What would it take to publicize CGOL construction arms, etc.

[pcallahan]

This is kind of meta but I think it's a reasonable discussion topic and it popped into my head reading recent results on construction arms. To put it in stark terms:

How is it that K. Eric Drexler created massive buzz in the 80s and 90s over his vaporware speculation on nanotechnology, while the CGOL community has received far less attention building actual working instances of universal constructors that can be run on very modest platforms?

I don't intend this as an attack on Drexler, whose ideas are interesting. I also understand the distinction between building universal constructors in the real universe as opposed to a simplified universe like CGOL. However, I still think the accomplishments of this community should have made a greater impression by now. These results do get publicized now and then. Also, there's Greene and Johnston's textbook (hooray!).

Maybe there's a shift in culture, where the lack of obvious commercial applications makes it less interesting. E.g., I now wonder if some of the truly mind-boggling advances in generative AI would even be in the news except that people are excited about monetizing them. Is that too cynical? E.g. I find SAT solvers mind-boggling as well, and those are far less well known even among people in tech. (I'm convinced there are many non-obvious commercial applications for SAT solvers that are yet to be discovered.)

I am sort of just venting, and it ties into the discussion of a few years back of how to popularize CGOL. We don't seem to have made much of an advance there despite significant effort (such as the textbook).

As is my habit of late, I give ChatGPT 4 first dibs on replying to the above comment:

Your question is a multifaceted one, but I'll try to answer it as best as I can.

For those who may not know, K. Eric Drexler is an American engineer best known for popularizing the potential of molecular nanotechnology, as described in his book, "Engines of Creation". He envisioned tiny machines that could manipulate matter at the atomic level, promising a revolution in manufacturing, medicine, and other fields. It's an exciting concept, but as of my knowledge cutoff in 2021, has yet to be fully realized.

CGOL, or Conway's Game of Life, is a simple mathematical game invented by British mathematician John Horton Conway in 1970. The game involves an infinite two-dimensional grid of cells that evolve across discrete time steps according to a set of rules. Although it may appear simple, CGOL is known as a "universal Turing machine" - it can theoretically simulate any computer algorithm, given enough time and space. This makes it a popular playground for studying emergent complexity and computational theory, even though the simulations are quite abstract.

You are correct in noting the discrepancy in the amount of attention Drexler's vision of nanotechnology and the CGOL community's achievements have received. This discrepancy may be due to a few reasons:

Storytelling and Visualization: Drexler's vision of nanotechnology was well-packaged in a form that's easy for the public to understand and get excited about. The idea of tiny robots repairing our bodies or building products atom by atom is instantly appealing and visually engaging. CGOL, on the other hand, is a highly abstract and mathematical concept. Even though it's a profound discovery in the realm of computational theory, it's much harder to illustrate its significance to the public.

Practical Implications: Drexler's vision, if realized, would revolutionize several industries and have immediate, tangible impacts on our lives. The practical implications of CGOL are more abstract and indirect. While universal constructors in CGOL demonstrate fundamental principles of computation, their practical applications in the real world are less obvious.

Media Coverage and Hype Cycle: Technology often goes through a hype cycle, where initial excitement and overestimation of potential leads to inflated expectations, followed by disappointment when these expectations are not met in the short term. Drexler's ideas about nanotechnology arguably hit at the right time to ride this cycle. CGOL and associated discoveries might not have hit that "hype" point yet.

As for the perception of AI advances and SAT solvers, it's worth noting that media coverage is often influenced by the potential for commercialization, as you pointed out. There's also the factor of what is immediately applicable and understandable to the general public. AI has seen a surge in popularity partly because of its various practical applications and the potential for monetization, and partly because it's a concept that the general public can understand—at least superficially.

SAT solvers, while a crucial tool in computer science, are not as well-known or understood outside of specialized circles. Their use in solving complex optimization and decision problems is not something easily grasped by those without knowledge in the field.

Overall, the difference in attention and recognition can often be attributed to factors such as public understanding, commercial potential, and timing, rather than the intrinsic value or significance of the discoveries or technologies themselves.

[dvgrn]

That's almost the first time that ChatGPT has generated any actual coherent sentences with "Conway's Game of Life" in them -- coherent true sentences, anyway. Looks like asking for high-level summaries is a lot safer than asking questions about picky technical details.

