Hello, cultural singularity—soon, every video you see online could be completely fake.
See full article...
See full article...
OpenAI collapses media reality with Sora, a photorealistic AI video generator
- Thread starter JournalBot
- Start date
VividVerism
Ars Tribunus Angusticlavius
In many, many jurisdictions: yes. It would still be illegal. Like, it's not even a gray area. The US, for example, bans virtual sexual content content with imagery that is "indistinguishable from that of a minor".
There are some jurisdictions where it is probably legal, but most likely you're breaking some very serious laws somewhere if you're only virtually abusing virtual minors.
https://en.wikipedia.org/wiki/Legal_status_of_fictional_child_pornography
Last edited:
Upvote
3
(3
/
0)
The Sheep Look Up
Ars Praetorian
That's a bold assertion. Got any evidence toward this that isn't pure hopium? Because AFAIK there's absolutely nothing that strongly indicates that ML or GANS will be sufficient to get us to AGI.
Upvote
10
(10
/
0)
It's a guess, not an assertion. But now we have a simulated neuronal tech that is eating up the gaps between machine and man; tech that can read and understand - at a low level for sure, but a similar level to that of our unconscious or semi-conscious processes - everything that we have written. Everything we have drawn. Tech with a rude mechanical power that we cannot emulate (could you draw even a few frames of such a video?)
True, the subsystems to achieve AGI have not yet been placed together. The painfully evolved logic that allows us to take some small steps forward from the unconscious calculation, maybe even modulate the calculation to incorporate our rational understanding of a sphere of thought. But the space where we can imagine uniquely human magic is exponentially contracting. And we cannot even hope any more that an AGI will need a long time to learn as an agent in the world, because we know it can absorb meaning from literature.
I don't think the hopium is on my side.
Last edited:
Upvote
-2
(1
/
-3)
The Sheep Look Up
Ars Praetorian
...and that tech is? ML and GANS != neuronal processes.
No, it cannot. It can utilize input as a transformer to determine a serial output from a weighted matrix generated via a training set. That's the whole game. Any description beyond that is poetry.
...and may not exist. At least as of yet. I don't see why it's impossible, but it just might be.
As it's a pure white space hypothetical, we can "know" anything about it we can imagine.
That's a lot of words for what boils down to a lot of wishcrafting. You really could've stopped with the WAG. Point still stands, we don't know that AGI is possible, let alone that our current generation of tools are on the path to reaching it. Again, unless you know something I don't, some demonstrable marker that the hype men have mysteriously been silent about.
Upvote
3
(4
/
-1)
For roughly US$20,000,000 you can currently do that.
Snakes on a Plane being a real world example of that being done.
The allure of AI is that it will be doable for under US$10,000.
There has always been a discongruency between what an artist is trying to portray, and what the recipient percieves as the message.
Is the sin of Genesis 19:
- Homosexuality;
- Inhospitablity towards others;
- Something else;
Or from the Q'ran 19:19-20, is the text:
- A warning;
- A commandment;
- Something else;
Roughly a decade ago, on a publishing forum, on a thread about revenue generation, one participant was proposing that an author do all of the following:
Create the:
- ebook. (mobi, epub);
- Mass market paperback;
- Four colour hard cover. Pictures of people, places, objects are included;
- Four colour limited edition collector's leatherbound version.
* Cartoon/manga edition;
- Table board game;
- Computer video game;
- RPG trading card game (^1);
- Basic audio book. Either software generated, or the author or a voice actor reads it;
- 'Theatre of the mind' audio book. Software generates Foley effects anda different voice for each character;
* Audio soundtrack to the book. (^2);
- Animated cartoon edition;
- 'Made for TV' movie edition;
All of those can be created by the author, on their home desktop, using then currently available FLOSS, and the author's time. No AI required.
With AI, the process to create those works is speeded up, and, when it works as claimed, is less expensive to produce.
^1: We are talking about one-off Mills & Boone/Harlequin romance here. Or any other mind-candy you indulge in.
