/sci/ - Science & Math

>>15434696
Automated resource extraction and conversion.

>>15434702

How close can we get? In the picture, the robots simply assemble more robots from prefab parts. Can we do that? Can we build a factory which, for example, has dual assembly lines? Line A produces parts for robots, which assemble those parts into more robots. Line B produces parts for factories like itself, which the robots assemble into new factories.

>>15434696
>>15434702
Because Von Neumann's conception of cellular automata, while clever, are fundamentally flawed
In biological life, the constructor, instructions, and universal copier are all the same thing. We abstract away DNA as being a set of instructions but in reality discoveries in epigenetics and biological chemistry show DNA isn't just a set of instructions, it part of the machine itself
So it won't work because the logical premise is already flawed. I'm not claiming to be smarter than Von Neumann, but he was a man and every man makes mistakes

>>15434711

Maybe we can't make it that small, but the human operated industrial substrate upon which we depend can self-replicating, self-repairing and self-expanding, given only a humanoid robot with sufficiently intelligent AI, able to use all the same vehicles and tools as humans. We actually seem pretty close to that now. It would be very big and clunky, but do basically all the things that cells do and more

>>15434711
pseuds should just lurk instead of making embarrassing posts

>>15434716
>Maybe we can't make it that small
You're misunderstanding the point, it's not about how big or small. Any organism that copies itself doesn't start from raw materials, it starts from itself. The machine and instructions are one in the same, you can't have a truly self-replicating self-evolving system if those qualia are discretized
>given only a humanoid robot with sufficiently intelligent AI, able to use all the same vehicles and tools as humans. We actually seem pretty close to that now.
We are nowhere close. The AI that people claim is going to kill us all in 10 years can't even understand how to visually represent math, try asking an AI image generator to generate text in a picture.
>We actually seem pretty close to that now. It would be very big and clunky, but do basically all the things that cells do and more
This is the most "practical" reason, at our current technology level

>>15434717
Wow, good response, nice argument to back it up, very convincing logic you've laid out. You're so smart anon!

>>15434723
>try asking an AI image generator to generate text in a picture.

This is like when people said it couldn't do hands. Then it could do hands. None of the current limitations are permanent, this is an emerging technology experiencing rapid growth. I would wager text would be easy to implement if only via a plugin, like how it can use Wolfram Alpha now to get math problems right, but then it would generate images that might incorporate copyrighted brand names

>>15434729
The particular task of representing text as an image is data fusion which AI sucks at. An AI image generator can "do hands" because images of hands are part of the data set it was trained on, the information is there the functions transforming it just needed to be refined. That's not the case with text which needs to be "understood" as an "image" by the AI

>>15434735
>The particular task of representing text as an image is data fusion which AI sucks at.

Right, so give it a text plugin. This is also how the hallucination problem was solved for math, they gave it a calculator.

>>15434729
Also like the last 30 years of computer vision have had understand text as one of the major goals and we've only just gotten it right the last 5 years or so

>>15434740

Gains are exponential, not linear.

>is muh soience fiction hollywood plot device real
IQ < 110

>>15434737
>Right, so give it a text plugin
You really don't understand how AI works. How are you going to make a "plug in" that makes your text into a grahpical patter, then trains the AI to recognize that pattern and that patter alone to correspond to the text you entered, but still be able to change that pattern enough to vary it according to other prompts while still being recognizable as the original pattern?

>>15434745
Until it isn't
see: Moore's law

>>15434749

Easy, the AI knows where it intends to generate text in the image. We know that text will be insufficient. Have it omit the text or edit over it using a standard set of fonts, according to the text specified in the prompt by the user. This is a bandaid but it gets the job done.

>>15434753
>Easy, the AI knows where it intends to generate text in the image
Not always, even now
> Have it omit the text or edit over it using a standard set of fonts, according to the text specified in the prompt by the user.
Ok, HOW?

