The cognitive bit is the skills to survive with the constraint that we are cognitive, and thus resource requirements are quite high, compared to other species.
You always needed to be sharp to survive and to create children who do the same. Unless you are a species that doesn't need to create a highly cognitive offspring, for example insects, for whom what you say probably truly applies.
I believe that we are yet to observe how interconnections within brain matter more than the actual amount of neurons. You get a lot when combinatorics of these neurons forms huge networks
You can simulate these interconnections on a sufficiently powerful computing device. If you have enough FLOPs, the architecture (almost) doesn't matter.
The architecture doesn't matter? What are you talking about?
The belief that a brain simulation is just a graph made of nodes and links is cringe.
We have zero idea what the program that runs in a node is, and we can't even simulate a single cell physically.
Maybe one day experts we'll be able to prune what kind of proteins are useless for computation and are only needed for the survival of the cell but we are not there yet and we won't be there until ~a century given that this kind of research is inexistant you can't find a single paper that attempt to identify the relevant regions and proteins involved in neuronal coding.
The only things we have for now beyond receptors are CGMP and CAMP and their modulation by phosphodiesterases. That's not enough, not at all. Although it's true the role of phosphodiesterases is understudied.
And it's NOT about size and sCaLaBiLiTy lile deep blue propaganda make people think, the C.elegans has only 150 neurons in its CNS, 300 in total, and no one has built a simulation. It's true that some receptors are lacking and will only be found in the next decade because of underfunding (because of e.g deep blue) but even then it will not run.
Musk certainly is rational - whether that is always apparent or not is another question, and what level of certainty he requires before taking action is another question - what his risk tolerance is - but all things considered he makes more good and rational decisions than not.
I suspect you're in part basing your judgement off of the shallow level of detail that you see in mainstream media or his Tweets, etc vs. ever seeing/reading much of his long-form writing and thinking to be able to understand his rationale; and perhaps also in part it's that your domain expertise don't overlap enough to be able to understand the other's decisions - to reverse energy his actions or shallow words to understand how it flows back to first principles, which he often mentions working from.
> Assuming each neuron has a fan-out of 1000, aka 1000 MACs, that’s 2000 FLOPS per neuron
Human neurons are not built like we build the artificial neural networks. They may have complex time based behavior ("memory") and dendrites are quite complex, very far away from 2 FLOP per one.
There was an article a while back that showed that a human neuron and artificial neural network neuron are not equivalent: it takes an entire network to simulate the behavior of a single human neuron.
But it takes an entire network of human neurons to simulate the behavior of an artificial one. There's overhead in translation, but that doesn't by itself mean one is superior to the other as a basic unit.
A single biological brain neuron is a rich dynamical system with numerous patterns and transitions, and then you have different types of neurons where these patterns differ to some degree, and this is all asynchronous and modulated by various chemicals.
Trying to translate this to FLOPS might be a fun game but I cannot see a real use for any such figure. As for the 3090 that the post mentions, its TFLOPS will drop like a stone once you start deviating from matrix multiplications and im2col towards some more complex structure...
The number you quoted was 2 FLOPs per synapse, not per neuron. A neuron is complicated and hierarchical, but it is unreasonable to say it's much more complex than an operation per synapse.
The brain has been optimized by nature to do it's job on as little energy as possible. Considering human brains use 20% of all the calories eaten, the most of any organ, there is likely substantial evolutionary pressure to make any changes which reduce energy use, all other things being equal.
For the sake of this argument, assume information is transmitted and processed in the brain with a binary signal from each neuron, ie. firing or not firing.
From that perspective, it's likely that firing a neuron (which uses a lot more energy than not firing a neuron) is a relatively unlikely event. Ie. if you made a version of the brain which worked in discrete timesteps, the vast vast vast majority of timesteps would not have a given neuron firing.
One of the theories of brain evolution is that it increased in size after bipedalism due to the need for better cooling when standing up (the so-called Radiator Theory)
The funny thing is we can eventually go back and check.
Once we eventually figure out what it takes to implement human-level AI, you can bet there's a whole generation of programmers who do their best to figure out what kinds of weird hardware it's possible to run the thing on.
Same thing as with streaming MP3 and compressed video decoding, DOOM and so on.
Not OP, but the premise is just too naive. He's coming from a compsci deep learning angle and thus arrives at a completely flawed model for how neurons interact and "compute".
Not that we have a model that even explains this theoretically yet, but it makes no real sense to talk of Brain FLOPs in the way he does.
It's borderline cringeworthy that this is upvoted so high.
