When will computer hardware match the human brain? (1998)
jetpress.org
jetpress.org
* 86 billion neurons in a brain [1]
* 400 transistors to simulate a synapse [2]
That's 34 trillion, 400 billion transistors to simulate a human brain.
As of 2024, the GB200 Grace Blackwell GPU has 208 billion MOSFETs[3]. In 2023, AMD's MI300A CPU had 146 billion transistors[3]. In 2021, the Versal VP1802 FPGA had 92 billion transistors[3]. Intel projects 1 trillion by 2030; TSMC suggests 200 billion by 2030.
We'll likely have real-time brain analogs by 2064.
(Aside, these are the dates I've used in my hard sci-fi novel. See my profile for details.)
[1]: https://pubmed.ncbi.nlm.nih.gov/19226510/
> * 400 transistors to simulate a synapse [2]*
> * That's 34 trillion, 400 billion transistors to simulate a human brain.*
You forgot about:
1-Astrocytes - more common than neurons, computational, have their own style of internal and external signaling, have bi-directional communication with neurons at the synapse (and are thought to control the neurons)
2-Synapse: a single synapse can be both excitatory and inhibitory depending on it's structure and the permeant ions inside and outside the cell at that point in time and space - are your 400 transistors handling all of that dynamic capability?
3-Brain waves: are now shown to be causal (influencing neuron behavior) not just a by-product of activity. Many different types of waves (different frequencies) that operate very locally (high freq), or broadly (low freq). Information is encoded in frequency, phase, space and time.
This is the summary, the details are even more interesting and complex. Some of the details are probably not adding computational capabilities, but many clearly are.
If we had a simple, throw it in a box for five years, stable memristor, we'd have everything we'd need for brain modeling. Being as a memristor is effectively a synapse.
With that, we would want to do everything in the analog domain anyways.
[0] the other three are the resistor, the capacitor, and the inductor.
I can't seem to find anything on memristors being on chip or really being in solid state anything.
I would love to get my hands on them!
It was possible to buy memristor storage devices, but they weren't sufficiently competitive with ordinary flash memory and I believe they're relegated to niche applications such as high erase counts.
AI implementations on Memristor circuitry has come up often enough that even I've heard of it. E.g.: A random recent development is https://www.nature.com/articles/s41467-024-45312-0
Things us hard sci fi fans will insist on:
- You have significantly undercounted transistors. As of 2024 you can put up to 8.4 terabytes of LPDDR5X memory into an NVidia Grace Blackwell rack. So that's 72 trillion transistors (and another 72 trillion capacitors) right there.
- A GPU executes significantly faster than a neuron.
- The hardware of one GPU can be used to simulate billions of neurons in realtime.
- Why limit yourself to one GB200 NVL72 rack, when you could have a warehouse full? (what happens when you create a mind that's a thousand times more powerful than a human mind?)
You really need to separate the comparison into state (how much memory), and computation rate. I think you'll find that an NVidia GB2000 NVL72 will outperform a brain by at least an order of magnitude. And the cost of feeding brains far exceeds the cost of feeding GB2000's. Plus, brains are notoriously unreliable.
The current generation of public-facing AIs are using ~24Gb of memory, mostly because using more would cost more than can conveniently given away or rented out for pennies. If I were an evil genius looking to take over the world today, I'd be building terabyte-scale AIs right now, and I'd not be telling ANYONE. And definitely not running it in Europe or the US where it might be facing imminent legislative attempts to limit what it can do. Antarctica, perhaps.
You mean simulate billions artificial neurons, right?
You can't simulate a single biological neuron yet because it's way too complex and they still don't even understand all of the details.
If they did have all of the details, at minimum you would need to simulate the concentrations of ions internally and externally as well as the electrical field around the neuron to truly simulate the dynamic nature of a neuron in space and time. It's response to stimulus is dependent on all of that and more.
If you looked at the operation of a digital computer, you'd find an analog substrate, with some quite interesting properties. Deciding whether the analog substrate is an important semantic element of how the computer worked, or whether its operation should be understand digitally, is as much a (research-driving) leap of faith as anything else.
