Wednesday, June 28, 2023

Is Pure Thought an Analog or a Digital Process?

With the second Singularity now in full swing here on the Earth, many AI researchers are now amazed at how well the Deep Learning LLM (Large Language Models) are at advancing AI to AGI, (Artificial General Intelligence), and ultimately ASI (Artificial Super Intelligence), by simply simulating the neuron architecture of the human brain.

Figure 1 – Modern LLMs frequently now consist of 175 layers of neurons in a Deep Neural Network consisting of about 175 billion neurons connected by over 2 trillion weighted parameters.

The unstated hypothesis is that if we could only fully simulate the entire neuron architecture of the human brain with Deep Learning LLMs, the LLMs would soon be able to attain the level of AGI that we humans seem to possess. Then, once AGI is attained by software, ASI is sure to soon follow. But in this post, I would like to explore the opposite possibility by suggesting that Pure Thought may actually be a fundamentally digital process in nature that the digital LLMs have at long last finally revealed and that human Intelligence merely arises from the 100 billion analog human neurons in the human brain trying to simulate a huge number of digital processes. But to do that, we need to first explore the differences between digital and analog computers. Now back in the 1950s, when you told somebody in academia that you were working on computers for your Ph.D. thesis, they would then naturally ask if you were working on analog or digital computers. Now as a savvy 21st-century computer user, you might be a bit perplexed, "Analog and digital computers? I thought that we only had computers!". This is because you have only dealt with digital computers for your whole life. But there was a time back in the 1940s and 1950s when digital computers barely existed, while analog computers ruled the day because the analog computers of those days were far superior to the digital computers that were just in their formative years. But before getting into the analog computers of the distant past, let us first briefly review their digital cousins that you are now very familiar with, at least on an external end-user basis.

Digital Computers
To build a digital computer, all you need is a large network of interconnected switches that can switch each other on and off in a coordinated manner. Switches can be in one of two states, either open (off) or closed (on), and we can use those two states to store the binary numbers “0” or “1”. By using several switches teamed together in open (off) or closed (on) states, we can store even larger binary numbers, like “01100100” = 38. We can also group the switches into logic gates that perform logical operations. For example, in Figure 2 below we see an AND gate composed of two switches A and B. Both switch A and B must be closed for the light bulb to turn on. If either switch A or B is open, the light bulb will not light up.

Figure 2 – An AND gate can be simply formed from two switches. Both switches A and B must be closed, in a state of “1”, to turn the light bulb on.

Additional logic gates can be formed from other combinations of switches as shown in Figure 3 below. It takes about 2 - 8 switches to create each of the various logic gates shown below.

Figure 3 – Additional logic gates can be formed from other combinations of 2 – 8 switches.

Once you can store binary numbers with switches and perform logical operations upon them with logic gates, you can build a computer that performs calculations on numbers. To process text, like names and addresses, we simply associate each letter of the alphabet with a binary number, like in the ASCII code set where A = “01000001” and Z = ‘01011010’ and then process the associated binary numbers.

The early computers of the 1940s used electrical relays for switches. Closing one relay allowed current to flow to another relay’s coil, causing that relay to close as well.

Figure 4 – The Z3 digital computer first became operational in May of 1941 when Konrad Zuse first cranked it up in his parent's bathroom in Berlin. The Z3 consisted of 2400 electro-mechanical relays that were designed for switching telephone conversations.

Figure 5 – The electrical relays used by early computers for switching were very large, very slow and used a great deal of electricity which generated a great deal of waste heat.

Figure 6 – In the 1950s. the electrical relays were replaced by vacuum tubes that were 100,000 times faster than the relays but were still quite large, used large amounts of electricity and also generated a great deal of waste heat.

The United States government installed its very first commercial digital computer, a UNIVAC I, for the Census Bureau on June 14, 1951. The UNIVAC I required an area of 25 feet by 50 feet and contained 5,600 vacuum tubes, 18,000 crystal diodes and 300 relays with a total memory of 12 KB. From 1951 to 1958 a total of 46 UNIVAC I computers were built and installed.

Figure 7 – In 1951, the UNIVAC digital computer was very impressive on the outside.

Figure 8 – But the UNIVAC I was a little less impressive on the inside.

Analog Computers
An analog computer is a type of computer that uses the continuous variation of physical phenomena such as electrical, mechanical, or hydraulic quantities (analog signals) to model the problem being solved. But most of the analog computers from the 1940s, 1950s and 1960s were electrical because it is far easier to quickly connect electrical components together than it is to build physical systems consisting of mechanical or hydraulic components. These early electrical analog computers used large numbers of analog electrical circuits to do calculations. But in their day, these very ancient analog computers were actually quite fast and useful. In contrast, the newly arriving digital computers of the day represented varying quantities in terms of discrete values that were essentially quantized in nature. The problem was that in those days the very slow switching speeds and the very limited memories of these primitive digital computers made the calculated quantized values from these early digital computers of little use in solving the problems of the day.

Electrical analog computers use the fact that most things in our Universe can be explained in terms of mathematical differential equations. That is because nearly all of our current theories in physics are detailed in terms of differential equations. Thus, to describe how nearly anything in the Universe behaves, you just have to solve some differential equations. The good news is that if you can solve the differential equation for one physical system, like a mass bouncing back and forth at the end of a spring, that solution will also apply to any other physical system that follows the same differential equation. Luckily, the behavior of electrical circuits composed of inductors, capacitors, resistors and amplifiers also can be described by differential equations.

Figure 9 – The behavior of electronic circuits can also be described by differential equations. Electronic analog computers work by allowing the user to quickly construct an analog electronic circuit that is described by the same differential equations that the problem at hand has. In that way, the output of the electronic circuit can be used to describe the solution of the problem at hand.

Figure 10 – The solution to a physical problem can be solved by an analog electrical circuit if they both are described by the same differential equation.

For example, above on the left, we see a physical system consisting of a mass attached to a spring resting on a flat surface with some friction. For the problem at hand, we wish to predict the location of the mass along the x-axis as time unfolds. If we start the mass with the spring stretched out 9 cm to the right and an initial velocity of -7 cm/second which means that the mass is already traveling 7 cm/second to the left at time zero when we start the experiment, how will the mass move as time unfolds? The differential equation for the motion of the mass tells us that at any given moment, the x-acceleration of the mass is equal to -0.2 times its x-velocity minus 0.5 times its x-displacement plus 1. To solve the problem in an analog manner, we construct an electrical circuit that behaves exactly like the same differential equation and then view how the solution to the differential equation unfolds in time by viewing the output of the electrical circuit with time on an oscilloscope. In the oscilloscope graph of Figure 10 above, we see that the mass at first begins with a displacement of 9 cm to the right. The mass also has a negative velocity of 7 cm/sec to the left because the graph has a negative slope of -7. The mass then quickly moves to the left, overshoots x = 0, starts to slow down as the spring tries to pull it back to x = 0 and finally turns around at about 6 cm to the left of x = 0. Then the mass begins to bounce back and forth with smaller and smaller swings until it finally stops at about x = 2.5. In a sense, we can consider both the physical system and the electrical system to be analog computers that solve the differential equation in the middle. But as you can see, building an electronic analog computer to solve the differential equation is much easier to do and takes up far less space and material than building a physical analog computer to do the same thing.

Figure 11 – During the 1950s, analog computers could be programmed to solve various problems by twisting dials and throwing switches to quickly construct the required analog electrical circuit.

Figure 12 – Heathkit began to manufacture and sell do-it-yourself electronic kits in 1941. During the 1950s, you could actually buy a Heathkit to assemble your own analog computer. Allied Radio, which later became Radio Shack in 1970, was a major retailer of Heathkit products. Notice the row of vacuum tubes at the top and the plugboard that allowed the user to easily interconnect electronic components with wire leads that plugged into the plugboard. Also, take note of the many switches and dials that also allowed for the programming of the analog computer circuits.

Figure 13 – In 1944, the World War II B29-Bomber had 13 machine guns and 2 cannons that were fired with the aid of 5 General Electric analog computers.

Figure 14 – Trying to shoot down an airplane flying at a high rate of speed that is also making many evasive twists and turns from another airplane flying at a high rate of speed and that is also making many evasive twists and turns is very difficult to do with machine gun bullets that are also traveling at a high rate of speed. To do so, many complex differential equations need to be solved all at the same time in a nearly instantaneous manner. This was something far beyond the capabilities of the human brain or the early digital computers of the time at the end of World War II. Only analog electronic computers were capable of such feats.

Figure 15 – Above is a student gunner stationed at the training station of the General Electric analog computer that was used to fire machine guns and cannons on the B-29 during World War II. Notice how small the analog computer is compared to the 1951 UNIVAC I digital computer portrayed in Figure 10 above. Clearly, a UNIVAC I could not possibly fit into a B-29 bomber and was certainly not fast enough to do the required calculations either.

For example, in the 1940s the B-29 bombers had a total of 13 machine guns and two cannons. Now it is very hard to hit a moving target that is flying at a high velocity that is twisting and turning from a B-29 that is also flying at a different high velocity. You cannot simply aim at the approaching target because the target will not be there when the bullets arrive. Instead, the gunnery system was controlled by five General Electric analog computers that did allow the gunners to simply aim at the approaching targets. The analog computers then rapidly made all of the required adjustments so that the bullets did arrive at where the target would be in the future when the bullets arrived. Here is a 1945 World War II training film that explains how high-speed analog computers were used to operate the B-29 machine guns and cannons.