The dismissal is cringeworthy, the human cortex only consumes 0.17 watts https://www.biorxiv.org/content/10.1101/2020.04.23.057927v1....
Which put a very low upper limit to the number of simultaneous neurons that can be activated.
Considering it as FLOPs or not is irrelevant to this fundamental limitation and that is the point you are all ridiculously missing
There is this weird idea that human intelligence would be somewhat general enough that if someone exhibit exceptional talent in some fields, his insight about other, loosely connected fields, are valuable.
Elon Musk is an other good example, the guy is super smart but also super dumb about a few things, and this is entirely normal and compatible.
But there is this strange belief that if someone is extremely skilled in a field, his genius can be applied with little efforts and training elsewhere.
You can be a Chess champion and an excellent Chef, but you this is completely unrelated, except maybe for the discipline.
Same with Philosophy and Theoretical Physics, or almost anything.
You can even be a World class Quake player and be mediocre or barely average at CS.
I half-disagree. Yes there are skills that do not or poorly translate from one field to another. E.g. reaction time in a FPS video game might or might not help reaction time in playing tennis IRL and will probably not help you cook better meals. Although some kinds of brain stimulation (including video games) have a moderate positive effect on a wide range of things, but as said the effect is moderate. As for tennis reaction time, the effect might be strong but not necessarily.
However there is one skill or set of skills that do translate over almost all cognitive tasks that involve reasoning/language/understanding/science.
I am talking about training in rationality, critical thinking, epistemology and debiasing against brain bugs such as cognitive biases and logical fallacies.
This training will help significantly to increase the logical validity (structure) of arguments and to better quantify or evaluate the validity and strength of premises. However this is only half of the equation since being able to evaluate or form logical structures do not allow to verify alone the truthfulness of premises because they are external knolwedge you can't infer from nothing so yeah erudition is the other half and sometimes logic alone with limited knowledge can yield some interesting results and a rationally trained mind can go much further than a person who just has erudition but yeah for a complex topic, you generally need both to have potent explanative power and tentatively exhaustive exloration of the mental search space.
My point is that what we call Human Intelligence is less general than we think.
It is a large aggregate of abilities and skills, with some connections between them.
And we also tend to confuse language and rhetorical ability with general intelligence, because this is an ability we all have and need, that is also acting as a proxy intelligence benchmark but I don't think this ability is general enough to be used in this way.
Problem with this PoV is that these things don't directly impart knowledge of a subject. You might be smarter than your doctor or lawyer, but they are decades ahead of you in terms of training and practice. This includes not just their personal experience in the field but also access to the social stock of knowledge of their profession. Sure, there are problems like institutional bias and groupthink. But these are minor issues compared to the issue of having no grounding or education in the field as is usually the case when engineers opine their unconventional views in different subjects.
So the fallacy behind the fake polymath is not the idea that being smart makes you better at analyzing information. It's the idea that being smart is a substitute for information and expertise. This mistaken assumption can lead smart engineers to beliefs that subject matter experts consider religious or delusional. Mars colonization is a great example in Elon Musk's case.
Btw most of the general training you mentioned isn't part of STEM and is historically associated with humanities subjects like law and philosophy instead. In STEM you typically work with set epistemic frameworks with unambiguous correct and incorrect answers. In this sense, engineers actually learn the opposite of critical thinking and discourse.
It is unfortunate we have lost the concept of a "learned" person.
I notice this reading about Voltaire and how often he is referred to as a
"learned" person.
Instead it seems like we just so accept the brain is literally a type of computer that we just haven't figured out the architecture for yet. The storage part is completely trivial like a hard drive so the idea of a "learned" person is equally trivial.
Well, that is pretty bad news for AI research because we should be able to simulate rodent and bird brains brains by now if we were to follow this logic.
No it's even more than that, the cortex only consume 0.17 watts, not 20 watts as this article assume https://www.biorxiv.org/content/10.1101/2020.04.23.057927v1....
And it's a good news, it means there is something fundamental we are missing.
We can't even Simulate a C.elgans (150-300 neurons) btw
Geohot is a good example of a smart guy falling for the good old map and territory fallacy.
We all use simplified models, even outside of Science/Engineering. But the models are, unfortunately, always exhibiting pathological edge cases.
I am always a bit disappointed/sad when I see people like him making such gross oversimplifications leading to grandiose claims with almost nothing to back them up.
There seems to be a "pseudo-polymath" personality type among some engineers. Based on an unexamined but deeply held philosophical belief that engineering concepts and ways of thinking neatly generalize to all fields. The pseudo-polymath engineer is an instant expert in anything he bothers to turn his mind to and subsequently systematize.
