMs won’t lead to AGI. The article focused on the mechanics and technology, but I feel that’s missing the point.
The point being, AGI is not going to be a direct outcome of LLM development, regardless of the efficiency or volume of data.
When we get to the point where LLMs are able to invoke these tools for a user, even if that user has no knowledge of them, and are able to translate the results of that reasoning back into the user's context... That'll start to smell like AGI.
The other piece, I think, is going to be improved cataloging of human reasoning. If you can ask a question and get the answer that a specialist who died fifty years ago would've given you because that specialist was a heavy AI user and so their specialty was available for query... That'll also start to smell like AGI.
The foundations have been there for 30 years, LLMs are the paint job, the door handles, and the windows.
I think OP meant other definition of reason, because by your definition calculator can also reason. These are tools created by humans, that help them to reason about stuff by offloading calculations for some of the tasks. They do not reason on their own and they can't extrapolate. They are expert systems.
> But of course we don’t actually care directly about performance on next-token prediction. The models already have humans beat on this loss function. We want to find out whether these scaling curves on next-token prediction actually correspond to true progress towards generality.
And:
> Why is it impressive that a model trained on internet text full of random facts happens to have a lot of random facts memorized? And why does that in any way indicate intelligence or creativity?
And:
> So it’s not even worth asking yet whether scaling will continue to work - we don’t even seem to have evidence that scaling has worked so far.
I don’t feel like we’re anywhere close given that we can’t even yet meaningfully define reasoning or consciousness… or as another commenter put it, what is it that differentiates us so significantly from other animals.
Never thought about it in this sense. Is he wrong?
> 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-intelligenc...
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.
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.
LLM may be a necessary step to get to AGI, but it (probably) won't be the one that achieves that goal.
Tardigrades might :)
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.
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.
"Are Emergent Abilities of Large Language Models a Mirage?"
https://arxiv.org/abs/2304.15004 https://blog.neurips.cc/2023/12/11/announcing-the-neurips-20...
What are you basing this claim on? There is no intelligence in an LLM, only humans fooled by randomness.
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
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.
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.
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.
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.
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.
Obviously scaling works in general, just ask anyone in HPC, haha.
I think Wittgenstein's ideas are pertinent to the discussion of the relation of language to intelligence (or reasoning in general). I don't meant this in a technical sense (I recall Chomsky mentioning that almost no ideas from Wittgenstein actually have a place in modern linguistics) but from a metaphysical sense (Chomsky also noted that Wittgenstein was one of his formative influences).
The video I linked is a worthy introduction and not too long so I recommend it to anyone interested in how language might be the key to intelligence.
My personal take, when I see skeptics of LLMs approaching AGI, is that they implicitly reject a Wittgenstein view of metaphysics without actually engaging with it. There is an implicit Cartesian aspect to their world view, where there is either some mental aspect not yet captured by machines (a primitive soul) or some physical process missing (some kind of non-language system).
Whenever I read skeptical arguments against LLMs they are not credibly evidence based, nor are they credibly theoretical. They almost always come down to the assumption that language alone isn't sufficient. Wittgenstein was arguing long before LLMs were even a possibility that language wasn't just sufficient, it was inextricably linked to reason.
What excites me about scaling LLMs, is we may actually build evidence that supports (or refutes) his metaphysical ideas.
1. https://www.youtube.com/watch?v=v_hQpvQYhOI&ab_channel=Philo...
Without the ability to interact with the physical world, LLMs will never be able to reach AGI. It can kind of simulate it to some extent, but it'll never get there
LLMs can't even form their own memories, the context has to be explicitly fed back to them.
That is one of the things that stood out to me in Searle's summary of his later work because I consider how the transformer architecture works and the way in which the surrounding context plays into the meaning of the words.
> Part of learning those language games is experimenting in the real world
It is interesting that the article we are responding to talks about how we have only just begun to experiment with RL on top of transformers. In the same way that Alpha Go engaged in adversarial play we can envision LLMs being augmented to play language games amongst themselves. That may result in their own language, distinct from human language. But it also may result in the formation of intelligence surpassing human intelligence.
