The way a plane flies is quite different than the way a bird flies in complexity - they share an underlying mechanism, but planes don't need to flap wings.
It's possible that scaling up does lead to generality and we've seen hints of that.
- https://deepmind.com/blog/article/generally-capable-agents-e...
Also check out GPT-3’s performance on arithmetic tasks in the original paper (https://arxiv.org/abs/2005.14165)
Pages: 21-23, 63
Which shows some generality, the best way to accurately predict an arithmetic answer is to deduce how the mathematical rules work. That paper shows some evidence of that and that’s just from a relatively dumb predict what comes next model.
It’s hard to predict timelines for this kind of thing, and people are notoriously bad at it. Few would have predicted the results we’re seeing today in 2010. What would you expect to see in the years leading up to AGI? Does what we’re seeing look like failure?
Few have predicted a reasonably-capable text-writing engine or automatic video face replacement, but many have predicted self-driving cars would have been readily available to consumers by now and semi-intelligent helper-robots being around.
Just because unforeseen advancements have been made, does not mean that foreseen advancements come true.
It's about very good general problem solving software that's way beyond the capabilities of humans while not being aligned with human interests. Not because the software is evil, but because aligning values is an unsolved problem (humans aren't even totally aligned - and values also change).
If you have an intelligence that's very good at achieving its goal and you don't have a good way to align its goal with human goals, you can very quickly get into trouble if that intelligence thinks much faster than you do.
https://www.lesswrong.com/posts/mMBTPTjRbsrqbSkZE/sorting-pe...
https://www.lesswrong.com/posts/4ARaTpNX62uaL86j6/the-hidden...
https://www.lesswrong.com/posts/BEtzRE2M5m9YEAQpX/there-s-no...
https://www.lesswrong.com/posts/5wMcKNAwB6X4mp9og/that-alien...
Also, self-driving cars were mostly hyped up by companies, FSD is quite obviously a hard problem, much closer to general intelligence than the average NN application.
When I was a kid, I remember wondering how Soviets were obsessed with faking photos. A few years later, I saw Terminator 2 and realized that faking videos was also a thing. The tools for it would clearly get better and better over time. When I studied ML in the early 2000s, it seemed obvious that pattern recognition tasks such as image manipulation would be "easy" for computers, once we found the right approach and made the ML systems big enough. In the end, I decided not to pursue ML, because jobs were still scarce and I found discrete problems more interesting. That was probably the worst career mistake I've ever made.
Or, watch the latest Veritasium
#2 see linked video, he sits in the backseat, there is no driver and no one to take control.
These are just overhyped drive assist tools that market themselves immorally as something they aren’t.
If you ask it to compute something that would never have been seen on the internet it’s likely to fail. E.g. add 2 extremely large/rare numbers together
A rough calculation, humans can feasibly consume 4-5 TJ/annum of energy, of which a lot is going to go into motion or whatever. And if devoted to mental activity, it has a shelf life of ~70 years before they die.
A distributed computer might theoretically burn TJ/hr and once the weights are known they may as well be in a permanent record. The upper limits of what computers can learn and get good at are much higher than what humans can. They won't need as much implementation trickery as biology to get results.
And a plane is a vastly simpler machine than a bird!
I think this for a couple reasons:
1. The current gap in complexity is so huge. Nodes in an ANN roughly correspond to neurons, and the brain has somewhere on the order of 100 billion of them.
Even if we built an ANN that big, we would only be scratching the surface of the complexity we have in the brain. Each synapse is basically an information processing unit, with behavioral characteristics much more complicated than a simple weight function.
2. The brain is highly specific. The structure and function of the auditory cortex is totally different to that of the motor cortices, to that of the hypothalamus and so on. Some brain regions depend heavily on things like spike timing and ordering to perform their functions. Different brain regions use different mechanisms of plasticity in order to learn.