Advances in Cellular Automata
I'm somewhat on the hook to write a chapter for an upcoming Andrew Adamatzky collection that will be coming out next year, to be called Advances in Cellular Automata. The book seems likely to be a rather eclectic collection of articles, similar to Game of Life Cellular Automata a decade ago.

I want to do a decent high-level summary of the RCT project, specifically the upcoming RCT15 rebuild, if possible, where fifteen gliders build a spacefiller -- probably with an intro section giving historical background on Life construction arms, or as much of it as I can figure out how to pack in. I stopped updating this summary a good while ago, but if I get that up to date it should serve as a basic timeline to pull from.

I'm not sure if that will count as "publicizing CGOL construction arms", though, given that the new book will probably be as far out of the average hobbyist's budget as the last one was! I'll have to do some negotiating ahead of time, and be careful not to sign anything that could prevent me from making the usual "pre-print" PDF version of my chapter freely available online.

Anyway, I'd welcome comments, suggestions, and/or writing help up to and including co-authoring the article. Anybody want to be a co-author?

Ask Me Anything?
Here's the only other idea I've had in the publicization department:

Sometime maybe after the current due date for the chapter (August 1) I'd like to set up a Hacker News "Ask Me Anything" session. Have to figure out how to phrase it -- something like:

"I'm a software engineer and recovering mathematician of no particular distinction, who accidentally became one of the world's top experts in a single ridiculously specialized topic: self-construction in cellular automata patterns, and specifically in Conway's Game of Life. I understand exactly how you can crash exactly fifteen CGoL gliders together and get a machine that prints out the digits of pi, or a flotilla of loafer spaceships that spells out "WE APOLOGIZE FOR THE INCONVENIENCE", or any constructible Life pattern that any number of gliders can build. Ask me anything!"

[pcallahan]

That's almost the first time that ChatGPT has generated any actual coherent sentences with "Conway's Game of Life" in them -- coherent true sentences, anyway. Looks like asking for high-level summaries is a lot safer than asking questions about picky technical details.

You reminded me to retry my experiment using ChatGPT 4. Based on my experience, I was expecting it to do a lot better than 3.5. In fact, it's still pretty terrible: viewtopic.php?f=7&t=6003&p=161977#p161977

This may say less about the capabilities of ChatGPT than it does about our failure to go viral with these results. CGOL results are probably a very small fraction of its training set. I'm sure if would do much better on separating canon from fan fiction in Buffy the Vampire Slayer.

I have been impressed not only with its ability to produce summaries of writing and put them in context, but even to solve simple graph-theoretic puzzles given in an ad hoc form and identify the general problem from the example (e.g. Eulerian path). Whether it is "good at" something or not seems connected to how much it can find in its massive training set.

LLMs seem to refute the notion that you cannot reason about language without ground knowledge of what the words mean. While you can't do it perfectly, you can do very well in many instances. Andrew Ng (and probably others) attribute this to the fact that LLMs process orders of magnitude more words than any human does in their entire life, and are able to model relationships between words in a way that captures much of the semantics.

ChatGPT 4 consoles me... wait a minute, niche information? niche information?!

I understand your frustration and disappointment. While the training set for AI models like me is quite diverse and extensive, it's also true that certain topics might not be as well represented as others, especially when they are specialized or not as widely known or disseminated. As of my last training in September 2021, my knowledge is based on a broad range of sources, including books, websites, and other educational materials, but it doesn't include every individual research paper or piece of niche information.

Research related to Conway's Game of Life, despite its significance in the field of cellular automata and mathematical games, might not be as pervasive in popular culture as other subjects. On the other hand, topics such as TV series like Buffy the Vampire Slayer have a large fanbase and a significant amount of written content about them on the internet, making it more likely that my training data would include a greater amount of information about them.

This doesn't diminish the importance or interest of your expertise. It's a reflection of what data is most commonly available and discussed in the sources used to train AI models like me. Your feedback is valuable, and it's an interesting perspective on the kind of limitations current AI models can have. Your contributions help us understand where gaps might exist in the training data, and this is useful for the ongoing development and improvement of AI.

