As you can imagine, AI is very complex programming and can
only run on the fastest computers available at the time, which
are known as supercomputers (despite Hollywood fantasies like
Electric Dreams (1984), AI does not run on just a
PC).
There have been AI "springs", when AI is the hot topic in research funding, startup investing, and the media, and AI "winters", when these all go cold for AI. We are currently in an AI spring.
These booms and busts are intimately related to speed advances in supercomputers. After AI has gone as far as it can go with current supercomputers, it must wait for significant advances in supercomputer speed before advancing itself.
For many years the speed of a supercomputer depended only on the speed of its single processor, as well as its memory, storage, and communication between these. This speed depended on the speed of the chips making up these parts. From the 1960's to the 1990's this followed Moore's Law and significant advances came relatively often (Ayrdos Board Member Dr. Steve Wozniak, co-founder and the engineering brains of Apple, can attest to this). This was the heyday of the Cray single-processor supercomputer.
There have been AI "springs", when AI is the hot topic in research funding, startup investing, and the media, and AI "winters", when these all go cold for AI. We are currently in an AI spring.
These booms and busts are intimately related to speed advances in supercomputers. After AI has gone as far as it can go with current supercomputers, it must wait for significant advances in supercomputer speed before advancing itself.
For many years the speed of a supercomputer depended only on the speed of its single processor, as well as its memory, storage, and communication between these. This speed depended on the speed of the chips making up these parts. From the 1960's to the 1990's this followed Moore's Law and significant advances came relatively often (Ayrdos Board Member Dr. Steve Wozniak, co-founder and the engineering brains of Apple, can attest to this). This was the heyday of the Cray single-processor supercomputer.
In the 1990's these speed advances plateaued, Cray
supercomputers were no more, and a new way to speed up
supercomputers was found (and Ayrdos Founder/CEO/Chairman
Dr. Duane
Thresher, Ayrdos Founder/President/Board Member
Dr. Claudia
Kubatzki, and Ayrdos Board Member Dr. Greg Newby,
CEO/Director of Project Gutenberg, all started working with
supercomputers). The era of parallel processing
supercomputers began and continues to this day.
If a program, like for AI, can be separated into pieces, each of which can be independently run at the same time, i.e. in parallel, on a different processor of a supercomputer with multiple processors, then a significant speedup can be achieved; seemingly, a speedup factor that is equal to the number of processors, so the more processors, the more the speedup.
Unfortunately, and what someone (like those at Ayrdos) will only know if they've parallel programmed on supercomputers, parallel programming is significantly more difficult than normal programming (if normal programming is chess, parallel programming is 3-dimensional chess) and the speedup is actually limited. Below is a graph of Amdahl's Law. The horizontal axis is the number of processors the program is running on and the vertical axis is the speedup compared to running on just one processor.
The colored curves show that the more (percent of total) the
program is parallelized, the more speedup can be achieved.
But again, parallel programming is difficult; sometimes it is
faster to run the program on one processor than to take the
time to parallel program (parallelize).
Further, as the flattening of the curves show, after some number of processors, there is no additional speedup and additional processors are a waste, of availability, money, and energy. This limitation is due to the overhead involved in using multiple processors, which increases with the number of processors; i.e. it takes a more quickly increasing number of processors to control an increasing number of processors (particularly the communication between them).
The optimal number of processors can only be determined by testing by those (like at Ayrdos) experienced in parallel programming on supercomputers. Those new to the field will just wastefully assume that the more processors the faster and there will be no indication that this is not the case. When thousands of processors are involved, it becomes a monopolizing waste of availability, a bankrupting waste of money, and an environmentally damaging waste of energy.
If a program, like for AI, can be separated into pieces, each of which can be independently run at the same time, i.e. in parallel, on a different processor of a supercomputer with multiple processors, then a significant speedup can be achieved; seemingly, a speedup factor that is equal to the number of processors, so the more processors, the more the speedup.
Unfortunately, and what someone (like those at Ayrdos) will only know if they've parallel programmed on supercomputers, parallel programming is significantly more difficult than normal programming (if normal programming is chess, parallel programming is 3-dimensional chess) and the speedup is actually limited. Below is a graph of Amdahl's Law. The horizontal axis is the number of processors the program is running on and the vertical axis is the speedup compared to running on just one processor.