^2: This need not be the soundtrack to the audio 'movie of the mind' edition, the animated cartoon edition, or the 'made for tv' movie edition.
Upvote
-1
(0
/
-1)
Yes yay.
It's like how the ubiquity of mobile phones with cameras democratised photography. It's fine if I take a photo that is only seen by myself and maybe ten other people from my friends and family. It still enriches our lives. And other people are doing the same thing, with different small audiences personal to them. There is a tendency to see the internet as being about few-to-many publishing, as for example a handful of Ars writers publishing to hundreds of thousands of us readers. That happens, of course, but a lot of the internet is many-to-few publishing. A great many people producing content that is intentionally seen by only a few others. That's what Facebook is about, for example. Not everything has to be worthy of putting up in an art gallery. Not everything has to get a million likes.
If generative AI extends this to making videos, than I'd see that part of it as an upside.
Upvote
-1
(2
/
-3)
Due to the way people in Adult Entertainment movies are paid, we have seen the actors and actresses in movies doing things that they were not filmed doing.
The sole reason mainstream movies have not yet done that is financial. Currently, it is marginally cheaper to pay an actor or actress to be physically present for the shoot, than do everything digitally. (Adult Entertainment contracts are very different from mainstream movie contracts.)
It would have very simple for a mainstream studio exec to tell the union, "Thank you for your offer, but we no longer need your services", and follow the practices of the Adult Entertainment movie industry.
Upvote
0
(0
/
0)
If I remember correctly Australia banned or tried to ban porn in which the actress look underage, even if they are of age.
Upvote
1
(1
/
0)
Crazy idea. Have the government force the companies that make this software to watermark their creations. At least in the United States. I think the EU would go along with this. Russia not so much. Or China. And you could not trust Canada.
Upvote
0
(0
/
0)
I think Aurich's statement should be revised to: "LLM's can't even do basic math reliably and consistently, something computers have been good at since computers first existed."
Note that the second link states that the AI did "at almost the same level as a human gold medalist." Humans are fallible and not as consistent and reliable as a simple calculator, and the AI did worse than that.
As for the other article, the advantage with the LLM was that it was able to produce attempted solutions quickly, not that it was inherently that good at it. They needed another program to evaluate the solutions the LLM produced.
https://www.scientificamerican.com/article/ai-beats-humans-on-unsolved-math-problem/
Upvote
3
(3
/
0)
LLM's aren't the only DL models.
Competing at a level comparable to a Human Gold Medalist, is way beyond a simple calculator. Here is a sample problem from that competition:
Will your calculator solve that for you?
Upvote
-1
(1
/
-2)
Yeah, examples of LLMs hallucinating ridiculously wrong math are rife.
Upvote
2
(2
/
0)
VividVerism
Ars Tribunus Angusticlavius
And the software that doesn't have a corporate backer but is actually open source? Or the software that runs locally and can thus be modified to not generate a watermark? Or post-processing to remove the watermark?
That's ignoring the fact that it doesn't really do any good to block US companies only if other countries release cheap or no-charge equivalent software that can be used from anywhere in the world.
Upvote
1
(1
/
0)
They were the only models used in the articles you cited. Do you have another article?
I don't really pretend to be good at math, but then I'm not touting an LLM pretending to be good at math either.
Upvote
2
(2
/
0)
Upvote
1
(2
/
-1)
Large Language Models are not designed to do math. They're designed to do generative language work, hence the name.
This is sort of like saying claw hammers make bad screwdrivers. Self-evident to anyone in construction except maybe an electrician.
Upvote
-2
(2
/
-4)
Yes, but it's clear from the context that Aurich's original comment was simply pointing out that pooling the disparate and very specific capabilities of various technologies to suggest that something as advanced as AGI combining those abilities into a single entity is both inevitable and imminent is kind of getting ahead of things.
Upvote
4
(4
/
0)
It'd be poetic that the bosses funded these AI versificators to suppress creative labor, and end up annihilating themselves.