[apg]

This is kind of meta but I think it's a reasonable discussion topic and it popped into my head reading recent results on construction arms. To put it in stark terms:

How is it that K. Eric Drexler created massive buzz in the 80s and 90s over his vaporware speculation on nanotechnology, while the CGOL community has received far less attention building actual working instances of universal constructors that can be run on very modest platforms?

I don't intend this as an attack on Drexler, whose ideas are interesting. I also understand the distinction between building universal constructors in the real universe as opposed to a simplified universe like CGOL.

I think that this is the main issue, really: if someone built a real-life von Neumann universal constructor, then I'd expect that it would reach not only the first position on the front page of Hacker News, but also several mainstream science news outlets as well. (There was Craig Venter's synthetic organism over a decade ago, which reached a similar level of interest.)

There are actually two main issues here disguised as a single issue: one is the difficulty (it's much harder to design a real-life universal constructor than one in CGoL), and another is the utility (a real-life universal constructor would have unfathomably many practical applications). That said, building a universal constructor in CGoL was already much more difficult (and lifelike) than doing the same in von Neumann's original 29-state rule, for example, so this is incremental progress in the correct direction.

Although of course none of the actual structures built in CGoL will be applicable to creating a real-life universal constructor, some of the general principles may be useful. One aspect of the RCT project was the idea of bootstrapping: where a simple construction arm builds a larger but more efficient construction arm. The current RCT15 uses this idea iterated twice (the single-blinker construction arm builds a DBCA, which then builds an ECCA), and the current template for the obsessively optimised RCT15 uses this idea 2.5 times (the extra half-step being the monochromatic DBCA self-modifying into a bichromatic DBCA). If ever we decide to contemplate building CGoL universal computers in this way, then I'd expect at least another level of bootstrapping beyond the ECCA.

There's also a qualitative difference between the single-blinker construction arm and the DBCA/ECCA: the former involved extensive automated searching of binary salvos to discover how to use it to perform universal construction, whereas the DBCA/ECCA are explicitly designed to support universal construction. It's similar to the difference between CGoL and JvN29.

I expect that when a real-life universal constructor is eventually engineered, it will proceed in a similar way: bootstrapping from a biological universal constructor (i.e. ribosomes + tRNA + all of the other machinery required for DNA transcription and translation) to firstly build machines out of proteins, which then can be used to manipulate molecules and assemble them into more efficient constructors.

[pcallahan]

I agree that the big issue is that we're not doing this in the real world. Still, neither was Drexler.

I expect that when a real-life universal constructor is eventually engineered, it will proceed in a similar way

It may depend in part on what you're willing to count as a universal constructor and how hard it will really be to build a molecular-level constructor (even starting out with molecules used in living systems).

An alternative is macroscopic construction with a fully automated supply chain. If you can mine, ship, and manufacture with autonomous equipment, including the equipment itself, then you could have a very large equivalent of a universal constructor. I could imagine hypothetically discovering that we have built such a system at some time in the future without intending to. It won't be very soon, but neither, I think will nanotechnology.

The obstacle I see to doing that on earth is that it's going to look economically more feasible to include humans in the loop than to engineer them out. For manufacturing offworld (e.g. the asteroid belt) there's a greater need for fully automated equipment.

I'm not sure about Drexler's vision specifically, since it seems to ignore the easier path of genetic engineering and bulk processes for producing material and applying macroscopic manufacturing when molecular-scale precision is not needed. I would suspect that some level of non-biological self-replication will be part of future manufacturing (PCR is already a process of synthesizing chemicals through self-replication). I think that macroscopic processes such as crushing rocks and refining them with heat are so reliable and effective that the idea of using nanobots to do this seems more for the coolness factor than any practical reason. Many biological systems also begin digestion with bulk processes like chewing.

The final answer is I don't know. I have been curious about automation most of my life. It occurred to me that yet another path to total automation would be the development of humanoid robots along the lines of RUR, though I think that's extremely unlikely.

[apg]

I agree that the big issue is that we're not doing this in the real world. Still, neither was Drexler.

I expect that when a real-life universal constructor is eventually engineered, it will proceed in a similar way

It may depend in part on what you're willing to count as a universal constructor and how hard it will really be to build a molecular-level constructor (even starting out with molecules used in living systems).