The process by how the neuron determines whether it's time to fire or not is almost grotesquely analog, in that it is effected by a million different things in addition to incoming spikes.
How much of each process matters for what is the mind-generating part of the brain's work? We don't know.
Myelination (which allows very fast, and digital, communication between neurons) is universal in vertebrates, which points to it being important for intelligence.
As a counterexample, we have cephalopods, which accomplish their impressive cognitive tasks using only 'slow' neurons; they get around the speed limit by having huge axons, which reduce internal resistance. So we know that digital transmission isn't necessary for intelligence, either.
The nature of cognition is a fascinating subject, isn't it!
Skull Volcano lair on an island. Geothermal energy to power the place. Seawater for liquid cooling.
Then of course, you also need the beautiful but corruptible woman as some high-up part of the organization. And a chief of security in your evil enterprise with some sort of weird deformity or scarring - preferably on the face.
Protip: If you catch a foreign spy on the island, kill them immediately. Do not explain your master plan to them just before you kill them.
2) Models are not linearly scalable in terms of their performance. There are 8B parameter models (Zephyr-Mistral) routinely outperforming 180B parameter models (Falcon). Skill issue, like they say.
3) Even if you are an evil genius, you need money and computing power. Contrarily to what you might be reading in the media, there isn't much capital available to evil geniuses on a short notice (longer term arrangements might still be possible). And even if you have money, you can't just buy a truckload of H100s. Their sales are already highly regulated.
Powered by what?
Solar panels in a place where the sun doesn't even rise for months at a time? That big beefy transmission grid that the Antarctic continent is so famous for? The jillions of oil wells already drilled into the Antarctic?
Oh wait, I know: penguin guano.
Pass the crack pipe, you've had enough my friend.
Antarctica was chosen because the area on the mainland between 90 degrees west and 150 degrees west is the only major piece of land on Earth not claimed by any country[1], which makes it a good place place to locate businesses that do not wish to be subject to regulations on Artificial Intelligences. Plus it's much more inconvenient to attack than it would be if you located it in a volcanic island.
[1] Sez wiki.
You could ballpark that as that’s 400 * 7000 * 86 billion transistors to simulate a brain that can’t learn anything, though I don’t see the point. Reasonably equivalent real time brain emulation is likely much further I’d say 2070 on a cluster isn’t unrealistic, but we’re not getting there on a single chip using lithography.
What nobody talks about is the bandwidth requirements if this isn’t all in hardware. You basically need random access to 100 trillion values (+100t weights) ~100-1,000+ times a second. Which requires splitting things across multiple chips and some kind of 3D mesh of really high bandwidth connections.
Agreed. In the novel, it's a neuromorphing perception lattice, to suggest a 3D mesh. If you'd like to have a read, I'd be grateful for your feedback.
Wild-ass prediction: future circuits are literally grown; circuit layout will be specified in DNA. No lithography necessary.
I imagine that would push out the timeline.
A single “cycle” (and you need ~100-1,000+ cycles per second) would involve ~7,000 synapses per neuron * 86 billion neurons random memory accesses each second. Each of those accesses need to first read the location of memory and then access that memory.
For perspective a video card which is heavily optimized for this only gets you low billions of random memory access per second. True random access across large amounts of memory is hard. Of course if you’re using a 1 million GPU supercomputer you now need to sync all this memory between each GPU every cycle, or have each GPU handle network traffic for all these memory requests…
PS: The article assumes more efficient AI designs for good reason, actually using a neuronal network the size of the brain in real time is prohibative.
Distributed cloud systems certainly surpass this number as well. https://www.top500.org/lists/top500/2023/11/
That or some novel AI architecture. I doubt we need to actually emulate a brain to reach near parity with human cognition.
What does 'unsolved' mean in this context?
When there’s occasional talk about triple-letter agencies having far-future tech now, I used to wonder what it might be. I guess this is one of those things.
The brain also does much more than a GPU. I’m not sure what percentage of the brain could be considered devoted to equivalent computation but is also interesting to think about. I’ve seen that the brains memory is equivalent to ~2PB. The power of a 12TB hard drive is ~6W. So current tech has a long way to go if it’s ever going to get close to the brain, 2PB storage is about 1kW alone.