> The fidelity of text is simply too low to communicate the amount of information humans use to build up general intelligence.
This does not at all follow from anything I've encountered in Wittgenstein. It is an empirical claim that we (as in humanity) are going to test and not something that I would argue either one of us can know simply reasoning from first principles.
What does follow for me is closer to what Steven Pinker has been proposing in his own critiques of LLMs and AGI, which is that there is no necessary correlation between goal seeking (or morality) and intelligence. I also feel this is concordant with Wittgenstein's own work.
> Without the ability to interact with the physical world, LLMs will never be able to reach AGI
Again, a confident claim that is based on nothing other than your own belief. As I stated in my last comment, I am excited to see if that is empirically true or false. We are definitely going to scale up LLMs in the coming decade and so we are likely to find out.
My suspicion is that people don't want this scaling up to work because it would force them to let go of metaphysical commitments they have on both the nature of intelligence as well as the nature of reality. And for this reason they are adamantly disbelieving in even the possibility before the evidence has been gathered.
I'm happy to stay agnostic until the evidence is in. Thankfully, it shouldn't take too long so I may be lucky enough to find out in my own lifetime.
lmao. I'd like to see you elaborate on what you think this means! The fact that you quote Searle, though, tells me all I need to know.
In his early work he appeared to believe that the way to do this categorization was to determine if the logical propositions corresponded in their structure to the real world (his picture theory of language). In his later work he appears to have believed the way to do this was to reference some "language game" among a set of communicating agents and then understand the words based on their context within that construct (use theory of language).
Both of these concepts do not reference any kind of subjective experience, aesthetics, morality, etc. They are simply a description of how to judge whether or not a proposition is "meaningful". In later Wittgenstein, meaning in language is entirely divorced from any necessity that the statements correspond to an independent objective reality. It is simply necessary that they are consistent amongst the participants in a particular language game.
I don't believe that means the entirety of our "consciousness" is related solely to our ability to categorize logical propositions. However, it may suggest that intelligence specifically is related to this ability.
In as much as we can say that an LLM is capable of participating in some particular human language game and can successfully categorize logical propositions within that language game - I would say that LLM is demonstrating "intelligence" within that language game. And if we can create LLMs that can participate in arbitrary (or general) language games across a wide variety of domains, we might call that LLM generally intelligent. I believe that current LLMs have achieved the first (demonstrating some intelligence in particular domains) but we have yet to achieve the second (demonstrating consistent intelligence in a wide range of general domains).
As for metaphysics, I would argue that Wittgenstein saw this general ability (to categorize logical propositions) as a subset of all possible experience. I believe he saw this categorization activity as the primary aim of philosophy. However, the kinds of experience that were outside of this categorization activity could not be spoken about at all.
I also don't understand the claims that it doesn't generalize. I currently use it to solve problems that I can absolutely guarantee were not in its training set, and it generalizes well enough. I also think that one of the easiest ways to get it to generalize better would simply be through giving it synthetic data which demonstrates the process of generalizing.
It also seems foolish to extrapolate on what we have under the assumption that there won't be key insights/changes in architecture as we get to the limitations of synthetic data wins/multi-modal wins.
You might want to reconsider your stance on emergent abilities in LLMs considering the NeurIPS 2023 best paper winner is titled:
"Are Emergent Abilities of Large Language Models a Mirage?"
https://arxiv.org/abs/2304.15004 https://blog.neurips.cc/2023/12/11/announcing-the-neurips-20...
Also, consider that some work gets a lot of positivity not for the work itself, but for the people who wrote it. Timnit Gebaru's work was effectively ignored until she got famous for her spat with jeff dean at google. Her citations have exploded as a result, and I don't think that most in the field think that the "stochastic parrot" paper was especially good, and certainly not her other papers which include significant amounts of work dedicated to claiming that LLM training is really bad for the environment (despite a single jet taking AI researchers to conferences being worse for the environment than LLM training circa that paper being written was taking). Doesn't matter that the paper was wrong, it's now highly cited because you get brownie points for citing her work in grievance studies influenced subfields of AI.