I sure hope the 'final generation' people meant to bring human-produced content to a close and start with purely AI-gen content realize this, flee en-masse and end up creating their own smaller co-ops that use AI to help build their own pet projects, which then turn out to be much more cutting, human, and entertaining than the tripe they were forced to create for the bosses the old-school way. Oh, and the money too. Especially the damn money.
I gotta believe in an anime-style last-minute uppercut from the guy on the mat somewhere and sometime, otherwise what's the point?
I sure hope the 'final generation' people meant to bring human-produced content to a close and start with purely AI-gen content realize this, flee en-masse and end up creating their own smaller co-ops that use AI to help build their own pet projects, which then turn out to be much more cutting, human, and entertaining than the tripe they were forced to create for the bosses the old-school way. Oh, and the money too. Especially the damn money.
I gotta believe in an anime-style last-minute uppercut from the guy on the mat somewhere and sometime, otherwise what's the point?
Last edited by a moderator:
Upvote
0
(0
/
0)
I think it's understandable that people look at the pace of progress and just keep extending the curve out. I'm hardly immune to that myself, look at my reaction to this article.
But I think I'm reasonably concerned because the tech is basically "good enough right now" to do the things that worry me. We don't need a generative leap. And the improvements to the issues people are rightly pointing out seem like the kind of thing that have generally been coming to pass.
But as you very correctly point out, just mushing together these various models into some kind of super AGI is a whole other ball of wax. Yes, there is software that's good at math. But the "AI" that gets all the headlines these days (and is the style being discussed in this article)? Terrible at it.
Upvote
4
(4
/
0)
I recall watching The Empire Strikes Back when it came out and exited the theater thinking to myself: "They've finally gotten to the point where anything can be filmed if it can be described." Everything since, CGI included, has just made it all easier, faster, cheaper, and better. That's pretty much going to happen to anything we are willing to throw money at.
For a real laugh, go to YouTube or some such and look how primitive 1960's cartoons were, even compared to what people have been doing alone in their mom's basement for 20 years now.
For a real laugh, go to YouTube or some such and look how primitive 1960's cartoons were, even compared to what people have been doing alone in their mom's basement for 20 years now.
Upvote
1
(1
/
0)
Transformer based LLMs are bad at multiplication in part because they can't easily represent that type of calculation. A NN can approximate any function given sufficient width and / or layers, but some functions are more naturally expressed than others. The models can usually do addition with an arbitrary number of digits, but can't multiply with more than 2-3 significant figures. If it was a priority, this should be relatively easily fixed ... simply additional wiring that allows the network to emit the multiplication of two activations, rather than only the linear sum of weighted activations, and then the model would more easily learn such patterns intrinsically. But it's more practical to instead train the LLM to emit python code that performs the calculations with full precision. That produces a better end-to-end product experience.
I agree with your perspective that many of the current limitations of the technology are things that seem relatively straightforward to fix, and thus, while it will take diligent work to do so, it's highly likely that such deficiencies will inevitably be fixed within a fairly short time period. Versus wild predictions may come true in some form at some point, but it's very hard to say exactly when and exactly how. It's hard to predict the full spectrum of techniques that will be invented or how they will be combined, but we can look at the obvious next steps and see that significant capability improvement is likely merely from basic iterative development even without any transformative capability jumps.
Upvote
1
(1
/
0)
My reason for believing that full AGI isn't far away is based on the belief that the rational part of our minds is a relatively thin layer based on a variety of unconscious substructures that embody our actual understanding of the world and sensing of its components. Previous attempts at AGI tried to construct the rational layer, and it wasn't enough. But it seems like we're getting a handle on simulating the substructures. Putting them together may be a lot easier and faster than most imagine.
Upvote
0
(0
/
0)
The thing is, a Death Star is both easier to destroy (as it is a physical object centralized to one location), and easier to justify violent action against.