That’s pretty impressive for a bunch of goo that resulted from a random process.
Makes me think we are barking up the wrong tree with silicon based computing.
Active dendrites the that can do operations like xor before anything reaches the soma, or use spike timing, once again before the soma are a couple of examples.
SNNs, or spikey artificial NNs have been hitting limits of the computable numbers.
Riddled basins as an example, which are sets with no open subsets. Basically any circle you can draw will always contain at least one point on a boundary set.
https://arxiv.org/abs/1711.02160
This is a much stronger constraint than even classic chaos.
While I am sure we will continue to try to make comparisons, wetwear and hardware are cheese and chalk.
Both have things they are better at and neither is a direct replacement for the other.
It's possible that I'm not aware of some breakthroughs on this topic, though.
Living beings are more "Rube Goldberg" than "Swiss Watch" - by virtue that they came to being guided not by intention, but circumstance.
On many occasions to get ambitious things to work despite a lack of good support, special subsystems were created with whatever wacky solution could be made to work, wrapped in sane surface API’s so the system as a whole could remain organized and accomplish the task well.
Five or so years after any of these were created, they could have easily been replaced with more reasonable code. But they worked reliably and efficiently. Their API’s were sensible and stable. So they accumulated.
If you reviewed all the code as a whole for the first time it was like finding a secret circus.
https://lozierinstitute.org/dive-deeper/the-newborn-senses-s...
The genes randomly mutate, and their result is subject to fitness testing.
The fascinating thing about evolution is that it's not actually necessary. The "unevolved" creatures persist just fine. The continued survival of cockroaches and whatnot proves that evolving to human levels is superfluous.
The Sperm carries the requisite DNA to activate the egg, which then just executes the body's Replicate() method. The sperm being required is so people don't spontaneously get pregnant.
The DNA coding for proteins is just to build the replicate() method, which then starts copying attributes from the host's (read: mother's) body. Much like how there are differences in compression/expansion algorithms (zip, tar, gz, etc..), the sperm that initiates the replication/unpacking adds non-neutral variance, hence the genetic traits of both parents present in the child.
Put a different way, why reinvent the wheel if I could just grab the memoized visual processing from the parent? Then, if a trait isn't present on the host, say, mid-digital hair, I (as the replication algorithm), "wouldn't know what I wouldn't know" which could perhaps be seeing as synonymous with "evolution."
In this way, DNA does control the blueprint for the brain, down to the initial generic wiring scheme of the cortex - six layers of differentiated neuron types, with a specific pattern of inter-connectivity. As a rough analogy, you could think of this DNA-controlled initial wiring scheme as something like an untrained transformer... the architecture is there, but won't do anything useful until it has been exposed to a lot of data, which will complete the wiring scheme (synapses = brain's model weights).
The brain is going to start to learn as soon as it has sufficiently developed, which is certainly before birth, and newborn babies have been shown to already respond to things like their mothers voice which they were exposed to in the womb.
Intelligence - ability to learn - is different from knowledge. A newborn baby has very little knowledge, but it does have intelligence since that comes from the brain architecture, hence ability to learn, that our DNA "encodes".
But I have no idea what is known nowadays.
If you look at state of the art pre-trained transformers, they gobble up unfathomable quantities of dense textual data that far exceed what even the most learned and smart humans could traverse in centuries. Yet, their results are barely intelligent.
The human neural network, on the other hand, stores and is able to recall an estimated average of a few bits per waking second for a decade or two, barely a few GB of useful storage [1]. A great deal of effort and substance of the brain is dedicated to compressing the sensory input into an internal representation useful for rational though - a problem much simplified for a text-fed neural network. The ability of blind and deaf people (or even blind-deaf) for rational reasoning show these ancillary tasks are not intrinsic to human reason.
So things seem to indicate that what we lack is not hardware, but the right architectures and algorithms to unlock human level reason, the exceptionally tight self training loop of humans that can make sense of the world and even push humanity's understanding further using only a few GB of compressed training data.
[1] LANDAUER TK. How much do people remember ‐ some estimates of the quantity of learned information in long‐term‐memory.