To me the most interesting aspect of LLMs is the way that they reveal cognitive 0-days in humans.
The human race needs patches to cognitive firmware to deal with predictive text... Which is a fascinating revelation to me. Sure it's backed up by psych analysis for decades but it's interesting to watch it play out on such a large scale.
I don't think what LLMs are currently doing is really generalizing, but rather:
1) Multiple occurrences of something in the dataset are mutually statistically reinforcing. This isn't generalization (abstraction) but rather reinforcement through repetition.
2) Multiple different statistical patterns are being recalled/combined in novel ways such that it seems able to "correctly" respond to things out of dataset, but really this only due to these novel combinations, not due to it having abstracted it's knowledge and applying a more general (or analogical) rule than present in it's individual training points.
Can you share the strongest example?
Maintaining the illusion is important to keep the money flowing in.
But if it were so settled and obvious there would be a clear line of reasoning to make that plain. And there is not. Instead, there is a very vibrant debate on the topic with tons of nuance and good faith (and bad) on each side, if we want to talk about sides.
And, of course, one of the implications of this very real and significant inquiry that needs to be made and that requires real contributions from informed individuals, is that whenever anyone is dismissive or reductive regarding the unresolved difficulties, you can be sure they have absolutely no clue what they are talking about.
In part, whether conscious or not, people see the bright future of LLMs as a kind of redemption for the world so far wrought from a Silicon Valley ideology; its almost too on-the-nose the way chatgpt "fixes" internet search.
But on a deeper level, consider how many hn posts we saw before chatgpt that were some variation of "I have reached a pinnacle of career accomplishment in the tech world, but I can't find meaning or value in my life." We don't seem to see those posts quite as much with all this AI stuff in the air. People seem to find some kind of existential value in the LLMs, one with an urgency that does not permit skepticism or critique.
And, of course, in this thread alone, there is the constant refrain: "well, perhaps we are large language models ourselves after all..." This reflex to crude Skinnerism says a lot too: there are some that, I think, seek to be able to conquer even themselves; to reduce their inner life to python code and data, because it is something they can know and understand and thus have some kind of (sense) of control or insight about it.
I don't want to be harsh saying this, people need something to believe in. I just think we can't discount how personal all this appears to be for a lot of regular, non-AI-CEO people. It is just extremely interesting, this culture and ideology being built around this. To me it rivals the LLMs themselves as a kind fascinating subject of inquiry.
What is experience? We are in state S, and take action A, and observe feedback R. The environment is the teacher, giving us reward signals. We can only increase our knowledge incrementally, by trying our many bad ideas, and sometimes paying with our lives. But we still leave morsels of newly acquired experience for future generations.
We are experience machines, both individually and socially. And intelligence is the distilled experience of the past, encoded in concepts, methods and knowledge. Intelligence is a collective process. None of us could reach our current level without language and society.
Human language is in a way smarter than humans.
What I say is not just regurgitation of my past experiences; there is a logic to it.
LLMs can work as data extraction already, so one can build some prolog DB and update it as it consumes data. Then translate any logic problems into prolog queries. I want to see this in practice.
Similar with usage of logic engines and computation/programs.
I also think that RL can come up with better training function for LLMs. In the programming domain for example one could ask LLM to think about all possible test for given code and evaluate them automatically.
I was also thinking about using diffusER pattern where programming rules are kinda hardcoded (similar to add/replace/delete but instead algebra on functions/variables). Thats probably not AGI path but could be good for producing programs.
1 - Babies learn much more with much less 2 - Video training data can be made in theory at incredible rates
The questions becomes: why is the author focusing on approaches in AI investigated in like 2012? Does the author think SOTA is text only? Are OpenAI or other market leaders only focusing on text? Probably not.
If babies learn much more from much less, isn't that evidence that the LLM approach isn't as efficient as whatever approach humans implement biologically, so it's likely LLM processes won't "scale to ago"?