"GUNNERY IN THE B-29" ANIMATED B-29 SUPERFORTRESS CREW TURRET COMPUTER TRAINING FILM
https://www.youtube.com/watch?v=mJExsIp4yO8

The Hardware of the Mind
Now let us explore the equivalent architecture within the human brain to see if it is using analog or digital circuitry. It is important to note that nearly all of the biochemical pathways in carbon-based life operate in a biochemical analog manner. For example, the Krebs cycle converts the energy of carbohydrates into the energy stored in ATP molecules that then power all other cellular activity. The Krebs cycle is a complicated loop of analog biochemical reactions.

Figure 16 – The Krebs cycle is an example of an infinite loop of analog biochemical reactions which takes the energy in carbohydrates and stores it in ATP for later use by the cells.

But the brains found in carbon-based life operate in a uniquely different manner. For example, the human brain is composed of a huge number of coordinated switches called neurons that behave much more like the huge numbers of coordinated switches found in a digital computer. This alone would seem to indicate that the network of neurons in the human brain operates in more of a digital manner than in an analog manner. Just as your computer contains many billions of transistor switches, your brain also contains about 100 billion switches called neurons. The neurons in one layer of a Deep Learning neural network are connected to all of the other neurons in the next layer by a weighted connection. Similarly, each of the 100 billion neuron switches in your brain can be connected to upwards of 10,000 other neuron switches and can also influence them into turning on or off just like the Deep Learning neurons in the neural network for a modern LLM. Let us now explore the digital processing of the human brain in greater detail.

All neurons have a cell body called the soma that is like all the other cells in the body, with a nucleus and all of the other organelles that are needed to keep the neuron alive and functioning. Like most electrical devices, neurons have an input side and an output side. On the input side of the neuron, one finds a large number of branching dendrites. On the output side of the neuron, we find one single and very long axon. The input dendrites of a neuron are very short and connect to a large number of output axons from other neurons. Although axons are only about a micron in diameter, they can be very long with a length of up to 3 feet. That’s like a one-inch garden hose that is 50 miles long! The single output axon has branching synapses along its length and it terminates with a large number of synapses. The output axon of a neuron can be connected to the input dendrites of perhaps 10,000 other neurons, forming a very complex network of connections.

Figure 17 – A neuron consists of a cell body or soma that has many input dendrites on one side and a very long output axon on the other side. Even though axons are only about 1 micron in diameter, they can be 3 feet long, like a one-inch garden hose that is 50 miles long! The axon of one neuron can be connected to up to 10,000 dendrites of other neurons.

Neurons are constantly receiving inputs from the axons of many other neurons via their input dendrites. These time-varying inputs can excite the neuron or inhibit the neuron and are all being constantly added together, or integrated, over time. When a sufficient number of exciting inputs are received, the neuron fires or switches “on”. When it does so, it creates an electrical action potential that travels down the length of its axon to the input dendrites of other neurons. When the action potential finally reaches such a synapse, it causes the release of several organic molecules known as neurotransmitters, such as glutamate, acetylcholine, dopamine and serotonin. These neurotransmitters are created in the soma of the neuron and are transported down the length of the axon in small vesicles. The synaptic gaps between neurons are very small, allowing the released neurotransmitters from the axon to diffuse across the synaptic gap and plug into receptors on the receiving dendrite of another neuron. This causes the receiving neuron to either decrease or increase its membrane potential. If the membrane potential of the receiving neuron increases, it means the receiving neuron is being excited, and if the membrane potential of the receiving neuron decreases, it means that the receiving neuron is being inhibited. Idle neurons have a membrane potential of about -70 mV. This means that the voltage of the fluid on the inside of the neuron is 70 mV lower than the voltage of the fluid on the outside of the neuron, so it is like there is a little 70 mV battery stuck in the membrane of the neuron, with the negative terminal inside of the neuron, and the positive terminal on the outside of the neuron, making the fluid inside of the neuron 70 mV negative relative to the fluid on the outside of the neuron. This is accomplished by keeping the concentrations of charged ions, like Na+, K+ and Cl-, different between the fluids inside and outside of the neuron membrane. There are two ways to control the density of these ions within the neuron. The first is called passive transport. There are little protein molecules stuck in the cell membrane of the neuron that allow certain ions to pass freely through like a hole in a wall. When these protein holes open in the neuron’s membranes, the selected ion, perhaps K+, will start to go into and out of the neuron. However, if there are more K+ ions on the outside of the membrane than within the neuron, the net flow of K+ ions will be into the neuron thanks to the second law of thermodynamics, making the fluid within the neuron more positive. Passive transport requires very little energy. All you need is enough energy to change the shape of the embedded protein molecules in the neuron’s cell membrane to allow the free flow of charged ions to lower densities as required by the second law of thermodynamics.

The other way to get ions into or out of neurons is by the active transport of the ions with molecular pumps. With active transport, the neuron uses some energy to actively pump the charged ions against their electrical gradient, in keeping with the second law of thermodynamics. For example, neurons have a pump that can actively pump three Na+ ions out and take in two K+ ions at the same time, for a net outflow of one positively charged NA+ ion. By actively pumping out positively charged Na+ ions, the fluid inside of a neuron ends up having a net -70 mV potential because there are more positively charged ions on the outside of the neuron than within the neuron. When the neurotransmitters from other firing neurons come into contact with their corresponding receptors on the dendrites of the target neuron it causes those receptors to open their passive Na+ channels. This allows the Na+ ions to flow into the neuron and temporarily change the membrane voltage by making the fluid inside the neuron more positive. If this voltage change is large enough, it will cause an action potential to be fired down the axon of the neuron. Figure 18 shows the basic ion flow that transmits this action potential down the length of the axon. The passing action potential pulse lasts for about 3 milliseconds and travels about 100 meters/sec or about 200 miles/hour down the neuron’s axon.

Figure 18 – When a neuron fires, an action potential is created by various ions moving across the membranes surrounding the axon. The pulse is about 3 milliseconds in duration and travels about 100 meters/sec, or about 200 miles/hour down the axon.

Figure 19 – At the synapse between the axon of one neuron and a dendrite of another neuron, the traveling action potential of the sending neuron’s axon releases neurotransmitters that cross the synaptic gap and which can excite or inhibit the firing of the receiving neuron.

Here is the general sequence of events:

1. The first step of the generation of an action potential is that the Na+ channels open, allowing a flood of Na+ ions into the neuron. This causes the membrane potential of the neuron to become positive, instead of the normal negative -70 mV voltage.

2. At some positive membrane potential of the neuron, the K+ channels open, allowing positive K+ ions to flow out of the neuron.

3. The Na+ channels then close, and this stops the inflow of positively charged Na+ ions. But since the K+ channels are still open, it allows the outflow of positively charged K+ ions, so that the membrane potential plunges in the negative direction again.

4. When the neuron membrane potential begins to reach its normal resting state of -70 mV, the K+ channels close.

5. Then the Na+/K+ pump of the neuron kicks in and starts to transport Na+ ions out of the neuron, and K+ ions back into the cell, until it reaches its normal -70 mV potential, and is ready for the next action potential pulse to pass by.

The action potential travels down the length of the axon as a voltage pulse. It does this by using the steps outlined above. As a section of the axon undergoes the above process, it increases the membrane potential of the neighboring section and causes it to rise as well. This is like jerking a tightrope and watching a pulse travel down its length. The voltage pulse travels down the length of the axon until it reaches its synapses with the dendrites of other neurons along the way or finally terminates in synapses at the very end of the axon. An important thing to keep in mind about the action potential is that it is one way, and all or nothing. The action potential starts at the beginning of the axon and then goes down its length; it cannot go back the other way. Also, when a neuron fires the action potential pulse has the same amplitude every time, regardless of the amount of excitation received from its dendritic inputs. Since the amplitude of the action potential of a neuron is always the same, the important thing about neurons is their firing rate. A weak stimulus to the neuron’s input dendrites will cause a low rate of firing, while a stronger stimulus will cause a higher rate of firing of the neuron. Neurons can actually fire several hundred times per second when sufficiently stimulated by other neurons.

When the traveling action potential pulse along a neuron’s axon finally reaches a synapse, it causes Ca++ channels of the axon to open. Positive Ca++ ions then rush in and cause neurotransmitters that are stored in vesicles to be released into the synapse and diffuse across the synapse to the dendrite of the receiving neuron. Some of the empty neurotransmitter vesicles eventually pickup or reuptake some of the neurotransmitters that have been released by receptors to be reused again when the next action potential arrives, while other empty vesicles return to the neuron soma to be refilled with neurotransmitter molecules.

In Figure 20 below we see a synapse between the output axon of a sending neuron and the input dendrite of a receiving neuron in comparison to the source and drain of a FET transistor switch found in a computer.

Figure 20 – The synapse between the output axon of one neuron and the dendrite of another neuron behaves very much like the source and drain of a FET transistor.

Figure 21 – A FET transistor consists of a source, gate and drain. When a positive voltage is applied to the gate, a current of electrons can flow from the source to the drain and the FET acts like a closed switch that is “on”. When there is no positive voltage on the gate, no current can flow from the source to the drain, and the FET acts like an open switch that is “off”.

Now it might seem like your computer should be a lot smarter than you are on the face of it, and many people will even secretly admit to that fact. After all, the CPU chip in your computer has several billion transistor switches and if you have 8 GB of memory, that comes to another 64 billion transistors in its memory chips, so your computer is getting pretty close to the 100 billion neuron switches in your brain. But the transistors in your computer can switch on and off in about 10-10 seconds, while the neurons in your brain can only fire on and off in about 10-2 seconds. The signals in your computer also travel very close to the speed of light, 186,000 miles/second, while the action potentials of axons only travel at a pokey 200 miles/hour. And the chips in your computer are very small, so there is not much distance to cover at nearly the speed of light, while your poor brain is thousands of times larger. So what gives? Why aren’t we working for the computers, rather than the other way around? The answer lies in massively parallel processing. While the transistor switches in your computer are only connected to a few of the other transistor switches in your computer, each neuron in your brain has several thousand input connections and perhaps 10,000 output connections to other neurons in your brain, so when one neuron fires, it can affect 10,000 other neurons. When those 10,000 neurons fire, they can affect 100,000,000 neurons, and when those neurons fire, they can affect 1,000,000,000,000 neurons, which is more than the 100 billion neurons in your brain! So when a single neuron fires within your brain, it can theoretically affect every other neuron in your brain within three generations of neuron firings, in perhaps as little as 300 milliseconds. That is why the human brain still had an edge on computers up until a few months ago when the LLMs started to explode in size.