>>15434752

Still seem to be in rapid ascent currently, nobody knows where the ceiling is. The workaround for moore's law wasn't further miniaturization, per the original definition, but out of the box solutions like multicore, 3D architecture, quantum computing and organoids.

>>15434753
>This is a bandaid but it gets the job done.
Yes, by not using AI and by using a brute-force solution implemented by humans
Seriously, fucking read a paragraph on how LLMs work

>>15434755
>Not always, even now

If it doesn't know where the text will appear, it can't draw it there to begin with. It does generate text, albeit nonsense, so it knows where it will put it during generation.

>Ok, HOW?

See above, but add a word processor able to edit in characters from a font set over the image, like meme generator websites. This is not the show stopper you imagine, every problem is solvable.

>>15434756
The CEO of the most successful chip maker on the planet and leader in AI hardware disagrees with you
>muh appeal to authority
He's the guy literally making them, you think anyone knows better than him?
Even if you need evidence, look at this graph. Are

>>15434758

I don't see what the problem is. Humans have to implement the workaround but they don't have to manually make it work every time after. Probably this sort of kludge is exactly how the limitations of LLMs will be overcome, in a bunch of different capacities, like how adding a calculator was the solution to hallucination re: math answers

>>15434762
>The CEO of the most successful chip maker on the planet and leader in AI hardware disagrees with you
>He's the guy literally making them, you think anyone knows better than him?

Produce a direct quote where he specifically says that LLMs will never get text right

>>15434761
>See above, but add a word processor able to edit in characters from a font set over the image, like meme generator websites.
It's not AI then, it's a human-generated solution or two AI working together through a bridge humans implemented
>>15434762
Raging so hard I forgot my graph. Even accounting for the log scaling does the delta look consistent to you at all?

>>15434767
>It's not AI then, it's a human-generated solution or two AI working together through a bridge humans implemented

Why, because humans had to create it? LLMs didn't create themselves, humans did.

>Raging so hard I forgot my graph. Even accounting for the log scaling does the delta look consistent to you at all?

This is not the only dimension of AI growth. Have you noticed the recent steep acceleration of breakthroughs?

>>15434765
We're talking about Moore's law you disingenuous little faggot, not LLMs
>>15434763
>. Probably this sort of kludge is exactly how the limitations of LLMs will be overcome
Two idiots talking to one another doesn't make a single genius, sorry. You just don't understand how AI works

>>15434771
>Have you noticed the recent steep acceleration of breakthroughs?
I've noticed an increase in flashy parlour tricks meant to impress midwits like you, yes
The internet is a really big place with billions of people. Anything "weird" an LLM chatbot does is the result of it being trained on a dataset including the description of that behavior. Until they cross-reference every AI result with every piece of data it used for training and prove me wrong, I'm quite confident there's 0 evidence of emergent properties from current LLMs

>Two idiots talking to one another doesn't make a single genius, sorry. You just don't understand how AI works

Nobody does at this point, but we can reason about it based on observed capabilities. Your problem is you have goalposts set up in your mind for what true AI is, and the goalposts are arbitrary. LLMs are already a workaround. They are not and may never be conscious, but that doesn't matter in terms of their impact on the economy and humanity if they can do many things conscious humans can do. It's the result that counts. Likewise, it doesn't matter if the text problem is solved by a workaround, if it is solved. That a human had to implement the workaround does not make it illegitimate, as humans implemented the LLM itself to begin with. No AI is self bootstrapping.

>>15434775

Why are you needlessly insulting to strangers? I don't hate you, we are not enemies. We share a common interest, even. Is everything ok in your life?