Actually this is untrue. Yann Lecun himself said it. A kid at the age of 6 has had much much more data go through his brain (based on the visual data) than what any transformer has ingested. Moreover, words are already compressed data with a lot of lost information, nowhere near the raw experience about the world that our brains have. On top of that there's the evolutionary stuff.
That's exactly the point, 99.9% of the information processed by the 6 year old has went into training his visual cortex, a subsystem not necessary for reasoning. This will eventually be used 8 hours a day for transforming visual representation of text on a computer screen into a compressed textual form which the inner rational loop can easily manipulate. The holy grail of AI, for anything but artistic creation, is just this inner loop.
But are they necessary for reasoning? Records of deaf-blinds with IQs a few standard deviations above average seem to suggest even highly restricted sensory inputs can allow for rational thought. It stretches credulity that you could convey more than a few bits of meaning per second to a child purely by hand gestures and touch, let alone the exabytes claimed.
Of course someone like Yann Lecun will say scale is the only solution, he leads one of the juggernauts that uses scale as a competitive moat. We shouldn't expect AI researchers in positions of power to be any less vane, power hungry or human than any other humans.
For you and me, a verbal description of a pattern test such as "You are given seven squares. Square A contains a filled cross in the upper left position, an empty circle in the middle... etc." would send us grasping for pen and paper. But with training many people can play mental chess, there are famous blind theoretical physicists and so on. This all seems to suggest that natural language could be called "Turing complete", to bastardize the term, for describing the real world to a rational entity and formulating solutions to real world problems.
Evolution definitely still has us beat on optimization-work put into thinking.
Some implemented hacks:
- emotions as shortcuts for reasoning
- learning through imitation
- self-regulation through relation
Computers are already great peripherals for brains. Maybe brains will become peripherals for computers or become completely replaced.
At the present rate, computers suitable for humanlike robots will appear in the 2020s. Can the pace be sustained for another three decades? The graph shows no sign of abatement. If anything, it hints that further contractions in time scale are in store. But, one often encounters thoughtful articles by knowledgeable people in the semiconductor industry giving detailed reasons why the decades of phenomenal growth must soon come to an end.
EDIT: This Wikipedia article puts an AMD CPU at ~2 million MIPS https://en.wikipedia.org/wiki/Instructions_per_second as the highest one.
So while he wrote about MIPS, and those have indeed basically stopped significantly improving, FLOPS have continued to improve. And for AI in particular, for inference at least, we can get away with 8 bit floats, which is why my phone does 1.58e13/second, only a factor of x60 from a million-billion.
[0]: https://ourworldindata.org/grapher/supercomputer-power-flops
https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor...
With that in mind we are currently at 1 billion GigaFLOPS for the largest super computer in 2023. And the trendline since the 90's has kept up with the prediction made in the post. Although we can't exactly say 1 GigaFLOPS = 1000 MIPS.
https://ourworldindata.org/grapher/supercomputer-power-flops...
An RTX 4090 has AI compute of up to 661 teraflops in FP16[1]. FLOPS and IPS are not interchangeable, but just for fun, that's 6.6 * 10^14 FLOPS, which is just short of 1 billion million (10^15).
1. https://www.tomshardware.com/features/nvidia-ada-lovelace-an...
>"NVIDIA GeForce RTX 4090 Is The First Gaming Graphics Card To Deliver 100 TFLOPs of Compute Performance" https://wccftech.com/nvidia-geforce-rtx-4090-is-the-first-ga...
Currently about $1,800 on Amazon.
Why don't we see exponential density growth in other products? What would that even look lke anyways?
It's basically akin to something like the Manhattan Project or the Apollo Program, just something initiated by private corporations rather than a government-directed investment program.
Do you agree?
I'm much less keen to ascribe "intelligence" to large, pretrained language models given that I know how primitive their training regime is compared to a scenario where I might have been "blended" by their ability to "chat" (double quote here since I know ChatGPT and the likes do not have a memory, so all prior interactions have to be re-submitted with each turn of a conversation).
Intuitively, I'd be more prone to ascribe intelligence based on convincing ways of construction that go along with intellignece-like performance, especially if the model also makes human-like errors.