For video data, that's not how LLMs work(or any NNs for that matter). You have to train them on what you want them to look at, so if you want them to predict the next token of text given an input array, you need to train it on the input arrays and output tokens.
You can extract the data in the form you need from the video content, but presumably that's already been done for the most part, since video transcripts are likely included in the training data for gpt.
a lot of people have tried to replicate this, I have tried. It's very hard to get GPT-4 to draw a unicorn, also asking it to draw an upside down unicorn is even harder.
The author also cited a few human assisted efforts as ML only.
The fact that the author also is surprised that GPT is better at falsifying user input while it struggles at new ideas demonstrates the fact that those who are hyping LLMs as getting us closer to strong AI don't know or ate ignoring the know limitations of problems like automated theorem solving.
I think generative AI is powerful and useful. But the AGI is near camp is starting to make it a hard sell because the general public is discovering the limits and people are trying to force it into inappropriate domains.
Over parameterization and double decent is great at expanding what it can do, but I haven't seen anything that justifies the AGI hype yet.
But - it's only a pre-print and I don't think that they have taken it forward to a publication so handle with care. Anyway, the relevant evidence and thinking that I would highlight and somewhat agrees with my findings is seen in figure 7 and section 5.6. I struggle to get the quality of results that they saw, but its believed by some people that ongoing development of GPT-4 and throttling of the reasoning for cost reasons may have limited some of its capabilities by the end of this year so that may be some of my problem.
I've did this and got a result. Asked for a cat wearing a hat. It drew a circle with dots as eyes and sorta-whiskers with lines and a triangle for the hat. All using just SVG vector code.
The claim is result shows that there is understanding of not just words about cats and hats and connections to shapes but also a bit of spatial awareness in the x,y coords needed on the SVG canvas to place the shapes.
I was able to do with with several different little characters / cartoons in SVG and while it completely fails every now and then it was better than I would have thought.
I think it has been exposed to SVG (via the web) way more than LaTeX vector drawings.
I think Google, Microsoft and facebook could easily have 5 OOM data than the entire public web combined if we just count text. Majority of people don't have any content on public web except for personal photos. A minority has few public social media posts and it is rare for people to write blog or research paper etc. And almost everyone has some content written in mail or docs or messaging.
Quality of data (which I believe is at least part of why synthetic data is being used) can perhaps make more of a difference and perhaps at least partly compensate in a crude way for these models lack of outlier rejection and any generalization prediction-feedback loop. Just feed them consistent correct data in the first place.
I’m worried that when people hear ‘5 OOMs off’, how they register it is, “Oh we have 5x less data than we need - we just need a couple of 2x improvements in data efficiency, and we’re golden”. After all, what’s a couple OOMs between friends?
No, 5 OOMs off means we have 100,000x less data than we need.
100,000x is only seventeen 2x improvements. This[1] says world volume of data doubles every two years, citing a 2016 McKinsey study as the source. That puts it 34 years away. 2056. This[2] says the estimated compound growth rate of data creation is around 61%, that's a doubling time of 1.32 years and puts it 22 years away. 2045.
It's not tomorrow, but it may not be all that far away.
[1] https://rivery.io/blog/big-data-statistics-how-much-data-is-...
[2] https://theconversation.com/the-worlds-data-explained-how-mu...
But every article/post of this kind immediately begs the question; What is AGI? I have yet to hear even a decent standard.
It always seems like I'm reading Greek philosophers struggling with things they have almost no understanding of and just throwing out the wildest theories.
Honestly, it raised my opinion of them seeing how hard it is to reason about things which we have no grasp of.
Assuming they have effectively a log of the internet, rather than counting the current state of the internet as usable data we should be thinking about the list of diffs that make up the internet.
Maybe this ends up like Millenium Management where a key differentiator is having access to deleted datasets.