Currently most ANN's we have are vaguely inspired by the visual cortex (which is probably why a lot of the most interesting things to come out of ML so far have been related to image processing) and use something roughly analogous to net firing frequency for signal processing. I would consider it highly likely that our current ANNs are just structurally incapable of performing some of the types of computation we would consider intrinsically linked to what we think of as general intelligence.
To make the airplane analogy, I believe we're probably closer to Leonardo da Vinci's early sketches of flying machines than we are to the Right Brothers. We might have the basic idea, but I would wager we're still missing some of the key insights required to get AGI off the ground.
edit: it looks like you added some lines while I was typing, so to respond to your last points:
> it’s hard to predict timelines for this kind of thing, and people are notoriously bad at it. Few would have predicted the results we’re seeing today in 2010. What would you expect to see in the years leading up to AGI? Does what we’re seeing look like failure?
I totally agree that it's hard to predict, that technology usually advances faster than we expect, and that tremendous progress is being made. But the road to understanding human intelligence has been characterized by a series of periods of premature optimism followed by setbacks. For instance, in the 20th century, when dyes were getting better, and we were starting to understand how different brain regions had different functions, it may have seemed like we were close to just mapping all the different pieces of the brain, and that completing the resulting puzzle would give a clear insight into the workings of the human mind. Of course it turns out we were quite far from that.
As far as what we can expect in the years leading up to AGI, I suspect it's going to be something that comes on gradually - I think computers will take on more and more tasks that were once reserved for humans over time, and the way we think about interfacing with technology might change so much that the concept of AGI might not seem relevant at some point.
As to whether the current state of things is a failure - I would not characterize it that way. I think we're making real progress, I just also think there is a bit of hubris that we may have "cracked the code" of true machine intelligence. I think we're still a few major revelations away from that.
That's hardly accurate - didn't Musk and Co. promise self-driving cars by 2012? We're in 2020, and the SDC's are great for making youtube videos, but not any good at piloting a vehicle without human intervention.
Since the 90s it has been clear that the only thing holding back what we have today is limited processing power. While there may be some new insights and directions in AI, they are not "general" and they require 3 orders of magnitude more processing power for a lot smaller improvement in performance.
What has been clear since 2010 is that this field has passed the point of diminishing returns already. We throw vastly more computational power at problems that we ever did before, and then call the result an improvement.
Deep blue beat the best human at chess using 11.8 GFLOPS of computational power. Alphago beat the best human at go using 720000 GFLOPS of power. The complexity difference between Chess and Go are within a single order of magnitude - 10x to 99x difference in complexity (https://en.wikipedia.org/wiki/Game_complexity). The difference in AI processing power to beat the best human between Chess and Go is between 4 and 5 orders of magnitude (1000x and 100000x).
This does not look like a success to me - it looks like a brute-force approach. If you spend 10000x more resources for a 10x more benefit, you're at the point of diminishing returns.
Here's a great paper that should be written (but won't be) - plot the improvements in AI and the usage of computational power for AI on the same chart.
From the 90s (https://en.wikipedia.org/wiki/History_of_self-driving_cars#1...): "The robot achieved speeds exceeding 109 miles per hour (175 km/h) on the German Autobahn, with a mean time between human interventions of 5.6 miles (9.0 km), or 95% autonomous driving."
Yup, 95% autonomous. Today we have 95.x% autonomous with roughly 10000x the resource power thrown at the problem.
So, yeah, your assertion that "Few would have predicted the results we’re seeing today in 2010." is wildly off mark, we predicted more than what we see today because we did not expect to hit a point of diminishing returns quite so quickly.
The people who did the 95% SDC in 1997 would have been disbelieving if anyone told them, in 1997, that even with 10000x more processing power thrown at the problem and new sensor hardware that was not available to them, it won't get much better than what they had.
- FSD is one example, but the improvement of computer vision since 2015 has been massive and deep learning approaches to general problem solving too. This wasn’t something people were predicting in 2010.