>>15434776
> That a human had to implement the workaround does not make it illegitimate, as humans implemented the LLM itself to begin with. No AI is self bootstrapping.
The entire point of this thread is "self-replicating machines" so unless you've completely moved your goalposts, having a human in the loop for the most advanced "machines" right now means universal constructors are DOA

>>15434777
>Is everything ok in your life
No, I'm trying to teach a retard how AI works and he isn't getting it and it's making me annoyed. Maybe I should just have compassion for the retarded though

>>15434779

It was my understanding the scope of the discussion had shifted to LLMs

>>15434781

I see, so that's just what you're like. My condolences to people in your life who cannot escape from you as easily as I can, like coworkers or family members. Likely everyone able to desert you already has, and evidently with good reason.

>>15434783
At the end of the day you can't escape how stupid you are

>>15434787

Nor can you escape how repellent and alone you are. I, at least, am loved.

>>15434790
There is no way for me to verify to you I have friends and family at all so it's pointless to try to argue against it or for it.
Meanwhile, the evidence that you are very dumb and have no understanding of LLms, universal constructors or AI in general is spread all across this thread and will be memorialized in the archive
Have a good night dude!

>>15434726
in his defence his post quality matched yours but took less space.

>>15434696
Principal roadblocks implies there is a functioning prototype based on sound engineering principles. There is none, Von Neumann machines describe an outcome, not a design. There is no known principle that maps some huge possible set of environments into a fixed few.

>>15434696
A family with lathe. The universal and fundamental questions are left unanswered in the age of specialization. Researchers can't even self-fund thus became paid shills now.

>>15434815
>angry midwit noises

>>15434707
That would be possible today already, given a suitable set of "prefab parts". One of those car assembly line hand robots could easily assemble itself with small modifications to the robot and the design. The issue is that final assembly is the easy part. The manufacture of components is much harder, some of these things require huge machines (like making chips), variety of materials that you have to mine and refine (your robot needs to be both a steelfoundry as well as a iron mine as well as an energy source). There's simply no way to make that type of thing economic or even interesting. Why try to make some universal robot to work in your iron mine when you can just have a guy sit in a dedicated miner who works for the fraction of a price and produces much more. Now it works the best in the actual von neuman probe scenario where you send a probe in the kilometer size range to another system. There economics don't matter and with something that big you have the space to put in your steelfoundry and asteroid miner and 0g litography and 3d printing units etc.

>>15434696
Three things:
1. Making computer chips is very hard and requires a huge amount of precise work, and for the older simpler versions those are much more unreliable in addition to preforming worse.
2. Producing the energy required would need a continuous supply of fuel, since renewable methods aren't ready for it and most people probably wouldn't trust a nuclear reactor without human oversight.
3. Making any modern component takes dozens or hundreds of raw materials, each of which are strewn unequally across the whole complicated world. To truely replicate itself autonomously it would have to either have so much power it can atomize shit to get access to rare stuff or it would have to comprehend geology and plan the search for and exploitation of materials all around the world, esentially making it's own world-spanning supply chain.
Maybe on another untouched world it could find what it needs but on earth it'll have to be fed with feedstock.

>>15434707
I don't know but no factory produces all the parts necessary for building another one because that would mean everything from ore extractors to smelters, forges, presses, wires, magnets, computers, &etc. This is before we even get to the problem of energy generation and routing.

Basically, it's an impossible task.

How's reprap going?

>>15434729
There is no "it". It's a bunch of people behind the scenes running in hamster wheels to make sure that retards like you continue to think AI is magic. There are knowledge workers in Kenya grading your retarded questions to make it seem like there is a new kind of intelligence when it's really just heroic human effort that is invisible to idiots that think AI is magic.

>>15434783
Women are not allowed here. Go back to plebbit and twatter

>>15434711
>We abstract away DNA as being a set of instructions but in reality discoveries in epigenetics and biological chemistry show DNA isn't just a set of instructions, it part of the machine itself
That's not really a problem, we can build machines with instructions built in.

>>15434711
Retard von neumann made his machines precisely like that.
Every cell follows the same rules.
Some function as memory, others as copying, etc.

>>15434696
>>15434696
>But research into self-replicating machines seems to have stagnated since the mid 1990s
Xenobots have achieve a lot of success lately.