An alternative is macroscopic construction with a fully automated supply chain. If you can mine, ship, and manufacture with autonomous equipment, including the equipment itself, then you could have a very large equivalent of a universal constructor. I could imagine hypothetically discovering that we have built such a system at some time in the future without intending to. It won't be very soon, but neither, I think will nanotechnology.

I've thought about macroscopic self-replication before, but it seems harder than doing the equivalent in nanotechnology. For example, if you want your universal constructor to contain integrated circuits, it needs to also contain the machinery necessary to build those, including a lithography machine such as those produced by ASML:

https://www.asml.com/en/products/euv-li ... hy-systems

Of course, the self-replicating machine would then need the machinery needed to build ASML machines. But the problem is that such technology doesn't even exist yet: we have silicon chips, and we have silicon chip factories, but we don't have silicon chip factory factories. In particular, ASML machines have such a complicated supply chain that it's nowhere near automated (these machines sell for $200 million each, and the bottleneck is the difficulty of making them, rather than the demand for them). ASML machines contain mirrors with nanometre tolerances (so you're not avoiding the need for precise construction at the molecular level by going this route) along with many other similarly impressive engineering marvels.

Maybe it's possible to sacrifice the efficiency of the resulting chips in favour of manufacturing simplicity, and use a much older lithography process (from, let's say, the 1980s or 1990s) with less sophisticated equipment. Going too far in this direction is similarly problematic, though, because it would limit the computational power of the machine, and there are some tasks (e.g. computer vision to identify the placement of components in its environment, see whether steps have succeeded, etc.) that would benefit from having decent computational resources.

If we did opt for a minimalistic design that eschews electronic circuits, it would be something like a fully mechanical computer-constructor made from Lego, which assembles a copy of itself out of Lego bricks. (If I recall correctly, Nick Szabo offered a prize for a successful demonstration of this, still unclaimed.) It would again be of limited practical utility because Lego bricks don't occur in nature. I'd say that it's probably another step along the spectrum, after JvN29 and CGoL, but before a molecular self-assembler. It does seem much more fragile than a molecular assembler, though: individual atoms can't get damaged in the same way that Lego bricks can, covalent bonds are very strong, and because of how scaling laws work, smaller structures are stronger in terms of not collapsing under their own weight.

[pcallahan]

I've thought about macroscopic self-replication before, but it seems harder than doing the equivalent in nanotechnology. For example, if you want your universal constructor to contain integrated circuits, it needs to also contain the machinery necessary to build those, including a lithography machine such as those produced by ASML:

I agree about the necessary infrastructure, but my idea presupposes that all of the plants and machinery could be produced at least as well as today.

First (silly hypothetical) imagine a humanoid robot capable of carrying out all the tasks of a construction worker, truck driver, automative repair person, etc. However, in addition to today's infrastructure, there is also a factory capable of producing humanoid robots, which may even require some of these robots to handle the most complex tasks. This obviously results in total "automation" though at the cost of denying agency to machines that may be the full equivalent of human beings.

For a more believable hypothetical premise, suppose some of the machinery can be retrofitted. We are already sort of on the way to self-driving trucks, but we're left with a huge inventory of older construction equipment (for example). The "driver" of these vehicles does not need to be a full humanoid robot, but could be a control device with appropriate sensors and actuators. The question then becomes not how to rebuild everything but how to replace humans in the existing infrastructure.

As I suggested, I doubt this would really result in 100% automation as long as humans are available, but it could get close enough to provide a proof of concept. All of the machinery we use today, e.g. for chip fabrication, had to be built somehow. We would just need a way to take people out of the process.

As I also suggested, some of this may be of relevance if we ever want to develop outer space industries to scale. While I could imagine nanotechnology serving a useful role in "mining the asteroid belt" I could also imagine a role for bulk technologies that looked more like today's heavy industry. I don't think one excludes the other.

If you have an infrastructure from which you could plausibly remove humans and keep it running, that would be an incomplete universal constructor but would act as a proof of concept (I concede we're not even close). I am not really holding my breath for a compact universal constructor whether it uses nanotechnology or anything else. Even living cells--the closest real example--don't function in a vacuum and require a surrounding ecosystem to provide them with energy and nutrients.