- The deep blue and stock fish style approaches vs. the alpha go or alpha zero approaches are categorically different - the latter being a lot more interesting and closer to general learning vs. the older approach which is more like brute force.
- GOFAI was a bad approach and the optimism in the 60s was wrong. Today’s looks more promising. Being wrong in the 60s doesn’t necessarily mean people are wrong now. It’s hard to know: https://intelligence.org/2017/10/13/fire-alarm/
For the AGI bit I’d recommend reading some of Eliezer Yudkowsky’s writing or Bostrom’s book (though I find Bostrom’s writing style tedious). There’s a lot of good writing about take offs and AGI/goal alignment that’s worth reading to get a base level understanding of the concepts people have thought through.
AGI doesn’t need to be human like to be dangerous - it can be good at general problem solving with poorly aligned goals and just act much faster. Brains exist everywhere in nature, simpler than human brains. A lot of that computation in training could be an analog of the genetic “pre-training” of evolution for humans that gets our baseline which could be one reasons humans don’t seem to require so much. There was a massive amount of “computation” over time via natural selection to get to our current state.
> plot the improvements in AI and the usage of computational power for AI on the same chart.
Would that be meaningful? I mean, I use an infinity times the computational power for writing a letter than people did 100 years ago, still producing more or less the same results.
A lot of birds don't need to flap their wings either.
And while our brain is objectively very impressive, I don’t see how our complex abilities are anything but emerging features.
It's probably true that you could imagine a "perfectly designed" brain which could perform better on some tasks with less complexity, but I think it's also true that there's been a lot of selection pressure towards increased intelligence, so this is probably fairly well optimized.
> why don’t we have a normal abstraction for sending signals? Instead, we have like 10s of slightly different ones with different failings each, but each having many repetitive machinery leading to inefficient “spaghetti code”.
What do you mean exactly by this? Like different neurotransmitter systems? Because I think it's actually quite elegant how the properties of different neurohormones lead to different processing modalities. It's like we have purpose-built hardware on the scale of individual proteins specialized for different purposes. I'm not so sure a more homogenized process for neural signaling would be an improvement.
As for how optimized is the human brain, well good question. I think not even a single biological cell is close to efficient, at most it is at a local minima. The reason is perhaps that “worse is better” in terms of novel functionality. But I don’t think there was a big evolutionary pressure on sufficient intelligence once it emerged - it is sort of a first past the post wins all.
I have to be honest, I would take any such comparison from the 1950's with a huge pinch of salt. I think perceptions about how "dumb" an individual neuron is as a processing unit have shifted quite a bit since then.
> Also, I don’t think that comparing the training of a neural network to the brain is fair from an energy usage point of view - compare the usage of the final NN with it.
I'm not considering this in terms of the training efficiency, I'm looking at it in terms of the ratio between operational utility and energy used. There's no trained ANN with anything remotely close to the overall utility of the human brain at any scale, let alone one that weighs 3lbs, fits into a human skull and runs on 20 Watts of power.
The fundamental difference between our current approach and biological brains is just as much a hardware one as it is theoretical. CPUs and GPUs are simply not best fit for this sort of usage — a “core” of them is way too powerful for what a single neuron can do (even with the more correct belief that they are not as dumb as we first thought), even if they can calculate multiple ones simultaneously. I’m not sure of specifics but couldn’t we print a pre-trained NN to a circuit that could match/beat a simple biological neural network in both speed and power efficiency? Cells are inefficient.
I just don't think this is a meaningful comparison, and I'm not convinced it's evidence of the "limitations" of biological computation.
Silicone beats biology in doing binary computation because they're a single-purpose machine built for this task. But a brain is capable of serving as a control system to operate millions of muscle fibers in parallel to navigate the body smoothly in unpredictable 3D space, while at the same time modulating communication to find the right way to express thoughts and advance interests in complex and uncertain social hierarchies, while at the same forming opinions about pop-culture, composing sonnets, falling in love and contemplating death.