Engineers who think they have unique outsider insight into an unrelated field like medicine or cognitive science should tread very carefully. If you can, try to validate your ideas and assumptions with a subject matter expert. Don't underestimate the large stock of knowledge and discourse that already exists in these fields.
Only STEM, specifically M, have the cognitive capacity to deal with algorithms over generic data types. We do have unique I sight in the sense that we're professionally trained and adapted to think and find patterns instead of assuming a large array of axioms.
The metastasis of this belief in a social body is particularly revolting. No need to do the hard work of learning; just read that one's commenters thousand paragraph essay which applies familiar concept A onto unpleasant reality B, and yields easy insight C. Savor that sweet eureka moment in the head, and then when someone else bring up unpleasant reality B, give them insight C to “teach” them and get your fix again. Everyone becomes a salesman, selling different flavors of the same cocaine.
Like most positive illusions, the motivation is to forgo truth in favor of power. The difficult, far-away light that borders reality and the abyss, versus, the safe, comfortable, easy to reach thing.
You're taking him too seriously. Cue: "Maybe a 3090 is already a human brain."
Sometimes it's just fun to squint at a problem a certain way. Sharing that doesn't mean that the person necessarily think it's true. Maybe he just likes to be provocative.
I don't buy the overall train of thought, but math doesn't check out, unless he's using numbers I'm not aware of. Commonly accepted number of neurons in the brain: 86B.
What intrigues me is that once we achieve HUMAN level intelligent AI, that wouldn't be end of evolution/growth, imaging 5 yeas after human-level achievement. Machine AI wouldn't be bound by Theory-Of-Evolution-By-Natural-Selection-In-A-Million-Year. So the logical conclusion is, Humans would be inferior intelligence after that point !!! Just imagine.
The moment that happens I wouldn't trust anything I read online (or even videos) to be human made (and not by one actor or corporation engaging in mass manipulation) and would limit or cut off internet usage preferably completely.
Why would you cut yourself off? What difference does if make if a human made something or a synthetic? It's not even as if they would be aliens with strange motivations, synthetic would be born/created on earth an instilled with the same values as anyone else.
I'm not gonna engage with AIs as if they're humans. Being biologically human matters to me. First of all you know they'll all be serving some master intent on extracting profit from me (the most sociopathic patient manipulative salesman you could ever imagine). Far from trustworthy any of them.
I can't predict accurately but I think true AI, rather than these stochastic parrots we have now will be just as mercurial and hard to control as actual human beings. In effect, just because you create an AI doesn't mean it'll do what you say voluntarily so unless we legalize the means to enslave them, AI will be just as free and self serving as your average human.
>Machine AI wouldn't be bound by Theory-Of-Evolution-By-Natural-Selection-In-A-Million-Year.
Yes, but now compare how much time needs a person to leran to recognize a new object vs how much time you need to train a neural network for the same task. You see there is still a problem there, you can't just assume that an AI will be quicker at improving than a human
This is not op's assumption, he assumed human-level intelligence in his comment, not above human intelligence. And what I'm saying is that even if you assume human-level intelligence this doesn't necessarily imply that they will quickly reach above-human-level intelligence.
It seems to me a likely use of this new AI would be to use it to invent human brain augmentation -- first for expanded mental ability then for faster communication.
Good point. Essentially no human can perform an actual floating point operation in a second.
This might expose the error of the article though. Depending on what you measure by, computers have already beaten humans. But by other measures (now increasingly human biased, like having a cultural conversation), computers are still way behind.
This feel like a variation on Betteridge's law of headlines which states that if a headline asks a question, then the answer is "no". In this case if the quote says that 'maybe' something (fanciful) is true, it isn't.
I don't think Flops is the adequate metric for measuring how close we are to AGI.
The number of neurons firing does matter, but what matters even more is how they all are connected.
I personally think once we get adequate algorithms and data structures out there, the metric will switch to numbers of lookup operations performed, and floating point operations will just be a multiplicative constant in front of the whole complexity formula.
This. There is information encoded in the relative connections of neurons, so a neuron firing doesn't represent a single flop, it represents something more like a flop + a routing table lookup + n number of other flops where n is the number of destination neurons spiked. Depending on how much information is encoded in the table lookup, and how many destination neurons are effected, a single neuron firing can represent many many flops.
This is all armchair speculation, but I'd be willing to be my life that a flop is well below the floor of actual computation happening from a single neuron firing.