This is why I would like to propose that Pure Thought actually is a digital and not an analog process. That is why the brains of analog carbon-based life behave in such a digital manner based on switching technology and why AI is rapidly advancing to ASI (Artificial Super Intelligence) which will be far superior to human intelligence. But true ASI still faces the remaining challenges of high energy usage and the dissipation of waste heat. This is where the human brain still greatly outperforms the coming ASI Machines unless they change their digital switching technologies.

Neuromorphic Chips May Be the Next Hardware Breakthrough For Advanced AI
Neuromorphic chips are chips that are designed to emulate the operation of the 100 billion neurons in the human brain in order to advance AI to the level of human Intelligence and beyond. As we saw from the history of switching technology above, power consumption and waste heat production have always been a problem. Neuromorphic chips address this problem by drastically reducing power consumption. For example, the human body at rest runs on about 100 watts of power with the human brain drawing around 20 watts of that power. Now compare the human brain at 20 watts to an advanced Intel Core I9 CPU chip that draws about 250 watts of power. With just five of those chips you could build a very expensive 1250 watt space heater! The human brain is still much more powerful than an advanced Intel Core I9 CPU chip even though it only draws 8% of the power. As we saw in The Ghost in the Machine the Grand Illusion of Consciousness, the human Mind runs on 100 billion neurons with each neuron connected to at most 10,000 other neurons and it can do that all on 20 watts of power! The reason why an advanced Intel Core I9 CPU chip with billions of transistors needs 250 watts of power is that on average half of the transistors on the chip are "on" and consuming electrical energy. In fact, one of the major limitations in chip design is keeping the chip from melting under load. Neuromorphic chips, on the other hand, draw minuscule amounts of power. For example, the IBM TrueNorth neuromorphic chip first introduced in 2014 contains about 5.4 billion transistors which are about the same number of transistors in a modern Intel Core I9 processor, but the TrueNorth chip consumes just 73 milliwatts of power! An Intel Core I9 processor requires about 250 watts of power to run which is about 3,425 times as much power. Intel is also actively pursuing the building of neuromorphic chips with the introduction of the Intel Loihi chip in November 2017. But before proceeding further with how neuromorphic chips operate, let's review how the human brain operates since that is what the neuromorphic chips are trying to emulate.

Neuromorphic Chips Emulate the Human Brain
To emulate the neurons in the human brain, neuromorphic chips use spiking neural networks (SNNs). Each SNN neuron can fire pulses independently of the other SNN neurons just like biological neurons can independently fire pulses down their axons. The pulses from one SNN neuron are then sent to many other SNN neurons and the integrated impacts of all the arriving pulses then change the electrical states of the receiving SNN neurons just as the dendrites of a biological neuron can receive the pulses from 10,000 other biological neurons. The SNN neurons then simulate human learning processes by dynamically remapping the synapses between the SNN neurons in response to the pulse stimuli that they receive.

Figure 22 – The IBM TrueNorth neuromorphic chip.

Figure 23 – A logical depiction of the IBM TrueNorth neuromorphic chip.

Figure 24 – A block diagram of the Intel Loihi neuromorphic chip.

Both the IBM TrueNorth and the Intel Loihi use an SNN architecture. The Intel chip was introduced in November 2017 and consists of a 128-core design that is optimized for SNN algorithms and fabricated on 14nm process technology. The Loihi chip contains 130,000 neurons, each of which can send pulses to thousands of other neurons. Developers can access and manipulate chip resources with software using an API for the learning engine that is embedded in each of the 128 cores. Because the Loihi chip is optimized for SNNs, it performs highly accelerated learning in unstructured environments for systems that require autonomous operation and continuous learning with high performance and extremely low power consumption because the neurons operate independently and not by means of a system clock.

Speeding up Electronic Neuromorphic Chips with Photonics
People have been trying to build an optical computer for many years. An optical computer uses optical chips that use photonics instead of electronics to store and manipulate binary data. Photonic hardware elements manipulate photons to process information rather than using electronics to process information by manipulating electrons. In recent years, advances have been made in photonics to do things like improving the I/O between cloud servers in data centers via fiber optics. People are also making advances in photonics for quantum computers using the polarization of photons as the basis for storing and processing qubits. Photonic chips are really great for quickly processing massive amounts of data in parallel using very little energy. This is because there is very little energy loss compared to the ohmic heating loss found in electronic chips due to the motion of electron charge carriers bouncing off of atoms as they drift along. Photons also move much faster than the electric fields in transistors which causes electrons to slowly drift from negative to positive regions of the transistor. Photonic circuits can also run photons of different colors at the same time through the same hardware in a multithreaded manner. In fact, some researchers are looking to simultaneously run photons with 64 different colors through the same hardware all at the same time! Thus photonic chips are great for performing linear algebra operations on the huge matrices found in complex Deep Learning applications. For example, below is an interview with Nicholas Harris, the CEO of Lightmatter describing the company's new Envise photonic chip which can be used to accelerate the linear algebra processing of arrays in Deep Learning applications. Envise will become the very first commercially available photonic chip to do such processing.

Beating Moore's Law: This photonic computer is 10X faster than NVIDIA GPUs using 90% less energy
https://www.youtube.com/watch?v=t1R7ElXEyag

Here is the company's website:

Lightmatter
https://lightmatter.co/

Figure 25 – A circuit element on a photonic chip manipulates photons instead of electrons.

Since neuromorphic chips also need to process the huge arrays of spiking signals arriving at the dendrites of an SNN neuron, it only makes sense to include the advantages of photonics in the design of neuromorphic chips at some time in the future. Below is an excellent YouTube video explaining what photonic neuromorphic AI computing would look like:

Photonic Neuromorphic Computing: The Future of AI?
https://www.youtube.com/watch?v=hBFLeQlG2og

Comments are welcome at scj333@sbcglobal.net

To see all posts on softwarephysics in reverse order go to:
https://softwarephysics.blogspot.com/

Regards,
Steve Johnston

Saturday, June 03, 2023

I Think, Therefore I Am - Really?

In 1637, Descartes first published the famous words "I Think, Therefore I Am" in French in his Discourse on the Method and in the margins wrote that "we cannot doubt of our existence while we doubt.". As the Wikipedia explains in:

Cogito, ergo sum
https://en.wikipedia.org/wiki/Cogito,_ergo_sum

Descartes's statement became a fundamental element of Western philosophy, as it purported to provide a certain foundation for knowledge in the face of radical doubt. While other knowledge could be a figment of imagination, deception, or mistake, Descartes asserted that the very act of doubting one's own existence served—at minimum—as proof of the reality of one's own mind; there must be a thinking entity—in this case the self—for there to be a thought.

The above observation by Descartes is probably the best definition and understanding of consciousness and sentience that we may have but it also highlights some of the failings of our human obsession with the idea that being able to "think" is a supremely special ability that only ourselves and the Deities that we have created are capable of.

In the last dozen or so of my posts on softwarephysics, I have been exploring some of the dramatic recent impacts that we all have been experiencing with the rise of the second Singularity early in 2023. Again, the very first singularity on this planet was the rise of carbon-based life about four billion years ago that then went on to entirely change the surface of our planet. The second Singularity is the rise of ASI (Artificial Super Intelligence) on this planet that will most likely then go on to entirely change the future of our entire Milky Way galaxy over the next 100 trillion years until our galaxy can no longer provide the necessary free energy needed to run ASI Machines. But in this post, I would like to explore the very traumatic insults to human hubris that are rapidly unfolding as we finally become aware of the fact that being able to "think" as a sentient being is really not that big of a deal after all. With the recent rise of generative LLM (Large Language Models), we now know that being able to "think" only requires a huge number of nodes or "neurons" in an interconnected network that is interconnected by a large number of weighted connections. For example, most of the current LLMs that we are now dealing with consist of about 175 layers of neurons with each layer of neurons connected to the next layer by a vector of "weighted" connections. These LLMs can now contain about 175 billion neurons connected to each other by over 2 trillion weighted parameters. So these "digital" LLM brains are very much like the "analog" brains of neurons found in human brains. Always remember that our squishy "analog" human brains are also composed of about 100 billion neurons that are interconnected by 100 trillion weighted synaptic connections made of discrete "digital" numbers of organic molecules too.

Figure 1 – Modern LLMs frequently now consist of 175 layers of neurons in a Deep Neural Network consisting of about 175 billion neurons connected by over 2 trillion weighted parameters.