Further, as the flattening of the curves show, after some number of processors, there is no additional speedup and additional processors are a waste, of availability, money, and energy. This limitation is due to the overhead involved in using multiple processors, which increases with the number of processors; i.e. it takes a more quickly increasing number of processors to control an increasing number of processors (particularly the communication between them).
The optimal number of processors can only be determined by testing by those (like at Ayrdos) experienced in parallel programming on supercomputers. Those new to the field will just wastefully assume that the more processors the faster and there will be no indication that this is not the case. When thousands of processors are involved, it becomes a monopolizing waste of availability, a bankrupting waste of money, and an environmentally damaging waste of energy.
A multi-processor parallel processing supercomputer is
basically a cluster of communicating single-processor
computers. Since the chips of these single-processor
computers have ceased to often have significant speed
increases, they can just be high-end commercially available
("commodity") computers.
However, a real supercomputer is not just a cluster of communicating single-processor computers. In real supercomputers, the discussed slowing overhead is greatly minimized, including by making the inter-processor communication (i.e. networking) as fast as possible and minimizing what each processor has to do (just the program calculations).
Such supercomputers were developed in the 1990's at university and government research institutions (particularly NASA, where Ayrdos Founder/CEO/Chairman Dr. Duane Thresher was). They later became available at big tech corporations.
Unfortunately and again, parallel programming on supercomputers is difficult and not easy to get experience at so not many could ever do it. Further, each of the single-processor computers clustered in a supercomputer can be used individually, providing a high-end computer to a user, who otherwise might not have one available.
The result was that supercomputers devolved into nothing more than computer clusters, including losing the overhead minimization so critical to real supercomputers — most users wanted their single-processor computers to do everything and didn't care about inter-processor communication.
Seizing a business opportunity, big tech corporations, like Microsoft, started to make the individual computers in a cluster available online (these online clusters became "the cloud"), for a fee. They used operating systems like Kubernetes, which can be characterized as being the opposite of overhead minimization (it stresses versatility and redundancy for example).
To cash in and control AI, big tech corporations, like Microsoft, pretended to AI startups that their computer clusters could be used as real supercomputers and encouraged these AI startups to use as many processors as possible, implying the more, the faster but really meaning the more, the more money. New to parallel programming and supercomputers, those at AI startups didn't know any better.
To make matters worse, these big tech corporations, like Microsoft, claimed that because so many processors were being used, there was a shortage, and they had to go looking for more energy sources, like nuclear, to run them (regardless of environmental damage). Thus they had to charge more for each processor.
Processors became one of the largest costs at AI startups. The only way they could see to deal with it was to accept processors from these big tech corporations, like Microsoft, as "investments" in their company, in return for control of it, and thus AI in general. Further, once a big tech corporation, like Microsoft, controls a startup, innovation dies.
However, a real supercomputer is not just a cluster of communicating single-processor computers. In real supercomputers, the discussed slowing overhead is greatly minimized, including by making the inter-processor communication (i.e. networking) as fast as possible and minimizing what each processor has to do (just the program calculations).
Such supercomputers were developed in the 1990's at university and government research institutions (particularly NASA, where Ayrdos Founder/CEO/Chairman Dr. Duane Thresher was). They later became available at big tech corporations.
Unfortunately and again, parallel programming on supercomputers is difficult and not easy to get experience at so not many could ever do it. Further, each of the single-processor computers clustered in a supercomputer can be used individually, providing a high-end computer to a user, who otherwise might not have one available.
The result was that supercomputers devolved into nothing more than computer clusters, including losing the overhead minimization so critical to real supercomputers — most users wanted their single-processor computers to do everything and didn't care about inter-processor communication.
Seizing a business opportunity, big tech corporations, like Microsoft, started to make the individual computers in a cluster available online (these online clusters became "the cloud"), for a fee. They used operating systems like Kubernetes, which can be characterized as being the opposite of overhead minimization (it stresses versatility and redundancy for example).
To cash in and control AI, big tech corporations, like Microsoft, pretended to AI startups that their computer clusters could be used as real supercomputers and encouraged these AI startups to use as many processors as possible, implying the more, the faster but really meaning the more, the more money. New to parallel programming and supercomputers, those at AI startups didn't know any better.