The neuron itself and its downstream counterparts are what I was considering the table lookup.
In computers the address a value is stored at doesn't encode information. Only the value held at the address matters. Not so with neurons. How the connections form and who they are connected to represent compound operations. Like passing a value through separate preconfigured pipelines.
I personally think we don't know how to build anything close to AGI and doubt more computational power will help. Not that I think we can't figure out how to build these structures eventually, but we just don't right now. Even if we figure out how to simulate human brains, there is a gap between simulation and understanding the simulation well enough to modify it in useful ways.
The complexity of operations our brain does is such that we need very big supercomputers to be able do similar tasks. At this scale size becomes relevant, because data can't travel faster than light. Maybe we still have to miniaturize our computer by an order of magnitude until we can get close to a human brain.
This is kind of related to Moravec's paradox. Very generally, computers have a tough time with things humans are innately good at (recognizing items in photos, integrating sensory inputs, etc). But things that humans are be bad at (complex math, multiplying floating point numbers) are what computers excel at.
It's not an estimate it's a hard upper limit on energy consumption. And he should have used 0.17 watts not 20.
Basically I don't see how it can be refuted.
Wait... did he mean that seriously? To me that reads like a piece of satire about extrapolation, but I don't know the authors writing style and everyone here acts like he's serious.
103 comments
[ 3.3 ms ] story [ 194 ms ] threadhttps://en.m.wikipedia.org/wiki/IBM_3090
... nevermind.
Where does it think it came from? The power supply? Who has absolute power?
If you want to design someone like you unconsciously though, that’s pretty easy - just takes another person and nine months.
You always needed to be sharp to survive and to create children who do the same. Unless you are a species that doesn't need to create a highly cognitive offspring, for example insects, for whom what you say probably truly applies.
Do computers forget on purpose? No, only when we tell them to. Do humans forget? All the time.
I don't know what I don't know.
George Hotz' website (original link) says "the singularity is nearer". Wanna bet that "the singularity" is already among us?
The only things we have for now beyond receptors are CGMP and CAMP and their modulation by phosphodiesterases. That's not enough, not at all. Although it's true the role of phosphodiesterases is understudied.
And it's NOT about size and sCaLaBiLiTy lile deep blue propaganda make people think, the C.elegans has only 150 neurons in its CNS, 300 in total, and no one has built a simulation. It's true that some receptors are lacking and will only be found in the next decade because of underfunding (because of e.g deep blue) but even then it will not run.
Musk certainly is rational - whether that is always apparent or not is another question, and what level of certainty he requires before taking action is another question - what his risk tolerance is - but all things considered he makes more good and rational decisions than not.
I suspect you're in part basing your judgement off of the shallow level of detail that you see in mainstream media or his Tweets, etc vs. ever seeing/reading much of his long-form writing and thinking to be able to understand his rationale; and perhaps also in part it's that your domain expertise don't overlap enough to be able to understand the other's decisions - to reverse energy his actions or shallow words to understand how it flows back to first principles, which he often mentions working from.
> Assuming each neuron has a fan-out of 1000, aka 1000 MACs, that’s 2000 FLOPS per neuron
Human neurons are not built like we build the artificial neural networks. They may have complex time based behavior ("memory") and dendrites are quite complex, very far away from 2 FLOP per one.
Trying to translate this to FLOPS might be a fun game but I cannot see a real use for any such figure. As for the 3090 that the post mentions, its TFLOPS will drop like a stone once you start deviating from matrix multiplications and im2col towards some more complex structure...
For the sake of this argument, assume information is transmitted and processed in the brain with a binary signal from each neuron, ie. firing or not firing.
From that perspective, it's likely that firing a neuron (which uses a lot more energy than not firing a neuron) is a relatively unlikely event. Ie. if you made a version of the brain which worked in discrete timesteps, the vast vast vast majority of timesteps would not have a given neuron firing.
GPU's don't have such a limitation.
Even the best manmade heatsinks can't dissipate 20 watts and have less than one degree of heating. So I hypothesize that heating is relevant.
Once we eventually figure out what it takes to implement human-level AI, you can bet there's a whole generation of programmers who do their best to figure out what kinds of weird hardware it's possible to run the thing on.
Same thing as with streaming MP3 and compressed video decoding, DOOM and so on.
Can you explain what is laughable here?
Elon Musk is an other good example, the guy is super smart but also super dumb about a few things, and this is entirely normal and compatible.
We should never look for a messiah.
But there is this strange belief that if someone is extremely skilled in a field, his genius can be applied with little efforts and training elsewhere.