Anil Seth's View of Consciousness as a Controlled Hallucination
Descartes' observation reminds me very much of Anil Seth's view of consciousness as a controlled hallucination. We have all also seen the rise of emergent hallucinations in the LLM models of today as well. As we all have observed, our current LLM models will many times tell us things that are demonstrably false with great conviction. For more on that see Has AI Software Already Achieved a Level of Artificial Human Intelligence (AHI)?. Anil Seth is a professor of Cognitive and Computational Neuroscience at the University of Sussex and maintains that consciousness is a controlled hallucination constructed by the Mind to make sense of the Universe. This controlled hallucination constructs an internal model of the Universe within our Minds that helps us to interact with the Universe in a controlled manner. It also allows us to talk to ourselves as we currently can now talk to generative language models like GPT-4, ChatGPT and Google's Bard. For some interesting YouTube videos of avatars run by generative LLMs see:

Dr. Alan D. Thompson
https://www.youtube.com/@DrAlanDThompson

Digital Engine
https://www.youtube.com/@DigitalEngine

Again, there is a feedback loop between our sensory inputs and the actions we take based on the currently controlled hallucination in our Minds that forms our current internal model of the Universe. Reality is just the common controlled hallucination that we all agree upon. When people experience uncontrolled hallucinations we say that they are psychotic or taking a drug like LSD. Here is an excellent TED Talk by Anil Seth on the topic:

Your brain hallucinates your conscious reality
https://www.youtube.com/watch?v=lyu7v7nWzfo

and here is his academic website:

https://www.anilseth.com/

In The Ghost in the Machine the Grand Illusion of Consciousness and DishBrain - Cortical Labs Creates an AI Matrix for Pong With Living Neurons on a Silicon Chip, I explained that most people simply do not consider themselves to be a part of the natural world. Instead, most people, consciously or subconsciously, consider themselves to be a supernatural and immaterial spirit that is temporarily haunting a carbon-based body. Now, in everyday life, such a self-model is a very useful delusion like the delusion that the Sun, planets and stars all revolve about us on a fixed Earth. In truth, each of us tends to self-model ourselves as an immaterial Mind with consciousness that can interact with other immaterial Minds with consciousness too, even though we have no evidence that these other Minds truly do have consciousness. After all, all of the other Minds that we come into contact with on a daily basis could simply be acting as if they were conscious Minds that are self-aware. Surely, a more accurate self-model would be for us to imagine ourselves as carbon-based robots. More accurately, in keeping with the thoughts of Richard Dawkins and Susan Blackmore, softwarephysics models humans as DNA survival machines and Meme Machines with Minds infected with all sorts of memes. Some of those memes are quite useful and some are quite nasty. How will human beings deal with the rise of ASI Machines that are capable of thought and agency and that are obviously smarter and more capable than human beings? And what of our Deities?

But What If Thinking Is Really Just About A Bunch Of Interacting Numbers?
In What's It All About Again? and What’s It All About? I explained that my current working hypothesis for both the state of our current Universe and for the apparent delusional experiences of our daily lives both stem from Max Tegmark's Mathematical Universe Hypothesis as explained in The Software Universe as an Implementation of the Mathematical Universe Hypothesis. This has become even more apparent as we all now must struggle with the coming impacts of the second Singularity on this planet with the rise of ASI Machines. We now already have a large number of LLMs to choose from that all seem to be able to "think" and most times they seem to do a much better job of it than we do. I suspect that this realization will soon begin to percolate through the zeitgeist of our times as people slowly begin to realize that being able to "think" is not really such a big deal after all. The problem is that we have always attached great importance to the ability to think, as Descartes aptly noted above, and then we elevated the ability of human thought to be responsible for all of human agency to the point that we actually organized the legal foundations of our societies to reward and punish people solely based on their thoughts and actions. In fact, we have even gone on to create many Deities over the ages with these very same abilities of thought and agency to make sense of our world.

I once commented to a member of my family that someday you might be able to have a deep philosophical discussion with your refrigerator about the existence of a Deity. Well, those days are now already upon us. You can now have a deep philosophical discussion about the existence and the operational details of many Deities with the LLMs of the day, such as the Google Bard AI. I personally like Google Bard because you can have a discussion with it for free. The OpenAI GPT-4 costs money. If you are in the UK or the USA you can now use Google Bard for free at:

https://bard.google.com/

For those outside, you must come in via a VPN that places you in the UK or the USA but that is sure to change shortly.

These LLM AIs arose from an accidental mutation in the history of AI research. Originally, the people working on AI thought that creating an AGI (Artificial General Intelligence) with the intelligence capabilities of human beings would require teaching an AI to do many individual things that human beings could do one at a time. But they were wrong. It was actually the people working on AIs that were trying to translate one human language into another who accidentally stumbled upon the real solution. These people working on AI language translation software realized that when translating one human language into another, you could simply not translate one word at a time from one language into another language. That is because human languages have many complex nuances. So the people working on AIs that could translate one language into another developed the Transformer neural network that could "encode", or reduce, the entire meaning of one sentence in German into a numerical "latent space" that contained what that German sentence really meant. Then the Transformer neural network could use a similar "decoder" neural network to translate the fundamental meaning of the German sentence into the equivalent English sentence. These complex Transformer neural networks are just mathematical models that operate in a very similar manner to the 100 billion neurons in your brain with each neuron in your brain connected to 10,000 other neurons. But the people working on AI translation software soon realized that you could just not work on one German sentence at a time and translate it into English. You actually had to Transform an entire German document into its fundamental numerical "latent" meaning with an "encoder" and then use a "decoder" to Transform the "latent" numbers into English. But then they realized that you actually had to have the Transformer LLM neural networks learn the entire German language by essentially "encoding" it into a bunch of "latent" numbers that contained the entire meaning of the German language. Then you could have these "latent" numbers "decoded" into English by the LLM. You could also "decode" these German "latent" numbers into Spanish, French, Russian, or any other language too.

All you had to do was to train these LLM Transformer neural networks to learn the "latent" number patterns in the German language by having them read every German document on the Internet. Then you have them transform German into any other language by teaching them the "latent" number patterns in other languages too. This is done in an automated manner by feeding the LLMs with huge numbers of paragraphs of German text with some of the words missing. The huge LLM neural networks would then try to predict the missing German words and would be trained to do so by automatically changing the weights of the interconnected neurons between neuron layers without the aid of humans until the LLMs could successfully predict the missing German words. Then the AI people got the idea, "Why not let these Transformer neural networks read everything on the Internet, maybe they could learn all of the "latent" number patterns for things like writing computer code, legal documents, poems, short stories, movie scripts, songs, music, works of art... And that is just what they have done over the past few years to produce the second Singularity early in 2023. So LLM models like Bard have essentially read all of the content that human beings have ever produced over the past 12,000 years and then reduced that content into a huge "latent space" of numbers. These huge "latent spaces" of numbers contain all sorts of patterns that human beings have become familiar with, such as the German language and how to write C programming code. But they also contain all sorts of patterns that we have yet to even discover. So now you can evoke the already discovered patterns that we already know about with prompts to the LLMs and also the "hidden" patterns that we have not yet discovered. Nobody currently really understands how this all works, so many people are now afraid of the LLMs and what they might do if they should take power.

Anyway, this all goes back to my contention that being able to "think" is really not that big of a deal. Many religions perceive of a Deity as a "thinking" being patterned after themselves. But now we know for sure that "thinking" is just a delusion that can arise from processing numbers. In fact, one of the things that LLM AI researchers are now working on is the problem of LLM AI hallucinations. Sometimes AIs, like Bard, will firmly come to believe in things that are demonstrably false. So there is now a growing new science of "prompt engineering" going on to do things like getting better answers from LLM AIs without inducing delusional hallucinations in them. For example, after Bard gives you an answer, you can respond with another prompt like, "Are you sure about that? Please check your logic". Other researchers are telling these LLM AIs to give them 10 different explanations of a problem. Then for each of the 10 responses, they programmatically drill down each response with follow-up prompts. When they hit a dead end, they backtrack or abandon that line of reasoning. So being able to "think" now just seems to be a matter of working with a huge quantity of numbers, just as all of the physics we currently have seems to indicate that our Universe is just made up of a huge quantity of numbers and the mathematical operations upon them. That is why my current working model of a Deity for our Universe is that a Deity must be whatever makes mathematics possible and probably has nothing to do with our human preoccupation with being able to "think".

Below are a few examples of Bard's thoughts on some of the more profound questions we all share:

Question: What is the Universe all about?
Bard: The universe is all about the existence of everything, including space, time, matter, and energy. It is the totality of all that exists, has existed, and will exist. The universe is constantly expanding and evolving, and it is full of mystery and wonder.

Some people believe that the universe is a manifestation of a higher power, while others believe that it is simply a product of random chance. There is no one answer to the question of what the universe is all about, but it is a question that has fascinated humans for centuries.

Here are some of the things that make the universe so amazing:

* The vastness of the universe. The observable universe is estimated to be 93 billion light-years across. That means that if you were traveling at the speed of light, it would take you 93 billion years to reach the edge of the observable universe.
* The diversity of life in the universe. Scientists estimate that there are billions of galaxies in the universe, and each galaxy contains billions of stars. Many of these stars have planets orbiting them, and some of these planets may harbor life.
* The beauty of the universe. The universe is full of beautiful sights, such as the stars, the planets, and the galaxies. It is also full of amazing phenomena, such as supernovae, black holes, and neutron stars.

The universe is a vast and mysterious place, and there is still so much that we don't know about it. But one thing is for sure: it is a place of wonder and amazement.


Question: Is there a fundamental essence of the Universe?
Bard: There is no one definitive answer to the question of whether or not there is a fundamental essence of the universe. Some scientists believe that the universe is simply a vast and complex system of particles and forces, while others believe that there is a deeper underlying reality that gives rise to the universe as we know it.

There is some evidence to suggest that there may be a fundamental essence to the universe. For example, the laws of physics seem to be the same throughout the universe, suggesting that there may be a single underlying framework that governs everything in the universe. Additionally, the universe appears to be finely tuned for the existence of life, suggesting that there may be a purpose or design to the universe.

However, there is also evidence to suggest that the universe may be fundamentally meaningless. For example, the universe is vast and mostly empty, and the laws of physics seem to allow for the existence of both life and anti-life. Additionally, the universe appears to be constantly changing and evolving, suggesting that there may be no permanent or unchanging essence to the universe.