>>15435204
He is probably a shcizo poltard type. Just ignore him.

>>15434711
>Unironically thinking you understand mathematics better than von Neumann, who was literally on of the greatest mathematicians in history.

lol, lmao even. I think you should probably take your meds.

>>15434696
I mean, I guess it’s this organism was as big as a factory… containing all the fabs needed to build itself. And had drones that could setup mining operations across the globe somehow… wait .. thats just describing a super organism…

>>15436553
>/pol/ derangement syndrome
>appeal to authority instead of critical thinking
Yup, it's an NPC

>>15434753
>the AI knows where it intends
Fucking halfwit retard LLMs don't "know" or "intend" anything.
Go back to /x/ or /b/ or better yet /r/eddit.

>>15434753
>knows
Cringe.

>>15434975
retarded post, look into running local models on your own hardware without even a net connection

>>15436545
Xenobots are a crime against nature

>>15434696
>What are the principle roadblocks inhibiting advances in this area?
the unavoidable fact a computer program is composed of an input, a program and an output, all of which must be defined by a programmer beforehand
>>15434702
the roaddblock is reality itself. consciousness is metaphysical, science being naturalistic, materialistic, it can never reach such a realm.
the evidence of that is despite even taking the neuroscience claim seriously if the brain being everything (clearly not the case due to medical literature) involved in the consciousness aspect (or "AGI"), is still able to achieve what not even the most powerful computer can and in a tiny space and weight (1260 cm3 and 1.5 kg), not to mention with a tiny, tiny energy consumption ("12 watts" https://press.princeton.edu/ideas/is-the-human-brain-a-biological-computer )

the AI field is plagged with the atheist problem, it ironically lacks intelligence. they have a problem with understanding the first cause argument, which reflects on their thinking everywhere, including STEM. They have been trying to create intelligence in machines by continuously creating machine made artificial thoughts or artificial thinking. problem is, that is no different in essence to writing or any typeof recording, yet you wouldn't mistake writing (as in the written text itself) with an intelligence. see the second section (não existe inteligência artificial)
https://olavodecarvalho.org/inteligencia-e-verdade/

>>15434711

>In biological life, the constructor, instructions, and universal copier are all the same thing. We abstract away DNA as being a set of instructions but in reality discoveries in epigenetics and biological chemistry show DNA isn't just a set of instructions, it part of the machine itself

This. And the "machine" is constantly self adapting, or rather moving within a very complex range of its own intercoiled equilibria. A mechanical replicator as envisioned today would simply be much too "rigid", too constrained to mimic this functionality. The same does in part apply to the "AI" research currently done.

>>15437075
>the roaddblock is reality itself. consciousness is metaphysical, science being naturalistic, materialistic, it can never reach such a realm.

Hahaha

>>15437153

>more than the sum of the individual parts

The really funny shit happens at the fluid intersection of these. Current approaches in the field are completely impotent in trying to achieve this. Only thing I could see wörking would be to have the machine getting "imprinted" through prolonged exposure to an actual wetware netwörk by a lernfähiges man-machine-interface. With the crucial part here being the "translator" function between the two systems. Current algos might be capable of covering this, not in general but on a very distinct individual level.

>>15437153
nice argument

>>15434711
What you lament matters only for the literally very first couple cells, at least among multicellular organisms, because there's no separation between creator and createe yet.
I.e. you need to boostrap yourself from one singular atom (the atom may be a cell, etc.)

Self-replicators literally don't need that. We can just activate 10 of them for the lolz. This is enough to ensure they can keep going and repairing each other indefinitely.
The code and construction doesn't need to be unified with non-atom-arising silicon-based constructors. You simply didn't even think this through. Just have the robot store the code on an Arduiono, and have the robot arms be the creator. Whatever. What is the issue? Your whining becomes more absurd the more you think about it.

Your IQ is 110><120.