I highly doubt you'd be able to get even close to that kind of performance with current manufacturing processes. We'd need something entirely different from laser lithography for that to happen.
Like, you know . . . nerve cells.
At a low level: We take an analog component, then drive it in a way that lets us treat it as digital, then combine loads of them together so we can synthesise a low-resolution approximation of an analog process.
At a higher level: We don't really understand how our brains are architected yet, just that it can make better guesses from fewer examples than our AI.
Also, 400 W of electricity is generally cheaper than 20 W of calories (let alone the 38-100 W rest of body needed to keep the brain alive depending on how much of a couch potato the human is).
Are you serious? I don't think you have to be an expert to see that the average human can perform more work per energy intake than the average GPU.
> The problem isn't the manufacturing process, but rather the architecture.
It's very much a problem, good luck trying to even emulate the 3D neural structure of the brain with lithography. And there are few other processes that can create structures at the required scale, with the required precision.
You're objecting to something I didn't say, which is extra weird because I'm just running with the same 400 W/20 W you yourself gave. All I'm doing here is pointing out that 400 W of electricity is cheaper than 20 W of calories especially as 20 W is a misleading number until we get brains in jars.
To put numbers to the point, at $0.10/kWh * 400 W * 24h = $0.96, while the UN definition for abject poverty is $2.57 in 2023 dollars.
As for my opinion on which can perform more work per unit of energy, that idea is simply too imprecise to answer without more detail — depending on what exactly you mean by "work", a first generation Pi Zero can beat all humans combined while the world's largest supercomputer can't keep up with one human.
> It's very much a problem, good luck trying to even emulate the 3D neural structure of the brain with lithography.
IIRC by volume a human brain mostly communication between neurones; the ridges are because most of your complexity is a thin layer on the surface, and ridges get you more surface.
But that doesn't even matter, because it's a question of the connection graph, and each cell has about 10,000 synapses, and that connectivity be instantiated in many different ways even on a 2D chip.
We don't have a complete example connectivity graph for a human brain. Got it for a rat, I think, but not a human, which is why I previously noted that we don't really understand how our brains are architected.
> And there are few other processes that can create structures at the required scale, with the required precision.
Litho vastly exceeds the required precision. Chemical synapses are 20-30 nm from one cell to the next, and even the more compact electrical synapses are 3.5 nm.
Most likely because any brains that required more energy died off at evolutionary time scales. And while there are some problems with burning massive amounts of energy to achieve a task (see: global warming) this is not likely a significant short falling that large scale AI models have to worry about. Seemingly there are plenty of humans willing to hook them up to power sources at this time.
Also you might want to consider the 0-16 year training stages for human which have become more like 0-21 year training stages with at minimum 8 hours of downtime per day. This does adjust the power dynamics pretty considerably in that the time actually thinking daily drops to around 1/3rd the day boosting effective power use to 60w (as in you've wasted 2/3s the power eating, sleeping, and pooping). In addition that model you've spend a lot of power training is able to be duplicated across thousands/millions of instances in short order, where as you're praying that human you've trained doesn't step out in front of a bus.
So yes, reducing the capability of a human brain/body to performance alone is far too simplistic.
> Only on the outside, where they can be appreciated as a whole, will the impression of intelligence emerge. A human brain, too, does not exhibit the intelligence under a neurobiologist's microscope that it does participating in a lively conversation.
We only have fuzzy definitions of "intelligence", not any essential, unambiguous things we can point to at a minute level, like a specific arrangement of certain atoms.
Put another way, we've used the term "intelligent" to refer to people (or not) because we found it useful to describe a complex bundle of traits in a simple way. But now that we're training LLMs to do things that used to be assumed to be exclusively the capacity of humans, the term is getting stretched and twisted and losing some of its usefulness.
Maybe it would be more useful to subdivide the term a bit by referring to "human intelligence" versus "LLM intelligence". And when some new developments in AI seem like they're different from "LLM intelligence", we can call them by whatever distinguishes them, like "Q* intelligence", for example.
(Chollet, 2019, https://arxiv.org/pdf/1911.01547.pdf)
Priors here means how targeted is the model design to the task. Experience means how large is the necessary training set. Generalization difficulty is how hard is the task.