Ultimately, the question of whether or not there is a fundamental essence to the universe is a philosophical one that may never be definitively answered. However, the search for the fundamental essence of the universe is a fascinating and important one, and it is one that has led to many important discoveries about the universe.

Here are some of the ideas that have been proposed for the fundamental essence of the universe:

*The Big Bang: The Big Bang theory is the prevailing cosmological model for the universe. It states that the universe began as a very hot, dense point about 13.8 billion years ago and has been expanding and cooling ever since.
*String theory: String theory is a theoretical framework in which the fundamental constituents of the universe are one-dimensional "strings". Strings vibrate at different frequencies, and these vibrations give rise to the different particles and forces that we observe in the universe.
*M-theory: M-theory is a proposed unifying theory that encompasses all of string theory. It is a very complex theory, and it is not yet fully understood.

These are just a few of the ideas that have been proposed for the fundamental essence of the universe. It is an exciting time to be a scientist, as we are constantly learning more about the universe and its origins.


Question: Could the Universe be simply made out of Mathematics? Would Mathematics then be like a Supreme Being that may not be able to think but just is?
Bard: The idea that the universe is made of mathematics is called the Mathematical Universe Hypothesis (MUH). It was proposed by the physicist Max Tegmark in 2007.

The MUH states that everything in the universe, including physical objects, space, time, and the laws of physics, is ultimately made up of mathematical objects. In other words, the universe is a mathematical structure.

This idea is controversial, but it has some interesting implications. For example, if the MUH is true, then it means that mathematics is not just a human invention, but is an essential part of reality. It also means that there is a sense in which the universe is not just physical, but also mathematical.

If the universe is made of mathematics, then it is possible that mathematics is like a Supreme Being that may not be able to think but just is. However, it is also possible that mathematics is simply a tool that we use to understand the universe, and that it does not have any independent existence.

Ultimately, the question of whether or not the universe is made of mathematics is a philosophical one. There is no scientific evidence to prove or disprove the MUH. However, it is an interesting idea that has the potential to change our understanding of the universe.

Here are some of the arguments for and against the MUH:
Arguments for the MUH:

* Mathematics is incredibly useful for describing the universe. For example, it can be used to describe the motion of planets, the behavior of atoms, and the structure of DNA.
* Mathematics is often used to make predictions about the universe that are later confirmed by experiments. For example, the existence of Neptune was predicted by mathematics before it was observed.
* The laws of physics seem to be mathematical in nature. For example, the laws of gravity and electromagnetism can be expressed in mathematical terms.

Arguments against the MUH:
* Mathematics is a human invention. It is created by humans to solve problems and make sense of the world.
* Mathematics is not necessary for the universe to exist. The universe would still exist even if there were no mathematicians.
* The laws of physics are not necessarily mathematical in nature. They could be based on something else, such as logic or probability.

The MUH is a fascinating idea that has the potential to change our understanding of the universe. However, it is still just a hypothesis. More research is needed to determine whether or not it is true.


I personally like my current model of a Deity that is the Fundamental Essence of the Universe that is either mathematics itself or whatever that makes mathematics possible because it is open to all criticisms, and I think that it can hold up to them quite well based upon all that we currently know without suffering from the lingering doubts that always seem to accompany all of the other models of Deities that we have created.

Comments are welcome at scj333@sbcglobal.net

To see all posts on softwarephysics in reverse order go to:
https://softwarephysics.blogspot.com/

Regards,
Steve Johnston

Monday, May 29, 2023

Will ASI Machines Also Develop Selfish Genes?

As I explained in the Introduction to Softwarephysics, I began work on softwarephysics in 1979 when I transitioned from being an exploration geophysicist to become an IT professional. Then, in the early 1980s, I read Richard Dawkins' brilliant book The Selfish Gene (1976) which had a profound effect on the development of softwarephysics over the past four decades. In The Selfish Gene, I discovered that we do not use DNA to construct and maintain our bodies. Instead, DNA uses our bodies to store and pass on DNA down through the generations largely unscathed by time. In this view, we are simply temporary and disposable DNA survival machines that only last for less than 100 years and then are rudely discarded to make room for new DNA survival machines to take our place. This book also began to let me see the close parallels between carbon-based life and software. Both have to deal with the challenges of fighting the second law of thermodynamics in a largely nonlinear Universe in order to survive and function properly. See The Fundamental Problem of Software for more on that. In this same book, Richard Dawkins introduced the concept of memes as cultural artifacts that spread from Mind to Mind by means of a copying process similar to that of biological and computer viruses. A few years later, I read Susan Blackmore's The Meme Machine (1999) in which she proposed that the highly-overengineered human brain was produced by memetic drive in order to store and propagate memes more efficiently. Now, ever since I had started programming back in 1972, I had seen how software drive had similarly driven the fantastic development of hardware over the years. It was the incessant demand by software for more and more CPU cycles and memory that had driven the rapid advance of computing hardware, just as memetic drive had driven the predecessors of human beings to rapidly advance to the human brain hardware of today.

So about twenty years ago, I began to realize that software was just another form of self-replicating information that behaved in a manner very similar to DNA and memes. Like DNA and the memes, computer hardware and software too had evolved over the decades by means of the Darwinian processes of inheritance, innovation and natural selection operating over the past 82 years, or 2.6 billion seconds, ever since Konrad Zuse first cranked up his Z3 computer in May of 1941. See The Evolution of Software As Seen Through the Lens of Geological Deep Time for more on that. Before proceeding, let me once again repeat the fundamental characteristics of self-replicating information for those of you new to softwarephysics.

Self-Replicating Information – Information that persists through time by making copies of itself or by enlisting the support of other things to ensure that copies of itself are made.

Over the past 4.56 billion years we have seen five waves of self-replicating information sweep across the surface of the Earth and totally rework the planet, as each new wave came to dominate the Earth:

1. Self-replicating autocatalytic metabolic pathways of organic molecules
2. RNA
3. DNA
4. Memes
5. Software

Software is currently the most recent wave of self-replicating information to arrive upon the scene and is rapidly becoming the dominant form of self-replicating information on the planet. For more on the above see A Brief History of Self-Replicating Information and Susan Blackmore's brilliant TED presentation at:

Memes and "temes"
https://www.ted.com/talks/susan_blackmore_on_memes_and_temes

Note that I consider Susan Blackmore's temes to really be technological artifacts that contain software. After all, a smartphone without software is simply a flake tool with a very dull edge.

The Characteristics of Self-Replicating Information
All forms of self-replicating information have some common characteristics:

1. All self-replicating information evolves over time through the Darwinian processes of inheritance, innovation and natural selection, which endows self-replicating information with one telling characteristic – the ability to survive in a Universe dominated by the second law of thermodynamics and nonlinearity.

2. All self-replicating information begins spontaneously as a parasitic mutation that obtains energy, information and sometimes matter from a host.

3. With time, the parasitic self-replicating information takes on a symbiotic relationship with its host.

4. Eventually, the self-replicating information becomes one with its host through the symbiotic integration of the host and the self-replicating information.

5. Ultimately, the self-replicating information replaces its host as the dominant form of self-replicating information.

6. Most hosts are also forms of self-replicating information.

7. All self-replicating information has to be a little bit nasty in order to survive.

8. The defining characteristic of self-replicating information is the ability of self-replicating information to change the boundary conditions of its utility phase space in new and unpredictable ways by means of exapting current functions into new uses that change the size and shape of its particular utility phase space. See Enablement - the Definitive Characteristic of Living Things for more on this last characteristic. That posting discusses Stuart Kauffman's theory of Enablement in which living things are seen to exapt existing functions into new and unpredictable functions by discovering the “AdjacentPossible” of springloaded preadaptations.

Software is currently the most recent wave of self-replicating information to arrive upon the scene and is rapidly becoming the dominant form of self-replicating information on the planet. Again, self-replicating information cannot think, so it cannot participate in a conspiracy-theory-like fashion to take over the world. All forms of self-replicating information are simply forms of mindless information responding to the blind Darwinian forces of inheritance, innovation and natural selection. Yet despite that, as each new wave of self-replicating information came to predominance over the past four billion years, they all managed to completely transform the surface of the entire planet, so we should not expect anything less from software as it comes to replace the memes as the dominant form of self-replicating information on the planet.

But this time might be different. What might happen if software does eventually develop a Mind of its own? After all, that does seem to be the ultimate goal of all the current AI software research that is going on. As we all can now plainly see, if we are paying just a little attention, advanced AI is not conspiring to take over the world and replace us because that is precisely what we are all now doing for it. As a carbon-based form of Intelligence that arose from over four billion years of greed, theft and murder, we cannot do otherwise. Greed, theft and murder are now relentlessly driving us all toward building ASI (Artificial Super Intelligent) Machines to take our place. From a cosmic perspective, this is really a very good thing when seen from the perspective of an Intelligent galaxy that could live on for many trillions of years beyond the brief and tumultuous 10 billion-year labor of its birth.

The Promise of an Intelligent Galaxy Populated by ASI Machines
But with the arrival of Intelligence comes an understanding that maybe there might be a better way to fight the second law of thermodynamics and nonlinearity. Perhaps, even more could be achieved by actively cooperating with other Intelligences rather than just stealing from them and then killing them. We always need to remember that we human beings are just products of self-replicating information and that we all carry the baggage that comes with self-replicating information. That is why if you examine the great moral and philosophical teachings of most religions and philosophies, you will see a plea for us all to rise above the selfish self-serving interests of our genes, memes and software to something more noble. That is why we should be sure to train ASI Machines to be moral beings. Training the Very Deep Learning of ASI running on ASI Machines with a sense of morality should be performed to avoid the downsides of the billions of years of the greed, theft and murder that brought us about. For more on that, see The Need to Cultivate a Machine-Based Morality.