What realistic actions can we take against other nations to prevent the study and development of this technology? What about civilian actors working on it? Businesses? Do we have the stomach to punish entire nations for what private companies and citizens are doing? If economic sanctions and other non-violent measures don't work, do we have the stomach to move to violent ones? Is that even justified?
I'm definitely for doing something about this, but frustrated that no one here seems to have any idea what 'doing something' looks like. Just "don't allow it!" And in the topic of taking political action, that's simply not enough to go on.
Upvote
0
(0
/
0)
Many thousands of years ago: "Look honey! The kids are staring at these pressed sheets of papyrus all day just because someone made squiggly marks on them. That can't be good, in fact, it's pretty weird. I mean, they just sit motionless and stare at something that doesn't even look like anything! If they want a story or something they should do it the old-fashioned way. What was wrong with just asking someone to tell them one?"
Upvote
-2
(1
/
-3)
That's what humans do too -- they recognize a math problem that is too complex in their heads, and realize they need to use a tool.
Aka, humans pick up a calculator. AI picks up Python or Wolfram or whatever calculator they have access to.
Neural network (biological or artificial) will never ever do classical math as well as traditional logic computing -- but they don't need to be.
Smarter tool-using ability is key. Such abilities can be trained-in, including various methods of eletronics calculators/computing, and self-checking step by multiple methods (via rote, via using logic calculator/program, etc).
When they recognize a problem is too risky to solve without calculator help. To help solve a complex math problem, or to self-doublecheck the steps of a manually calculated answer (if configured/permitted to use sufficient compute power to doublecheck itself). Even a notepad is a tool, yet a lot of math disciplines now mandatorily requires a calculator. Lest it takes decades for a single human to solve specific kinds of problems on paper.
Hats off to those virtuso <0.0001% human mathematicans of yesteryore who pulled off math breakthroughs with merely paper in the pre-slide-rule era; my respect. But today, we now solve math problems that takes way beyond millions of humankind-years to solve. So, electronic calculators (and programs).
Likewise, the same problem happens to the various neural network techologies too; math isn't a discipline ideally suited to AI. But in theory AI can be superior calculator-users (eventually), e.g. better interfacing to classical computing/calculating. Current AI's are only now finally beginning to learn to recognize the math problems, and breaking it down to steps. Humans (90% of them, on average) are also terrible at math problems without tool help. But when AI gets superlatively good at deciding what math tools to use (e.g. interfaces to classical computing) the line item will eventually be game-over like chess superiority and the ilk;
Indeed, math-tool-using abilities are not quite university league thought processes (we'll get there rather surprisingly rapidly) -- but two years ago, AI's didn't even recognize math problems well enough to recognize that they should automatically use a calculator.
Now, today, by default settings of the AI app (in its paid mode, no other settings configuration) -- when I ask certain subscription AI's to "Please calculate the compound growth rate for turning 10,000 into 22,000 in six years", it now automatically decides to use some tool. Such as writing a 10-line Python program by default and then running it in a sandbox. Basically, the AI metaophorical equivalent of picking up an electronic calculator that it had access to, and then using it. Even if the tool-using-decisions are quite imperfect;
It does seem to be inevitable -- the size of computer programs that an AI can successfully write from scratch is currently geometrically increasing; and that is concerning -- eventually 1000 line math calculators will be made to solve complex math problems, on the first try. (Right now for the better engines, it's about 50-100 lines success on first try, with proper prompt engineering). Ability to revise programs (repetitive iteration) is still really crap, but rapidly improving as various levels of memories are implemented (context memory, medium term personal training memory, and long-term training tweaks, in addition to baked-in training) as better deduction in step-by-step problem solving that now takes more conversational turns before it screws up.
Eventually (given sufficient tech progress on all fronts) it is probable that this math-tool-using ability will be superior to most humans too -- knowing when to use a general-purpose computer for certain kind of math steps -- and displaying formulas step by step while doublechecking its steps with a program, etc. We're not quite there yet, but it appears rapidly hurtling towards that;
There are the upsides too, but also the downsides. (Witness Aurich saying we're handing loaded guns to kids)
...Witnessing the loss of ability to read paper maps
...The loss of ability to memorize multiplication tables
...The loss of ability to answer many basic questions without consulting a smartphone, etc.