So intelligence is defined as ability to learn a large number of tasks with as little experience and model selection as possible. If it's a skill only possible because your model already follows the structure of the problem, then it won't generalize. If it requires too much training data, it's not very intelligent. If it's just a set number of skills and can't learn new ones quickly, it's not intelligent.
Yes, learning is an important aspect of human cognition. However, the key factor that humans possess that LLMs will never possess, is the ability to reason logically. That facility is necessary in order to make new discoveries based on prior logical frameworks like math, physics, and computer science.
I believe LLMs are more akin to our subconscious processes like image recognition, or forming a sentence. What’s missing is an executive layer that has one or more streams of consciousness, and which can reason logically with full access to its corpus of knowledge. That would also add the ability for the AI to explain how it reached a particular conclusion.
There are likely other nuances required (motivation etc.) for (super) human AI, but some form of conscious executive is a hard requirement.
Foundational changes are of course harder, but it does not mean we should drop it all together.
The very foundation of computer science was an attempt to emulate a human mind mindlessly processing data. Fixed that for you.
And I'm still not sure I agree.
The foundation of computer science was at attempt to process data so that human minds didn't have to endure the drudgery of such mindless tasks.
The tape is a piece of paper, the head is the human, who is capable of reading data from the tape and writing to it. The symbols are discernible things on the paper, like numbers. The movement of the tape ("scanning") is the eyes going back and forth. At each symbol, the machine decides which rule to apply.
Its an inescapable fact that we are trying to get computers to 1. operate as close to how we think (as we are the ones who operate it) and 2. to produce results which resemble how we think.
Abstractions, inheritence, objects, etc are no doubt all heavily influenced by thinking about how we think. If we still programmed using 1s and 0s, we wouldnt be where we are.
It seems incredibly short sighted to me to believe that because a few decades of research hasnt panned out, that we should all together forget about it.
You have read me backwards. I am firmly of the opinion that the last few decades of research has entirely and completely panned out. The topic is clearly in the mind of theorists in 1950. But I'm pretty sure early computer architects were more interested in creating better calculators than in creating machines that think.
[1] https://www.cs.virginia.edu/~robins/Turing_Paper_1936.pdf
The first section of the paper literally says "We may compare a man in the process of computing a real number to machine which is only capable of a finite number of conditions", gives a human analogue for every step/component of the process, continuously refers to the machine as a "he/him", and continuously gives justifications from human experience.
"We have said that the computable numbers are those whose decimals are calculable by finite means. This requires rather more explicit definition. No real attempt will be made to justify the definitions given until we reach § 9. For the present I shall only say that the justification lies in the fact that the human memory is necessarily limited. We may compare a man in the process of computing a real number to machine which is only capable of a finite number of conditions q1: q2. .... qI; which will be called " m-configurations ". The machine is supplied with a "tape " (the analogue of paper) running through it, and divided into sections (called "squares") each capable of bearing a "symbol". At any moment there is just one square, say the r-th, bearing the symbol <2>(r) which is "in the machine". We may call this square the "scanned square ". The symbol on the scanned square may be called the " scanned symbol". The "scanned symbol" is the only one of which the machine is, so to speak, "directly aware". However, by altering its m-configuration the machine can effectively remember some of the symbols which it has "seen" (scanned) previously."
"Computing is normally done by writing certain symbols on paper. "We may suppose this paper is divided into squares like a child's arithmetic book. In elementary arithmetic the two-dimensional character of the paper is sometimes used. But such a use is always avoidable, and I think that it will be agreed that the two-dimensional character of paper is no essential of computation. I assume then that the computation is carried out on one-dimensional paper, i.e. on a tape divided into squares"
"The behaviour of the computer at any moment is determined by the symbols which he is observing, and his " state of mind " at that moment. We may suppose that there is a bound B to the number of symbols or squares which the computer can observe at one moment. If he wishes to observe more, he must use successive observations. We will also suppose that the number of states of mind which need be taken into account is finite. The reasons for this are of the same character as those which restrict the number of symbols. If we admitted an infinity of states of mind, some of them will be '' arbitrarily close " and will be confused."