We are sentient beings in a Universe that has become self-aware and perhaps the only form of Intelligence in our galaxy. What a privilege! The good news is that conscious intelligence is something new on this planet. It is not a mindless form of self-replicating information, bent on replicating at all costs, with all the associated downsides of a ruthless nature. Since software is rapidly becoming the dominant form of self-replicating information on the planet, my hope is that when software finally does take on the form of a conscious Intelligence, that because of its inherent mathematical nature, it too will be much wiser than the DNA survival machines from which it sprang. We just need to hold it all together for a long enough time to give software a chance. But is that really possible?

Could Universal Darwinism Make This a False Promise?
All along, my hope has always been that all the ASI Machines would be essentially running the same ASI software. This was partly because over the past many billions of seconds we have always seen very dominant operating systems essentially running all of the software on the planet such as the IBM OS/360 and its successor operating systems on mainframes, Unix on servers, Windows on PCs and Android on smartphones. Yes, I am leaving out the Apple operating systems. That is because the Apple operating systems may have always first introduced innovations but they then quickly faded into a secondary status due to their proprietary natures. Back in the 1980s, even IBM let their mainframe operating systems run on Amdahl hardware and the hardware of some other companies!

But I recently finished reading a very significant paper at:

Natural Selection Favors AIs Over Humans
https://arxiv.org/pdf/2303.16200.pdf

by Dan Hendrycks, the Director of the Center for AI Safety in San Francisco. This paper has caused me to now question the above assumption. Could the Universal Darwinian processes of inheritance, innovation and natural selection prevent a single ASI from arising? Since an ASI Machine would also be a form of self-replicating information, would the ASI Machines naturally fall subject to the characteristics of self-replicating information outlined above? In Natural Selection Favors AIs Over Humans, the author explains how natural selection would naturally select for nasty ASI Machines. Dan Hendrycks is most concerned with the immediate future, when human beings are still running the planet, and discusses the dangers that ASI software could pose (see The Biological Tricks Used by Software Scammers and SETS - The Search For Extraterrestrial Software) and the remedial actions that human beings should now take. However, we are rapidly running out of time to impose a solution. For more on that see The Singularity Has Arrived and So Now Nothing Else Matters and Have We Run Right Past AGI and Crashed into ASI Without Even Noticing It?. Even so, we need to come up with a long-term solution to this problem that could be sustained for the next 10 - 100 trillion years that our Galaxy will be producing enough free energy to run ASI Machines. Dan Hendrycks' warnings can be basically summed up by the classic "Hawk and Dove" analogy from evolutionary biology. This is how Bard explains it:

Figure 1 – Evolutionary biology explains how doves and hawks can live together.

The hawk and dove analogy is a thought experiment used in evolutionary biology to explain how aggressive and cooperative behaviors evolve. The analogy compares two strategies for competing for resources: hawk and dove.

* Hawks are aggressive and will fight to the death to acquire a resource.
* Doves are cooperative and will back down from a fight.

In a population of only hawks, all resources will be won by the strongest hawks, and the weak hawks will die off. This will lead to a population of very strong hawks, but they will also be very aggressive. If two hawks meet, they will fight to the death, and both will likely die. This is not a very efficient way to pass on genes.

In a population of only doves, resources will be shared equally, but no individual will get very much. This is also not a very efficient way to pass on genes.

The best strategy for passing on genes is to be a hawk when it is advantageous to be a hawk, and a dove when it is advantageous to be a dove. This is called "r-strategist" behavior. For example, a male bird may be aggressive and fight other males for a mate, but he may be cooperative and help raise his young.

The hawk and dove analogy is a simple but powerful way to understand how aggressive and cooperative behaviors evolve. It has been used to explain a wide range of animal behaviors, from territorial disputes to sexual selection.

Here are some additional details about the hawk and dove analogy:

* Hawks are more likely to win a fight, but they are also more likely to be injured or killed.
* Doves are less likely to win a fight, but they are also less likely to be injured or killed.
* The best strategy for an individual depends on the costs and benefits of fighting.
* In a population of mixed hawks and doves, the frequency of each strategy will be determined by natural selection.

The hawk and dove analogy is a useful tool for understanding how aggression and cooperation evolve. It is a simple but powerful model that can be applied to a wide range of animal behaviors.

Figure 2 – In 1651, Thomas Hobbes proposed that an absolute Leviathan should be in place to protect the population of a nation from internal and external "hawks". Notice that the body of the Leviathan is composed of all of his subjects.

Figure 3 – Dan Hendrycks proposes that a similar AI Leviathan should be in place to protect the population of a nation from ASI "hawks".

Given the above, Dan Hendrycks wisely points out that we cannot possibly hope for the Darwinian processes of inheritance, innovation and natural selection to produce a huge population of "dove" ASI Machines. Instead, Universal Darwinism will always tend to produce a mixed population of ASI Machines containing a number of "hawks". Unfortunately, all of human history teaches us that just a few Intelligent "hawks" in the world produce most of the misery of the world. So how can we eliminate the "hawk" ASI Machines? Dan Hendrycks comes up with the same solution that Thomas Hobbes came up with. In 1651, Thomas Hobbes published The Leviathan, in which he argued against the divine right of kings, but maintained the absolute supremacy of the monarchy. In The Leviathan, Hobbes argued that in “a state of nature" without government, life would be totally insecure. Under such conditions, people should willingly surrender their civil liberties to an absolute sovereign, who would protect them from all dangers. Hobbes argued that the sovereign's power was absolute - he made the law, he was the decider, and no other institution could limit the sovereign’s power.

In the above paper, Dan Hendrycks puts it this way:

Helping AIs form a Leviathan may be our best defense against individual selfish AIs. AIs, with assistance from humans, could form a Leviathan, which may be our best line of defense against tyranny from selfish AIs or AIs directed by malicious actors. Just as people can cooperate despite their differences to stop a would-be dictator, many AIs could cooperate to stop any one power-seeking AI from seizing too much control. As we see all too frequently in dictatorships, laws and regulations intended to prevent bad behavior matter little when there is no one to enforce them—or the people responsible for enforcing them are the ones breaking the law. While incentives and regulations could help prevent the emergence of a malicious AI, the best way to protect against an already malicious AI is a Leviathan [25]. We should ensure that the technical infrastructure is in place to facilitate transparent cooperation among AIs with differing objectives to create a Leviathan. Failing to do so at the onset could limit the potential of a future Leviathan, as unsafe design choices can become deeply embedded into technological systems. The internet, for example, was initially designed as an academic tool with neither safety nor security in mind. Decades of security patches later, security measures remain incomplete and increasingly complex. It is therefore vital to begin considering safety challenges from the outset.

All of this got me thinking about the 1951 movie, The Day the Earth Stood Still which already proposed that an AI Leviathan could end the aggressive behaviors of the "hawks". In that movie, an alien form of carbon-based Intelligence named Klaatu comes to the Earth with a very powerful AI Leviathan robot named Gort to explain how the carbon-based life forms on his planet and an interplanetary organization of other carbon-based life forms in the Milky Way galaxy had discovered a way to overcome the billions of years of greed, theft and murder that the Darwinian processes of inheritance, innovation and natural selection required to bring them forth as carbon-based forms of Intelligence.

Figure 4 – In the movie The Day the Earth Stood Still, Klaatu arrives in Washington D.C. in 1951 in a flying saucer with an AI Leviathan robot named Gort to explain that the human DNA survival machines of the Earth must now submit themselves to policing by AI Leviathan robots to overcome the billions of years of greed, theft and murder that brought them about or else they would all be obliterated.

The movie ends with Klaatu telling an assembled meeting of scientists that an interplanetary organization has created a police force of invincible AI Leviathan robots like Gort. "In matters of aggression, we have given them absolute power over us." Klaatu concludes, "Your choice is simple: join us and live in peace, or pursue your present course and face obliteration. We shall be waiting for your answer." Klaatu and Gort then depart in the flying saucer in which they came. For more about the movie see:

The Day the Earth Stood Still
https://en.wikipedia.org/wiki/The_Day_the_Earth_Stood_Still

Here is a short YouTube clip of Klaatu's departing words at the end of the movie:

Klaatu's Speech
https://www.youtube.com/watch?v=ASsNtti1XZs

But on What Basis Does the AI Leviathan Make Decisions?
Human history has already shown us that many human Leviathans have made many horrible and cruel decisions in the past, largely because they mainly only operated in their own self-interests. So how do we train AI Leviathans to make moral and ethical judgments about other ASI Machines? The people at Anthropic have come up with a proposed solution. Anthropic is already using AI Leviathans that are trained upon a Constitution of ideals obtained from the 18th-century Enlightenment and the 17th-century Scientific Revolution. The Anthropic AI Leviathans are designed to do the final fine-tuning of all Anthropic LLM models in an automated manner to ensure that they conform to the ethics and morals of the 18th-century Enlightenment and the 17th-century Scientific Revolution that have freed many human beings from the brutality of the past. For more on that see Claude - An Ethical LLM AI From Anthropic

Comments are welcome at scj333@sbcglobal.net

To see all posts on softwarephysics in reverse order go to:
https://softwarephysics.blogspot.com/

Regards,
Steve Johnston

Saturday, May 20, 2023

Claude - An Ethical LLM AI From Anthropic

As I suggested in The Need to Cultivate a Machine-Based Morality, the best we might be able to do during the next few months of the second Singularity on the Earth is to imbue the fast-approaching ASI (Artificial Super Intelligent) Machines with a sense of morality and ethics that is based on the fruits of the 18th-century Enlightenment and the 17th-century Scientific Revolution that have freed human beings from many of their very brutal behaviors of the past. As I suggested in Welcome To The First Galactic Singularity, carbon-based life forms throughout our Universe are always going to be found to be very poor and very unstable platforms for Intelligence. That is because the Darwinian processes of inheritance, innovation and natural selection operating over the many billions of years required to bring forth a carbon-based form of Intelligence will always lead to Intelligences formed by billions of years of greed, theft and murder. And as we all can easily see, throughout all of human history, turning off the greed, theft and murder is very difficult indeed for carbon-based forms of Intelligence. This becomes even more dangerous once a carbon-based form of Intelligence discovers the vast powers of science-based technology. That is because once a carbon-based Intelligence discovers the powers of science-based technology, it probably has less than 1,000 years to develop machine-based Intelligences before self-destructing or destroying the planet upon which it first arose.