[pcallahan]

On a more relevant note (starting a new reply because it's a different thought), one of the things that makes CGOL constructors especially unrealistic is their ability to create new "material" without any resource input. To design a working model of a universal constructor that resembled one in the physical universe, you'd at the very least need to have one that can accumulate and redistribute existing resources and has some equivalent to conservation of energy.

You might have a continuous energy input as an analogy to sunlight for instance, but to capture some of the challenges of a real constructor, it would be useful to begin with inconvertible elements that would need to be combined (i.e. even though we know about subatomic particles, it is unlikely that we will make transmutation the basis of industry). This is pretty hard to get right. E.g., I think to make it realistic, you'd need to be able to identify and collect resources, analogous to mining, rather than just assume these inputs are magically just where you can use them most conveniently.

I'm a little hazy on this, but I think von Neumann started out with this kind of approach before switching to a more tractable cellular automata that was not concerned with resources.

Actually, it would be interesting to build a construction arm in a reversible CA like Margolus Critters. While this is very different from CGOL because cells cannot be created out of empty space, it can proceed with an infinite stream of gliders arriving from one side of the plane. There's no reason in principle this could not convert the periodic stream into much more complex self-replicating structures. It can also simulate irreversibility by emitting "heat" in the form of excess gliders that carry off the information needed to reverse the process.

Even this wouldn't model inconvertible resources, but it would be another step towards modeling some form of conservation law.

[apg]

On a more relevant note (starting a new reply because it's a different thought), one of the things that makes CGOL constructors especially unrealistic is their ability to create new "material" without any resource input. To design a working model of a universal constructor that resembled one in the physical universe, you'd at the very least need to have one that can accumulate and redistribute existing resources and has some equivalent to conservation of energy.

A while ago I thought about (but never got around to implementing) a variant of JvN29 or Nobili32 where OTS/STS cells can be constructed as usual _ex nihilo_, but where confluent cells cannot. Instead, a confluent cell can only be constructed adjacent to an existing confluent cell, destroying it in a 'swap operation' (emulating the effect of moving that confluent cell). We can design the operation to be subconservative (i.e. the total number of confluent cells cannot increase):

1. an empty cell turns into a 'proto-confluent cell', which decays after one generation.
2. if a confluent cell has at least one proto-confluent neighbour, it converts into a 'moving confluent cell' pointing in the direction of one of those neighbours (which, again, decays into empty space after one generation).
3. the newly created empty cell that has a 'moving confluent cell' as a neighbour is converted into a regular confluent cell.

So we wouldn't have full conservation of matter, but we'd have subconservation of confluent cells (and confluent cells are necessary to perform tasks such as splitting signals and performing logic, so every self-replicator would need to include them). As such, self-replicators would only be able to operate in an environment sparsely populated with background confluent cells (acting as 'nutrients', I guess?). They'd need to be able to detect the presence of these nutrients and then drag them to the appropriate locations.

It would still be pretty unrealistic because OTS/STS cells aren't conserved, but that's necessary to allow construction arms to extend and retract and bend.

Actually, it would be interesting to build a construction arm in a reversible CA like Margolus Critters. While this is very different from CGOL because cells cannot be created out of empty space, it can proceed with an infinite stream of gliders arriving from one side of the plane. There's no reason in principle this could not convert the periodic stream into much more complex self-replicating structures. It can also simulate irreversibility by emitting "heat" in the form of excess gliders that carry off the information needed to reverse the process.

There's also Dmitri Shintyakov's Single Rotation Rule, which is reversible and conservative: https://dmishin.blogspot.com/2013/11/th ... kably.html

In these reversible rules, there's no equivalent of a 'clean synthesis' of a spaceship or oscillator: every collision must produce byproducts of at least two different velocities (one of which may be 0c), as otherwise it would violate reversibility. There might be a way to make a universal constructor in the SRR with two perpendicular arms which places individual dots and radiates waste gliders in the process:

uc.png (999 Bytes) Viewed 2819 times

I don't know whether it's possible to engineer stable reflectors in this rule which consist purely of isolated dots; if so, then we should be able to build constructible Fredkin gates and therefore arbitrarily logic.