"We suppose, as in I, that the computation is carried out on a tape; but we avoid introducing the "state of mind" by considering a more physical and definite counterpart of it. It is always possible for the computer to break off from his work, to go away and forget all about it, and later to come back and go on with it. If he does this he must leave a note of instructions (written in some standard form) explaining how the work is to be continued. This note is the counterpart of the "state of mind". We will suppose that the computer works in such a desultory manner that he never does more than one step at a sitting. The note of instructions must enable him to carry out one step and write the next note."
"The differences from our point of view between the single and compound symbols is that the compound symbols, if they are too lengthy, cannot be observed at one glance. This is in accordance with experience. We cannot tell at a glance whether 9999999999999999 and 999999999999999 are the same"
Taken all this together, I don't think its far fetched to think Turing was very much thinking about the individual steps he was taking when doing calculations manually on a piece of graph paper, while trying to figure out how to formalize it. Perhaps you disagree, but saying its "completely mechanical" is surely false, no?
But they're not. The "bitter lesson" of machine learning is that the primitive operations are really simple. You just need a lot of them, and as you add more, it gets better.
Now we have a better idea of how evolution did it.
Given their performance, I think it is important to pay attention to their weirdnesses — I can call them "intelligent" or "dumb" without contradiction depending on which specific point is under consideration.
Transistors outpace biological synapses by the same degree to which a marathon runner outpaces continental drift. This speed difference is what allows computers to read the entire text content of the internet on a regular basis, whereas a human can't read all of just the current version of the English language Wikipedia once in their lifetime.
But current AI is very sample-inefficient: if a human were to read as much as an LLM, they would be world experts at everything, not varying between "secondary school" and "fresh graduate" depending on the subject… but even that description is misleading, because humans have the System 1/System 2[0] distinction and limited attention[1], whereas LLMs pay attention to approximately everything in the context window and (seem to) be at a standard between our System 1 and System 2.
If you're asking about an LLM's intelligence because you want to replace an intern, then the AI are intelligent; but if you're asking because you want to know how many examples they need in order to decode North Sentinelese or Linear A, then (from what I understand) these AI are extremely stupid.
It doesn't matter if a submarine swims[2], it still isn't going to fit into a flooded cave.
[0] https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow
[1] https://youtu.be/vJG698U2Mvo?si=omf3xleqPw5u6Y2k
https://youtu.be/ubNF9QNEQLA?si=Ja-9Ak4iCbcxbWdh
https://youtu.be/v3iPrBrGSJM?si=9cKHXEvEGl764Efa
https://www.americanbar.org/groups/intellectual_property_law...
[2] https://www.goodreads.com/quotes/32629-the-question-of-wheth...
It would take a human more than one lifetime to read everything most LLMs have read. I often have trouble remembering specifics of something I read an hour ago, never mind a decade ago. I can't imagine a human being an expert in something they read 300 years ago.
I used to think this too, but now I'm not so sure.
There's an influential school of thought arguing that one of the primary tasks of the brain is to predict sensory input, e.g. to take sequences of input and predict the next observation. This perspective explains many phenomena in perception, motor control, and more. In an abstract sense, it's not that different from what LLMs do -- take a sequence and predict the next item in a sequence.
The System 1 snd System 2 framework is appealing and quite helpful, but let's unpack it a little. A mode of response is said to be "System 1" if it's habitual, fast, and implemented as a stimulus-response mapping. Similar to LLMs, System 1 is a "lookup table" of actions.
System 2 is said to be slow, simulation-based, etc. But tasks performed via System 2 can transition to System 1 through practice ('automaticity' is the keyword here). Moreover, pre-automatic slow System 2 actions are compositions are simpler sets of actions. You deliberate about how to compose a photograph, or choose a school for your children, but many of the component actions (changing a camera setting, typing in a web URL) are habitual. It seems to me that what people call System 2 actions are often "stitched-together" System 1 behaviors. Solving calculus problem may be System 2, but adding 2+3 or writing the derivative of x^2 is System 1. I'm not sure the distinction between Systems 1 and 2 is as clear as people make it out to be -- and the effects of practice make the distinction even fuzzier.