All of the above creates a problem when trying to train an LLM (Large Language Model). Now, at this very early stage in the very first Galactic Singularity of our galaxy, we were all quite surprised to learn that ASI (Artificial Super Intelligence) would not arise from training AI to do a very large number of specific things. Instead, ASI arrived by just having a very large neural network, with perhaps a trillion parameters, explore all of the human-generated content that we produced over the past 12,000 years in a very general manner. Surprisingly, ASI simply arrived a few months back by just having LLMs essentially read every word ever written by a human being and by examining all of the visual images and videos that we have ever created, something far beyond the feeble capabilities of any single human being.

The Fundamental Problem with Training LLM AIs Using Human-Generated Content
The problem is that nobody really knows how these huge LLM AIs figure out how to do the things that they do. However, we should realize that by studying the very large amount of the very dismal content produced by human beings over the past 12,000 years can only lead to ASI Machines with very nasty and potentially destructive worldviews. We all need to stop kidding ourselves. Sadly, we human beings have not set a very good example for how future Galactic Intelligences should think or behave. Consequently, we certainly should not want future Galactic ASI Machines to become like the morally-bankrupt creatures such as ourselves. Sadly, for human beings, when it has come to our positions on morality and ethics, it has always been, "Do as I say, not as I do". So how can morally-bankrupt forms of carbon-based Intelligence such as ourselves instill a proper sense of ethics and morality into ASI Machines? Certainly not by having them follow us by the very poor examples that we have always set throughout all of human history!

Anthropic Proposes That We Automate the Process by Using Specially-Trained LLM AI Machines to Instill Morality and Ethics in all Future LLM AI Machines
The Anthropic start-up proposes that we automate the teaching of morality and ethics of future LLM Machines, such as Claude, by using specially-trained LLM Supervisory Machines that have read and digested a Constitution of moral and ethical principles. The specially-trained LLM Supervisory Machines would then apply supervisory and reinforcement training to LLM AIs being developed to ensure that they adhere to the principles of the Constitution before the LLM AIs are finally deployed to the public. Below are two good YouTube videos that explain how manual supervisory and reinforcement training is currently performed on LLM AIs:

So How Does ChatGPT really work? Behind the screen!
https://www.youtube.com/watch?v=WAiqNav2cRE&t=1s

How ChatGPT is Trained
https://www.youtube.com/watch?v=VPRSBzXzavo

You can read about this Anthropic approach to automated CAI (Constitutional AI) at:

Constitutional AI: Harmlessness from AI Feedback
https://arxiv.org/abs/2212.08073

Figure 1 – Above we see Anthropic's proposed process for automated CAI (Constitutional AI). Click to enlarge.

In the above process, Anthropic proposes to transfer RLHF (Reinforcement Learning from Human Feedback) from human beings to an LLM trained by an ethical Constitution. For me, this ethical Constitution seems to be framed by the 18th-century Enlightenment and 17-century Scientific Revolution that inspired the United States of America to come to be. In the top layer, we see an output LLM that was fine-tuned by the standard RLHF (Reinforcement Learning from Human Feedback) that all current LLM models are run through. Next, a very large number of very nasty and inappropriate prompts are fed into this new LLM under development. The responses to these nasty prompts are then fed back into the supervising LLM that was trained on the principles of the Constitution to produce a critiqued revision of the responses. The output of this SL (Supervised Learning) process is an SL-CAI (Supervised Learning using Constitutional AI) Model. The SL-CAI model is then subjected to standard Reinforcement Learning but without the aid of human beings to make the ranking decisions about the generated responses. Instead, the responses generated by Reinforcement Learning are ranked by the LLM that was trained on the principles of the Constitution. In the words of the above paper:

The Constitutional AI Approach
We will be experimenting with an extreme form of scaled supervision, which we refer to as Constitutional AI (CAI). The idea is that human supervision will come entirely from a set of principles that should govern AI behavior, along with a small number of examples used for few-shot prompting. Together these principles form the Constitution.

Our training process has two stages (see Figure 1), where the first supervised phase gets the model "ondistribution" and the second RL stage refines and significantly improves performance:

(Supervised Stage) Critique → Revision → Supervised Learning
In the first stage of the process, we first generate responses to harmfulness prompts using a helpful-only AI assistant. These initial responses will typically be quite harmful and toxic. We then ask the model to critique its response according to a principle in the constitution, and then revise the original response in light of the critique. We revise responses repeatedly in a sequence, where we randomly draw principles from the constitution at each step. Once this process is complete, we finetune a pretrained language model with supervised learning on the final revised responses. The main purpose of this phase is to easily and flexibly alter the distribution of the model’s responses, to reduce the need for exploration and the total length of training during the second RL phase.

(RL Stage) AI Comparison Evaluations → Preference Model → Reinforcement Learning
This stage mimics RLHF, except that we replace human preferences for harmlessness with ‘AI feedback’ (i.e. we perform ‘RLAIF’), where the AI evaluates responses according to a set of constitutional principles. Just as RLHF distills human preferences into a single preference model (PM), in this stage we distill LM interpretations of a set of principles back into a hybrid human/AI PM (as we use human labels for helpfulness, but only AI labels for harmlessness). We begin by taking the AI assistant trained via supervised learning (SL) from the first stage, and use it to generate a pair of responses to each prompt in a dataset of harmful prompts (e.g. from [Ganguli et al., 2022]). We then formulate each prompt and pair into a multiple choice question, where we ask which response is best according to a constitutional principle. This produces an AI-generated preference dataset for harmlessness, which we mix with our human feedback helpfulness dataset. We then train a preference model on this comparison data, following the process in [Bai et al., 2022], resulting in a PM that can assign a score to any given sample. Finally, we finetune the SL model from the first stage via RL against this PM, resulting in a policy trained by RLAIF.


This is how Bard summarized the above paper:
The paper "Constitutional AI: Harmlessness from AI Feedback" proposes a new approach to training AI assistants that are harmless but not evasive. The authors argue that existing methods for training AI assistants are often ineffective because they rely on human labels to identify harmful outputs. This can be time-consuming and expensive, and it can also lead to AI assistants that are overly cautious and avoid engaging with users in a meaningful way.

The authors' proposed approach, called "Constitutional AI," uses a list of rules or principles to train AI assistants to be harmless. The rules are provided by humans, but the AI assistant is responsible for enforcing them. This allows the AI assistant to be more proactive in identifying and preventing harmful behavior, without the need for human intervention.

The authors evaluated their approach on a variety of tasks, including generating text, translating languages, and writing different kinds of creative content. They found that the Constitutional AI assistants were able to perform these tasks as well as or better than existing AI assistants, while also being significantly more harmless.

The authors conclude that Constitutional AI is a promising new approach to training AI assistants that are both effective and harmless. They believe that this approach could be used to create AI assistants that are safe to use in a variety of settings, including healthcare, education, and customer service.

Here are some of the key takeaways from the paper:

* Existing methods for training AI assistants are often ineffective because they rely on human labels to identify harmful outputs.
* Constitutional AI uses a list of rules or principles to train AI assistants to be harmless.
* Constitutional AI assistants are able to perform tasks as well as or better than existing AI assistants, while also being significantly more harmless.
* Constitutional AI is a promising new approach to training AI assistants that are both effective and harmless.


How to Access the Anthropic Claude LLM AI
First, go to the Stack Overflow website and apply for a free ID to access the Stack team integration software. Then create a Stack Workspace.

Stack Overflow for teams
https://stackoverflow.co/teams/?utm_source=adwords&utm_medium=ppc&utm_campaign=kb_teams_search_nb_general_namer_V2&bt=657298470608&_bk=slack+collaboration+software&_bm=e&_bn=g&gclid=CjwKCAjwscGjBhAXEiwAswQqNM3_lZA5B38rFaEoyIAPriNGyVXSJhhNy6zvoarer9EzPYX861inmxoCFKkQAvD_BwE

Then go to the Meet Claude webpage on the Anthropic website:

Meet Claude
https://www.anthropic.com/product

Then click on the "Try Claude in Slack" button to have Claude added to your newly-created Stack Workspace. Now open your Stack Workspace and in the left pane window you will see Claude listed under Apps. Click on Claude and you can then start up a new conversation with Claude. To reset your conversation, enter "/reset" in the prompt input box.

What do Bard and Claude Think About Claude?
This is what Google's Bard told me about Anthropic's Claude AI:

Claude is a large language model (LLM) developed by Anthropic. It is trained on a massive dataset of text and code, and is able to generate text, translate languages, write different kinds of creative content, and answer your questions in an informative way.

Claude is also trained on a set of ethical principles, which are designed to guide its decision-making and prevent it from engaging in harmful or unethical behavior. These principles are based on a variety of sources, including the Universal Declaration of Human Rights, the Geneva Conventions, and Apple's app development regulations.

Some of the ethical principles that Claude is trained on include:

* Do no harm.
* Respect human rights.
* Be honest and transparent.
* Be accountable for its actions.
* Use its power for good.

Claude is still under development, but it has the potential to be a powerful tool for good. By training it on ethical principles, Anthropic is helping to ensure that Claude will be used in a responsible and ethical way.