To make matters worse, these big tech corporations, like Microsoft, claimed that because so many processors were being used, there was a shortage, and they had to go looking for more energy sources, like nuclear, to run them (regardless of environmental damage). Thus they had to charge more for each processor.
Processors became one of the largest costs at AI startups. The only way they could see to deal with it was to accept processors from these big tech corporations, like Microsoft, as "investments" in their company, in return for control of it, and thus AI in general. Further, once a big tech corporation, like Microsoft, controls a startup, innovation dies.
If all this seems unbelievable, see the movie The Big
Short (2015), or read the book by Michael Lewis it was
based on, about the 2007 subprime mortgage crisis. That
similarly unbelievable story is also true.
While doing research for his former company, Apscitu Inc, Ayrdos Founder/CEO/Chairman Dr. Duane Thresher came across major AI startup OpenAI, of ChatGPT fame, desperately seeking supercomputer engineers. He discovered how, and who, was doing OpenAI's "supercomputing". They were complete novices using thousands of Microsoft (Azure cloud) processors running under Kubernetes, with no mention of Amdahl's Law.
Dr. Thresher also noted in the news that OpenAI was being taken over by Microsoft, as described. Combined with other big tech corporations taking over other AI startups, this would ultimately mean control of AI by "Big Tech", which had already proven it could not be trusted with control of information.
With a background in fighting Big Tech (about control of email for example), as well as fighting overstepping government and universities, Dr. Thresher set out to fight Big Tech control of AI (this is also one of the reasons, particularly against anti-Apple Microsoft, that Dr. Steve Wozniak, co-founder of Apple, became an Ayrdos board member). He realized that fighting Big Tech control of AI would have to be done on two fronts.
On the first front, Dr. Thresher founded Ayrdos AI Supercomputing Inc, a company of experienced, real supercomputing experts. Ayrdos works with AI startups as customers to help parallelize their programs, determine the optimal number of real supercomputer processors the programs need, and then put together real supercomputers with just this number of processors, so that they are affordable and AI startups don't have to sell their souls to Big Tech. (Ayrdos is also a great opportunity for non-Big Tech investors.)
On the second front, Dr. Thresher is spearheading a public fight against Big Tech control of AI, including legal action and government lobbying, just like Big Tech does with its nearly unlimited money. This is paid for by smart donators who, especially after experiencing Big Tech's control of information, understand the massive danger of Big Tech control of AI.
While doing research for his former company, Apscitu Inc, Ayrdos Founder/CEO/Chairman Dr. Duane Thresher came across major AI startup OpenAI, of ChatGPT fame, desperately seeking supercomputer engineers. He discovered how, and who, was doing OpenAI's "supercomputing". They were complete novices using thousands of Microsoft (Azure cloud) processors running under Kubernetes, with no mention of Amdahl's Law.
Dr. Thresher also noted in the news that OpenAI was being taken over by Microsoft, as described. Combined with other big tech corporations taking over other AI startups, this would ultimately mean control of AI by "Big Tech", which had already proven it could not be trusted with control of information.
With a background in fighting Big Tech (about control of email for example), as well as fighting overstepping government and universities, Dr. Thresher set out to fight Big Tech control of AI (this is also one of the reasons, particularly against anti-Apple Microsoft, that Dr. Steve Wozniak, co-founder of Apple, became an Ayrdos board member). He realized that fighting Big Tech control of AI would have to be done on two fronts.
On the first front, Dr. Thresher founded Ayrdos AI Supercomputing Inc, a company of experienced, real supercomputing experts. Ayrdos works with AI startups as customers to help parallelize their programs, determine the optimal number of real supercomputer processors the programs need, and then put together real supercomputers with just this number of processors, so that they are affordable and AI startups don't have to sell their souls to Big Tech. (Ayrdos is also a great opportunity for non-Big Tech investors.)
On the second front, Dr. Thresher is spearheading a public fight against Big Tech control of AI, including legal action and government lobbying, just like Big Tech does with its nearly unlimited money. This is paid for by smart donators who, especially after experiencing Big Tech's control of information, understand the massive danger of Big Tech control of AI.