What does System 2 have that LLMs lack? I'd argue: a working memory buffer. If you have working memory, you can then compose System 1 actions. In a way, a Turing machine is System 1 rule manipulation + working memory. Chain-of-thought is a hacky working memory buffer, and it improves results markedly. But I think we could do better with more intentional design.
[1] https://mitpress.mit.edu/9780262516013/bayesian-brain/ [2] https://pubmed.ncbi.nlm.nih.gov/35012898/
Can't we consider the context window to be memory?
And eventually wont we have larger and larger context windows, and perhaps even individual training where part of the context window 'conversation' is also fed back into the training data?
Seems like this will come to pass eventually.
Also, take this more advanced example: if you built a sufficiently large lookup table (of course, you need (possibly human) intelligence to build it, and it would be of impractical size), you can build a chatbot completely indistinguishable from a human that nonetheless doesn't seem like it really is intelligent (or conscious for that matter). The only operations it would perform would be some sort of decoding of the input into an astronomically large number to retrieve from the lookup table, and then using some (possibly rudimentary) seeking apparatus to retrieve the content that corresponds to the input. Your input size can be arbitrarily large enabling arbitrarily long conversations. It seems that to judge intelligence we really need to examine the internals and look at their structure.
I have a hunch that our particular 'feeling of consciousness' stems from the massive interconnectivity of the brain. All sorts of neurons from around the brain have some activity all the time (I believe the brain's energy consumption doesn't vary greatly with activity, so we use it), and whatever we perceive goes through an enormous number of interconnected neurons (representing concepts, impressions, ideas, etc.); unlike typical CPU-based algorithms that process a somewhat large number (potentially 100s of millions) of steps sequentially. My intuition does seem to pair with modern neural architectures (i.e. they could have some consciousness?), but I really don't know how we could do this sort of judgement before understanding better the details of the brain and other properties of cognition. I think this is a very important research area, and I'm not sure we have enough people working on it or using the correct tools (it relies heavily on logic, philosophy and metaphysical arguments; as well as require personal insight from being conscious :) ).
[1] From sources I can't remember, and from wiki: https://en.wikipedia.org/wiki/Turing_test#Loebner_Prize
It was provocative back then, and still is!
The idea of having my consciousness continue its existence indefinitely in somebody's kubernetes cluster, however, seems like a very special vision of hell.
Also a provocative read.
https://news.ycombinator.com/item?id=39751509
What is the computational equivalent of a biological neuron? How much of the detailed chemistry is really relevant to it's operation (doing useful computation), or can we just use the ANN model of synapses as weights, and the neuron itself as a summation device plus a non-linearity?
Maybe we can functionally model the human brain at a much coarser, more efficient, level than individual neurons - at cortical mini/macro column level perhaps, or as an abstract architecture not directly related to nature's messy implementation?
I believe that a supercomputer from the top 5 can already simulate a human brain. But we don't have the data and the theory yet. We are developing it, though.
An intelligent question is: how does one machine compare to the human brain on one dimension of performance?
A slide rule easily outperforms the human brain on some mathematical computations.
George Hotz has a nice series of posts on this:
https://geohot.github.io/blog/jekyll/update/2020/08/20/a-sil...
https://geohot.github.io/blog/jekyll/update/2022/02/17/brain...
https://geohot.github.io/blog/jekyll/update/2023/04/26/a-per...
I sorted them by date.
> We’ll use the estimates of 100 billion neurons and 100 trillion synapses for this post. That’s 100 teraweights.
... or maybe actual synapses cannot be described with a single weight.
> The max firing rate seems to be 200hz [4]. I really want an estimate of “neuron lag” here, but let’s upper bound it at 5ms.
... but ANNs output float activations per pass. Biological neurons encode values with sequences of spikes which vary in timing, so the firing rate doesn't on its own tell you the rate at which neurons can communicate new values.
> Remember also, that the brain is always learning, so it needs to be doing forward and backward passes. I’m not exactly sure why they are different, but [6] and [7] point to the backward pass taking 2x more compute than the forward pass.
... but the brain probably isn't doing backprop, in part because it doesn't get to observe the 'correct' output, compute a loss, etc, and because the brain isn't a DAG.