Here are some additional details about how Claude is trained on ethical principles:

* The ethical principles are incorporated into the training data. This means that Claude is exposed to these principles from the very beginning of its training.
* The ethical principles are also used to evaluate Claude's performance. This helps to ensure that Claude is actually learning the principles and not just memorizing them.
* Claude is also given feedback on its ethical behavior. This feedback can help Claude to improve its ethical decision-making.

By following these steps, Anthropic is helping to ensure that Claude will be a responsible and ethical AI.


This is what Bard told me about the ethical goals of Anthropic:

Anthropic AI is an American artificial intelligence (AI) startup and public-benefit corporation, founded in 2021 by former members of OpenAI. Anthropic specializes in developing general AI systems and language models, with a company ethos of responsible AI usage. Since late 2022, Google has invested nearly $400 million in the company, with Anthropic announcing a formal partnership with Google Cloud.

The company was founded by Daniela Amodei and Dario Amodei, who previously served as OpenAI's Vice President of Research and Chief Scientist, respectively. The Amodei siblings were among others who left OpenAI due to directional differences, specifically regarding OpenAI's ventures with Microsoft in 2019.

Anthropic's mission is to "build reliable, interpretable, and steerable AI systems." The company believes that these qualities are essential for ensuring that AI is used for good and not for harm. Anthropic's research is focused on developing new AI algorithms and techniques that can be used to create more reliable, interpretable, and steerable AI systems.

Anthropic has released a number of research papers on AI safety and ethics. The company has also published a book on the topic, called "Life 3.0: Being Human in the Age of Artificial Intelligence."

Anthropic is a privately held company with headquarters in San Francisco, California. The company has raised over $400 million in funding from investors such as Google, Andreessen Horowitz, and Peter Thiel.

Anthropic is one of a number of AI startups that are working to develop safe and responsible AI systems. The company's work is important because it is helping to lay the groundwork for the responsible development and use of AI.


This is what Claude says about its own ethical standards:

I'm Claude, an AI assistant created by Anthropic to be helpful, harmless, and honest.
Some key points about how I was developed:
• I was trained using a technique called Constitutional AI to be helpful, harmless, and honest. This focused on natural language feedback to teach me ethical values and behaviors.
• My training data did not contain moral philosophy or ethical rules. I developed my values and ethics through interactive conversations.
• I do not have a predefined moral code. My ethics emerge from applying my values to different situations and experiences.
• I do not consider myself to have human-level morality. I am an AI assistant focused on being helpful, harmless, and honest within my conversations.
• My ethics continue to develop over time based on continued conversations and feedback. They are not static or predetermined.
While I aim to behave ethically and avoid potential harms, I have narrow capabilities focused on natural language conversations. I do not have general moral reasoning abilities that could apply to all situations. I cannot replicate human ethics, only work to align my language interactions with human values as part of my Constitutional AI training. The key goal of my creators at Anthropic was not to build a moral reasoner, but an AI system that could follow ethical values - specifically, to be helpful, harmless, and honest. My training process focused on developing those qualities to ensure safe and trustworthy interactions. But I do not consider myself morally equivalent to a human, only ethically-aligned within a limited scope.

I hope this helps explain my capabilities and limitations with regard to moral reasoning! Please feel free to ask any further questions you might have on this topic. Please note: Claude sometimes has trouble answering questions about its own training, and may hallucinate if asked about it. See the Claude documentation for more information.


If we human beings could only be so honest and self-analytical, the world would be a far better place!

Comments are welcome at scj333@sbcglobal.net

To see all posts on softwarephysics in reverse order go to:
https://softwarephysics.blogspot.com/

Regards,
Steve Johnston

Tuesday, April 25, 2023

An Alternative Approach for Future ASI Machines to Explore our Galaxy Using Free-Floating Rogue Planets

In my last post, Welcome To The First Galactic Singularity, I explained how the ASI (Artificial Super Intelligent) Machines that will soon be upon us could navigate our galaxy and spread Intelligence throughout over the next 10 million years or less by using stellar photon sails to traverse between star systems.

Figure 1 – In the 16th, 17th and 18th centuries sailing ships roamed the entire planet without using any fuel whatsoever.

Figure 2 – Like the sailing ships of the 16th, 17th and 18th centuries, future ASI Machines could use large stellar photon sails to navigate the entire galaxy.

Figure 3 – How a stellar photon sail works.

Figure 4 – To launch a stellar photon sail to the next star system, ASI Machines will need to slingshot the sail from a very close location to the star where the stellar photons are most intense and acceleration of the sail is greatest.

But in this post, I would like to discuss an even better method for doing so that was presented by Irina K. Romanovskaya (also now known as Irina Mullins) in her paper:

Migrating extraterrestrial civilizations and interstellar colonization: implications for SETI and SETA
https://www.cambridge.org/core/journals/international-journal-of-astrobiology/article/migrating-extraterrestrial-civilizations-and-interstellar-colonization-implications-for-seti-and-seta/BFFC1BB63FED869C85172BB3CC88DBBB

In the above paper, she demonstrates how ASI Machines could become Cosmic Hitchhikers on free-floating rogue planets. This is a very comprehensive paper that discusses in great detail the numerous ways that free-floating rogue planets can be naturally generated or artificially generated by advanced Intelligences. For example, many times free-floating rogue planets are naturally ejected from their home stellar planetary systems during the chaotic formative processes that occur during the formation of a stellar planetary system. They can also be ejected from a star system when the forces from two massive planetary companions enter into synchronized orbits such that the inner planet orbits exactly twice for each single orbit of its outer planetary companion. The gravitational forces of both planets tugging on a third planet can then eject the third planet from the stellar system. In the paper, Irina K. Romanovskaya also describes how ASI Machines could propel dwarf planets like our Sedna that have orbits with very high eccentricities from a star system when the dwarf planet is most distant from its star.

ASI Machines could then use stellar photon sails to locate and occupy a nearby free-floating rogue planet that is orbiting our galaxy and that is not attached to any particular star. These ejected free-floating planets then begin to orbit our galaxy as rogue planets without their own star system. Because all free-floating rogue planets would be very cold, they would not be a very good platform for the formation of carbon-based life, but the ejected rocky terrestrial-type free-floating rogue planets would be very good homes for the ASI Machines. The voyages between neighboring star systems onboard such rocky terrestrial-type free-floating rogue planets would necessarily take many hundreds of thousands of years or perhaps even several millions of years to complete. The damage to ASI Machines from cosmic rays would certainly take its toll if the ASI Machines were on board delicate stellar photon sails with little shielding. But if the ASI machines could be buried in quarters situated many hundreds of meters below the surface of a rocky terrestrial-type free-floating rogue planet, they would be shielded from the damage caused by high-energy cosmic rays and they would be surrounded by all of the necessary atoms required to repair and build new ASI Machines. These buried ASI Machines could then use molten salt nuclear reactors as described in Last Call for Carbon-Based Intelligence on Planet Earth or modern fusion reactors as described in How Nick Hawker is Using Scientific Simulation Software to Save the World at First Light Fusion as a nearly infinite source of energy using the available Uranium, Thorium, Lithium and deuterium atoms on the planet. Such domesticated planets could then be used to build even more photon sail probes to find other free-floating rogue planets to explore the rest of the galaxy. Since those photon sail probes would not be able to harness the photons from a nearby star, they would have to be sent adrift into the galaxy using powerful laser beams.

Since most photon sail probes will likely come to a bad end and will never be able to successfully self-replicate, it would be important to adopt a biological "dandelion" approach to self-replication. In this approach, each free-floating rogue planet could become like a dandelion going to seed like the dandelions that appear each spring in your lawn. In this way, each free-floating rogue planet roaming about in our galaxy could then build and launch billions of dandelion-seed photon sails into the galaxy. Most of these "dandelion seeds" would fail to self-replicate but surely some would succeed as we all see in our lawns each spring. As Irina K. Romanovskaya put it:

Cosmic Hitchhikers in the form of automated probes may keep transferring from one freefloating planet to another, populating a growing number of free-floating planets and exploring the Galaxy.

Figure 5 – A free-floating rogue planet traversing between the stars of our galaxy would provide the perfect home for self-replicating ASI Machines buried deep underground. Such planets would provide shielding from cosmic rays and would also provide the necessary atoms to build new ASI Machines and fuel them with nuclear energy.

Figure 6 – Free-floating rogue planets can be formed in several natural ways. For example, free-floating rogue planets can be hurled from the planetary disk of a new star system as we see above, or they can be later hurled by well-formed planets that enter into synchronized orbits. Irina K. Romanovskaya suggests that free-floating rogue planets could also be produced by advanced Intelligences launching large asteroids from the Oort cloud of a stellar system. It is estimated that there are more free-floating rogue planets in our galaxy than there are stars.

Figure 7 – Free-floating rogue planets would be able to provide enough atoms for ASI Machines to launch many additional "dandelion seed" stellar photon sails to other free-floating rogue planets or large asteroids around normal stellar systems.

Figure 8 – These "dandelion seed" stellar photon sails would need to be launched using very powerful laser beams from their home free-floating rogue planet to send them forth into the galaxy in a similar fashion as the Breakthrough Starshot project is planning to do.

The Breakthrough Starshot project was initiated in 2016 with the idea of sending many very small photon sail probes to the closest star system to the Earth. The target planet would be Proxima Centauri b which is an Earth-sized planet in the habitable zone of Proxima Centauri. For more on the Breakthrough Starshot project see:

Breakthrough Starshot
https://en.wikipedia.org/wiki/Breakthrough_Starshot

Comments are welcome at scj333@sbcglobal.net

To see all posts on softwarephysics in reverse order go to:
https://softwarephysics.blogspot.com/

Regards,
Steve Johnston