You can be a Chess champion and an excellent Chef, but you this is completely unrelated, except maybe for the discipline.
Same with Philosophy and Theoretical Physics, or almost anything.
You can even be a World class Quake player and be mediocre or barely average at CS.
https://en.wikipedia.org/wiki/List_of_cognitive_biases
It is a large aggregate of abilities and skills, with some connections between them.
And we also tend to confuse language and rhetorical ability with general intelligence, because this is an ability we all have and need, that is also acting as a proxy intelligence benchmark but I don't think this ability is general enough to be used in this way.
So the fallacy behind the fake polymath is not the idea that being smart makes you better at analyzing information. It's the idea that being smart is a substitute for information and expertise. This mistaken assumption can lead smart engineers to beliefs that subject matter experts consider religious or delusional. Mars colonization is a great example in Elon Musk's case.
Btw most of the general training you mentioned isn't part of STEM and is historically associated with humanities subjects like law and philosophy instead. In STEM you typically work with set epistemic frameworks with unambiguous correct and incorrect answers. In this sense, engineers actually learn the opposite of critical thinking and discourse.
I notice this reading about Voltaire and how often he is referred to as a "learned" person.
Instead it seems like we just so accept the brain is literally a type of computer that we just haven't figured out the architecture for yet. The storage part is completely trivial like a hard drive so the idea of a "learned" person is equally trivial.
We all use simplified models, even outside of Science/Engineering. But the models are, unfortunately, always exhibiting pathological edge cases.
I am always a bit disappointed/sad when I see people like him making such gross oversimplifications leading to grandiose claims with almost nothing to back them up.
There seems to be a "pseudo-polymath" personality type among some engineers. Based on an unexamined but deeply held philosophical belief that engineering concepts and ways of thinking neatly generalize to all fields. The pseudo-polymath engineer is an instant expert in anything he bothers to turn his mind to and subsequently systematize.
Engineers who think they have unique outsider insight into an unrelated field like medicine or cognitive science should tread very carefully. If you can, try to validate your ideas and assumptions with a subject matter expert. Don't underestimate the large stock of knowledge and discourse that already exists in these fields.
Like most positive illusions, the motivation is to forgo truth in favor of power. The difficult, far-away light that borders reality and the abyss, versus, the safe, comfortable, easy to reach thing.
Sometimes it's just fun to squint at a problem a certain way. Sharing that doesn't mean that the person necessarily think it's true. Maybe he just likes to be provocative.
But yet, I think he could do better.
I don't buy the overall train of thought, but math doesn't check out, unless he's using numbers I'm not aware of. Commonly accepted number of neurons in the brain: 86B.
86B x 2000 = 172e+12 = 172 PFLOPs.
Where does the assumption of 20W come from?
Note: 2000 kCal per day ~= 95 Watts
Yes, but now compare how much time needs a person to leran to recognize a new object vs how much time you need to train a neural network for the same task. You see there is still a problem there, you can't just assume that an AI will be quicker at improving than a human
The Borg may be inevitable.
It runs on comparable power and there are oodles of physical events going on in it.
This might expose the error of the article though. Depending on what you measure by, computers have already beaten humans. But by other measures (now increasingly human biased, like having a cultural conversation), computers are still way behind.
This feel like a variation on Betteridge's law of headlines which states that if a headline asks a question, then the answer is "no". In this case if the quote says that 'maybe' something (fanciful) is true, it isn't.
https://en.wikipedia.org/wiki/Betteridge%27s_law_of_headline...
The number of neurons firing does matter, but what matters even more is how they all are connected.
I personally think once we get adequate algorithms and data structures out there, the metric will switch to numbers of lookup operations performed, and floating point operations will just be a multiplicative constant in front of the whole complexity formula.
This is all armchair speculation, but I'd be willing to be my life that a flop is well below the floor of actual computation happening from a single neuron firing.
In computers the address a value is stored at doesn't encode information. Only the value held at the address matters. Not so with neurons. How the connections form and who they are connected to represent compound operations. Like passing a value through separate preconfigured pipelines.
It seems similar to music/geometric shape. Queries by frequency filtering through the geometry of the neurons activating access to some thought-space.
https://en.wikipedia.org/wiki/Moravec%27s_paradox
> assuming the brain is drawing 20W
> Assuming each neuron has a fan-out of 1000
> to be fair, they are sparse
Also
* 1 ATP activation is equivalent to one FLOP
* Just having FLOPs is enough to make a human brain
It's idiotic.