"Faith and Fate: Limits of Transformers on Compositionality"
https://arxiv.org/abs/2305.18654
"Comparing Humans, GPT-4, and GPT-4V On Abstraction and Reasoning Tasks":
https://arxiv.org/abs/2311.09247
"Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve"
https://arxiv.org/abs/2309.13638
"Pretraining Data Mixtures Enable Narrow Model Selection Capabilities in Transformer Models"
You can feed your dog 100,000 times a day, but that won't make it a 1,000kg dog. The whole idea that AGI can be achieved by predicting the next word is just pure marketing nonsense at best.
Is there any existing cases of LLMs coming up with novel, useful and namely better architectures? Either related to AI/ML itself or any other field.
The entire subject of the article is concerned with what it will take and how likely it is than an AI will ever will able to generate improvements like this.
In my view, domain modeling, managing state, knowing when to transition between states, techniques for final decision making, consideration for the time domain, and prompt engineering are the real challenges.
It's a fact of neural networks that to train them supervised you need the training data in the expected input for(vector of n thousand preceding tokens for LLMs) with the expected output(the next token for LLMs). "Training them on video" would mean converting the video to a format we can train the llm with, then training the LLM with that info.
This would probably be a 1 OOM increase at maximum, if the video transcripts aren't already a part of the training data for gpt.
My estimation is about 200 years in future to have a "human-brain AI" that works.
All idea should be treated equally, not based on revenue metrics. If everyone could make a Youtube clone, the revenue should be divided equally to all of creator, that's the way the world should move forward, instead of monopoly.
Everything will be suck, forever.
There must be a reason we can do so much while consuming so little, and then again struggling with other tasks.
What is the success if we build a machine that consumes just heaps of energy and then is as bad in maths as us?
- to say that because we're "more optimised" we must be the most optimised. Our brains are optimised well for certain things, sure, but computers are far more efficient at e.g. crunching numbers than we are
- to say that there's no success in a machine that can't currently beat us at math - this year has already proven that false
Hint for everyone else here:
It's about scaling LLMs.
I think a clearer answer is that scaling alone will certainly NOT get us to AGI. There are some things that are just architecturally missing from current LLMs, and no amount of scaling or data cleaning or emergence will make them magically appear.
Some obvious architectural features from top of my list would include:
1) Some sort of planning ahead (cf tree of thought rollouts) which could be implemented in a variety of ways. A simple single-pass feed forward architecture, even a sophisticated one like a transformer, isn't enough. In humans this might be accomplished by some combination of short term memory and the thalamo-cortical feedback loop - iterating on one's perception/reaction to something before "drawing conclusions" (i.e. making predictions) based on it.
2) Online/continual learning so that the model/AGI can learn from it's prediction mistakes via feedback from their consequences, even if that is initially limited to conversational feedback in a ChatGPT setting. To get closer to human-level AGI the model would really need some type of embodiment (either robotic or in a physical simulation virtual word) so that it's actions and feedback go beyond a world of words and let it learn via experimentation how the real world works and responds. You really don't understand the world unless you can touch/poke/feel it, see it, hear it, smell it etc. Reading about it in a book/training set isn't the same.
I think any AGI would also benefit from a real short term memory that can be updated and referred to continuously, although "recalculating" it on each token in a long context window does kind of work. In an LLM-based AGI this could just be an internal context, separate from the input context, but otherwise updated and addressed in the same way via attention.
It depends too on what one means by AGI - is this implicitly human-like (not just human-level) AGI ? If so then it seems there are a host of other missing features too. Can we really call something AGI if it's missing animal capabilities such as emotion and empathy (roughly = predicting other's emotions, based on having learnt how we would feel in similar circumstances)? You can have some type of intelligence without emotion, but that intelligence won't extend to fully understanding humans and animals, and therefore being able to interact with them in a way we'd consider intelligent and natural.
Really we're still a long way from this type of human-like intelligence. What we've got via pre-trained LLMs is more like IBM Watson on steroids - an expert system that would do well on Jeopardy and increasingly well on IQ or SAT tests, and can fool people into thinking it's smarter and more human-like than it really is, just as much simpler systems like Eliza could. The Turing test of "can it fool a human" (in a limited Q&A setting) really doesn't indicate any deeper capability than exactly that ability. It's no indication of intelligence.
