On Fri, Mar 29, 2024 at 09:55:28AM -0400, John Clark wrote: > On Thu, Mar 28, 2024 at 9:27 PM Russell Standish <li...@hpcoders.com.au> > wrote: > > > >"So to compare apples with apples - the human brain contains around 700 > trillion (7E14) synapses" > > > I believe 700 trillion is a more than generous estimate of the number of > synapses in the human brain, but I'll let it go. > > > > >"which would roughly correpond to an AI's parameter count > > > > NO! Comparing the human brain's synapses to the number of parameters that an > AI > program like GPT-4 has is NOT comparing apples to apples, it's comparing > apples > to oranges because the brain is hardware but GPT-4 is software. So let's > compare the brain hardware that human intelligence is running on with the > brain > hardware that GPT-4 is running on, that is to say let's compare synapses to > transistors. I'll use your very generous estimate and say the human brain has > 7*10^14 synapses, but the largest supercomputer in the world, the Frontier > Computer at Oakridge, has about 2.5*10^15 transistors, over three times as > many. And we know from experiments that a typical synapse in the human brain > "fires" between 5 and 50 times per second, but a typical transistor in a > computer "fires" about 4 billion times a second (4*10^9). That's why the > Frontier Computer can perform 1.1 *10^18 floating point calculations per > second > and why the human brain can not.
There is a big difference between the way transistors are wired in a CPU and the way neurons are wired up in a brain. The brain is not optimised at all to do floating point calculations, which is why even the most competent "computer" (in the old fashioned sense) can only manage sub 1 flops. Conversely, using floating point operations to perform neural network computations is not exactly efficient either. We're using GPUs today, because they can perform these very fast, and its a massively parallel operation, and GPUs are cheap, for what they are. In the future, I would expect we'd have dedicate neural processing units, based on memristors, or whatever. Indeed Intel is now flogging chips with "NPU"s, but how much of that is real and how much is marketing spin I can't say. The comparing synapses with ANN parameters is only relevant for the statement "we can simulate a human brain sized ANN by X date". Kurzweil didn't say that (for some reason I thought he did), he said human intelligence parity (which I supose could be taken to be avergae intelligence, or an IQ of 100). In a human brain, a lot of neurons are handling body operations - controlling muscles, interoception, proprioception, endocrine control etc, so the actual figure related to language processing is likely to be far smaller than the figure given. But only by an order of magnitude, I would say. > > I should add that although there have been significant improvements in the > field of AI in recent years, the most important being the "Attention Is All > You > Need" paper, I believe that even if transformers had never been discovered the > AI explosion that we are currently observing would only have been delayed by a > few years because the most important thing driving it forward is the brute > force enormous increase in raw computing speed. > > > > "He [Ray Kurzweil] was predicting 2029 to be the time when AI will > attain human level intelligence." > > > It now looks like Ray was being too conservative and 2024 or 2025 would be > closer to the Mark, and 2029 would be the time when an AI is smarter than the > entire human race combined. > 2025 should see the release of GPT5. It is still at least two orders of magnitude short of the mark IMHO. It is faster though - training GPT5 will have taken about 2 years, whereas it takes nearly 20 years to train a human. > > > > "I would still say that creativity (which is an essential prerequisite) > is still mysterious" > > > It doesn't matter if humans find creativity to be mysterious because we have > an > existence proof that a lack of understanding of creativity does not prevent > humans from making a machine that is creative. That may be the case, but understanding something does accelerate process dramatically over blind "trial and error". It is the main reason for the explosion in technical prowess over the last 400 years. > Back in 2016 when a computer > beat Lee Sedol, the top human champion at the game of GO, the thing that > everybody was talking about was move 37 of the second game of the five game > tournament. When the computer made that move the live expert commentators were > shocked and described it as "practically nonsensical" and "something no human > would do", and yet that crazy "nonsensical" move was the move that enabled the > computer to win. Lee Sedol said move 37 was "an incredible move" and was > completely unexpected and made it impossible for him to win, although it took > him a few more moves before he realized that. If a human had made moves 37 > every human GO expert on the planet would've said it was the most creative > move > they had ever seen. > Yes - I have said there is a glimmering of creativity. There are numerous such examples, I don't discount that. > > > "But singularity requires that machines design themselves" > > > Computers are already better at writing software than the average human, and > major chip design and manufacturing companies like NVIDIA, AMD, Intel , > Cerebras and TSMC are investing heavily in chip design software. > The average human doesn't code. This is not a valid comparison. Computers are still not at the level of an inexperienced intern. At best, they can provide suggestions, which if appropriate, can save a developer time, so could be used in a form of pair programming. As for unsupervised coding, they might be able to help writing unit tests, or other fairly boilerplate, but to actually let one loose on a codebase would be a net negative, if I'm to believe the reports. I haven't tried the tech yet - mainly because it will take some time out of my schedule to even set things up to work in my environment - but I do intend to when my current crunch period has receded somewhat. As for chip design software, this is software that assist a human designer. Performing circuit layouts is combinatorically difficult problem that is hard even for computers. Actually designing circuits from scratch without a human engineer in the loop is still a way off. It'll be some sort of evolutionary algorithm for doing this I expect - John Koza has done a lot of interesting work in this area with genetic programming for example. > > > > Anyway my 2c - I know John is keen to promote the idea of singularity > this decade - but I don't see it myself. > > > One thing I know for certain, whenever the Singularity occurs most people will > be surprised, otherwise it wouldn't be a Singularity. > Most people - sure - they don't tend to think about these things. But not us. We're expecting it - and there will be certain milestones that need to be hit first. Once hit, though, there will otherwise little warning. > John K Clark See what's on my new list at Extropolis > oib > > > > > > > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email > to everything-list+unsubscr...@googlegroups.com. > To view this discussion on the web visit https://groups.google.com/d/msgid/ > everything-list/ > CAJPayv3HNhk6ufAiQjjeK419CqpSubiJp%3DnTpPefCSADYs0Osg%40mail.gmail.com. -- ---------------------------------------------------------------------------- Dr Russell Standish Phone 0425 253119 (mobile) Principal, High Performance Coders hpco...@hpcoders.com.au http://www.hpcoders.com.au ---------------------------------------------------------------------------- -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to everything-list+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/everything-list/20240330022822.GB2268%40zen.