For me to buy the argument that ANN's can be more efficient than biology, you'd have to show me a computer which can do all of that using less resources than the human brain. Currently we have an assembly line for math problems.
> a “core” of them is way too powerful for what a single neuron can do
I just think you're vastly under-counting the complexity of what happens inside a single neuron. At every synapse, there's a complex interplay of chemistry, physics and biology which constitutes the processing of the neurotransmitter signal from the presynaptic neuron. To simulate a single neuron accurately, we actually need all the resources of a very powerful computer.
So it may be the case that we can boil down intelligence to some kind of process which can be printed in silicon. But I think it's also entirely likely that the extreme parallelism (vast orders of magnitude greater than the widest GPU) of the brain is required for the kind of general intelligence that humans express, and the "slowness" of biological computation is a necessary trade-off for the flexibility we enjoy. If that's the case, it's going to be very hard for a serial computer to emulate intelligence.
Also, while indeed we can’t simulate the whole of neuron, why would we want to do that? I think that is backwards. We only have to model the actually important function of a neuron. If we were to have a water computer, would it make sense to simulate fluid dynamics instead of just the logical gates? Due to the messiness of biology, indeed some hard to model factors will effect things (in the analogy, water will be spilt/evaporated) but we should rather overlook the ones that have a minimal influence on the results.
Yeah so I think this is where we fundamentally differ. It seems like your assumption is that neurobiology is fundamentally messy and inefficient, and we should be able to dispense with the squishy bits and abstract out the real core "information processing" part to make something more efficient than a brain.
So if that's your assertion, what would that look like? What would be the subset of a neuron that we could simulate which would represent that distillation of the information processing part?
Because my argument would be, the squishy, messy cellular anatomy is the core information processing part. So if we try to emulate neural processing with the assumption that a whole neuron is the base unit, we will miss a lot of that micro-level processing which may be essential to reaching the utility and efficiency achieved by the human brain.
I'm not against the idea that whatever brains we happened to evolved are not the most efficient structure possible. But my position would be, we're probably quite far in terms of current computing technology from being able to build something better. I would imagine we might have to be able to bioengineer better neurons if we really want to compete with the real thing, rather than trying so simulate it in software.
But I feel I may be misrepresenting your point now. To answer your question, maybe a sufficient model (sufficient to be able to reproduce some core functionality of the brain, eg. make memories) would be one that incorporates a weight for each sort of signal (neurotransmitter) it can process, complete with a fatigue model per signal type, as well as we can perhaps add the notable major interactions between pathways (eg. activation of one temporarily decreasing the weight of another, but in a way bias is sorta this in the very basic NNs). But to be honest, such a construction would be valuable even with arbitrary types of signals, no need to model it exactly based on existing neurotransmitters. I think most properties interesting from a GAI perspective are emerging ones, and whether dopamine does this and that is an implementation detail of human brains.
You only need to accurately simulate the input and the output.
Frankly, if that can’t be done with a Markov process I’d be very surprised, and we already know that Markov chains can be simulated with ANNs
For instance one of those is spike-timing dependent plasticity. Basically the idea is that the sensitivity of a synapse gets up-regulated or down-regulated depending on the relative timing of the firing of the two neurons involved. So in the classic example, if the up-stream neuron fires before the down-stream neuron, the synapse gets stronger. But if the down-stream neuron fires first, the synapse gets weaker.
Another one is synchronization. It appears that the firing frequency of groups of neurons which are - for instance representing the same feature - become temporally synchronized. I.e. you could have different neural circuits active at the same time in the brain, but oscillating at different frequencies.
Another interesting mechanism is how dopamine works in the Nucleus Accumbens. Here you have two different types of receptors at the same synapses: one of them is inhibitory, and is sensitive at low concentrations of dopamine. The other is excitatory, and is sensitive at high concentrations. What this means is, at a single synapse, the same up-stream neuron can either increase or decrease the activation of the down-stream neuron: if the up stream neuron is firing just a little, the inhibitory receptors dominate. But if it's firing a lot, the excitatory receptors take over, and the down-stream neuron starts to activate more. Which kind of connection weight in an ANN can model that kind of connection?
My overall question would be, do you think back-propogation and markov chains are really sufficient to account for all that subtlety we have in neural computation, especially when it comes to specific timing and frequency-dependent effects?
To boil it down, if you really want to argue that the behaviour of a neuron can’t be simulated by an ANN, you’re arguing that a neuron is doing something non-computable. At which point you might as well argue it’s magical.
1. Can ANN's (in their current iteration) achieve general intelligence
2. Can they do it more efficiently than a biological brain
It certainly has not been established that a Turing machine can achieve general intelligence.
The biggest thing is that computing is just now finally able to have enough data and large enough networks to really start to create more generalized models. With a computer 20 years ago you might be able to do squeeze out simple pattern recognition, but for every layer and every neural node you add to a neural network, the more complex the model becomes and the more edge cases it can fold into itself.
Take a look at Universal approximation theorem and how, with enough nodes in a neural network, you can solve pretty much any problem given the right weights.
I also find it improbable that intelligence will emerge from modern ML without some major leap. But you have added nothing to the discussion, beyond some impressions from undergrad, when we are talking about something that is a very active and evolving research area. It's insulting to researchers and practitioners who have devoted years to studying ML to just dismiss broad areas of applicability because you took a course once.
Do you disagree substantively with anything I have said, or do you just think I could have phrased it better?
If your conclusion is that current gradient based methods probably won't scale up to AGI, you're probably right. But if you want to get involved in the discussion of why this is true, what ML actually can and can't do, etc. I would encourage you to learn more about the subject and the current research areas, and draw on that for your discussion points.
Otherwise, it comes across as "I once saw a podcast that said..." type stuff that is hard to take seriously.
No doubt I come across as condescending, please take what I say with the usual weight you'd assign to the views of a random guy on the internet :)
If you can’t characterize the technical problem that creates a limitation then you are just expressing an uninformed opinion.
Even if you were an expert!
Once we finally had the tools to even start trying, in the late 80s/early 90s, it took us a very long time to "calibrate" these general ideas and figure out the "devils in the details" that were necessary to make certain ideas viable (for example, neural networks were discarded as a dead end in the 80s, and only considerably later were we able to discover that multi-layer networks essentially "salvaged" the idea).
Machine learning without the era of "modern computers" was a bit like flight before we'd really mastered the internal combustion engine - we understood quite a bit about it, and had theories about a lot of stuff (like the basic shape of a wing), and could successfully build gliders and such. Contrary to a lot of propaganda, the Wright Brothers didn't just arrive in the world like "lightning from a clear sky", but ... it had to become practical to do for us to then move on to putting the ideas through the paces, and all of the established theory from beforehand ran into the usual treatment of "no plan of battle survives contact with the enemy".
However until the hardware and software support for mainstream massively parallel execution became available it was a niche tool.
So the level of adoption, experimentation, deployment and research resources available are multiple orders of magnitude greater than 20 or 30 years ago.
As a practitioner (for my entire career) the field still operates as a new field, with enormous areas for new experimentation and interesting new creative advances happening quickly.
So we are still at the beginning.
The more you learn about neurobiology, the more apparent it is that there are so many levels of computation going on - everything from dendritic structure, to cellular metabolism, to epigenetics has an effect on information processing. The idea that we could reach some approximation of "general intelligence" by just scaling up some very large matrix operations just seemed like a complete joke.
However, as you say, that doesn't mean what we've done in ML is not worthwhile and interesting. We might have over-reached thinking ML is ready to drive a car without major fourth-coming advancements, but use-cases like style transfer and DLSS 2 are downright magical. Even if we just made marginal improvements in current ML, I'm sure there is a ton of untapped potential in terms of applying this tech to novel use-cases.