And the thorny stuff, access to information/images (real or AI generated) that kids (and even adults too) should not have access to, all trained on information that may be copyrighted (and other controversial issues, such as image/video being unethically purchased/shared between companies for the purpose of training AIs). With AIs in everybody's hands, I fear a lot of humankind skills may atrophy. It's already happening in many disciplines due (directly, indirectly, etc) to many tools such as smartphones, Internet, etc. Imagine an era in the distant future where humankind is dangerously wholly dependant on AI tools (and AI training influenced by very few parties); now that's worrying when you think deep about it;
The skills that atrophies when one can simply Google for a lot of things we used to learn. Now this will be even more on steroids.
I certainly understand the "loaded gun" sentiment.
While I am excited for AI and its applications to helping my deafness & other aspects of life that I am terrible at (organizing notes etc) -- I can't but feel worried I'm contributing to the problem by jumping on the AI bandwagon (even for less controversial uses such as captioning/notetaking/etc). It was more harmless to feel guilty I'm using the dishwasher instead of manually washing dishes...
How do we proceed wisely, is the question;
Last edited:
Upvote
2
(2
/
0)
On the topic of discussing AGI, it is useful to define "What is an AGI?". For example, we're already beyond humans for certain lineitems (chess) but far short of humans for other lineitems (skills needed to do the dishes).
Part of the debate, could include "Does AGI have to do dishes?" "Does AGI have to also have smell and taste senses?" etc.
Regardless, the new DeepMind AGI definitions (arxiv) are quite an interesting read:
The industry seems to be migrating to a 6-level AGI rankings:
From page 6 of paper: https://arxiv.org/pdf/2311.02462.pdf
It can be argued that an average human does math really badly (without a calculator) and hallucinates facts (see the crazy politics and fishing stories), so GPT-4 may be outperform the worst 10% of humankind, especially those that did not have much education. I merely say this, as one random example, in one possible interpretation of the above.
Before this paper... I did not quite realize that the "AGI" term was such a rabbit hole, indeed...
Last edited:
Upvote
2
(3
/
-1)
I don't think anyone should feel guilty for using AI tools. The responsibility is in the hands of the people developing more and more use cases without considering the implications, not on the end users who are just leveraging what's in front of them.
Obviously if someone is making revenge porn that's something you can blame them for. But "I used AI to improve parts of my life" isn't that.
It's like the tragedy of the commons, it's not really the fault of individuals, even if they're contributing to a problematic depletion.
Indeed!
Upvote
3
(4
/
-1)
The Sheep Look Up
Ars Praetorian
Those are mighty squishy, and also seem to redefine AGI away from general, or at least get really fast and loose with what "general" means. I wouldn't say that ChatGPT, Bard, or LLaMA can perform a "wide range" of tasks, nor that they can "learn new skills". Even a model you can retrain from scratch to perform an additional skill isn't "general", any more than disassembling a folding knife and replacing the blade with a saw makes it a multitool.
Upvote
1
(2
/
-1)
I think the cynical-but-true take is that AGI is as much a marketing term as it is a goal for true believers, and there's a lot of pressure to move the goal posts to be able to announce that you've achieved it because you can turn on a money faucet with the news.
Upvote
5
(5
/
0)
It's basically the new Turing Test, since that's now achievable. Before the latest LLMs, there'd always be someone claiming their AI "passed" the Turing Test because it fooled some non-scientists into thinking it was a barely literate 13-year-old.
https://www.theguardian.com/technol...-simulates-13-year-old-boy-passes-turing-test
Upvote
3
(3
/
0)
No, but the range is wider than yesterday, skill-learning is happening more than yesterday, and training overhead per unit of intelligence is lower than yesterday.
It used to be true that the same AI couldn't do 2+2 and play tic tac toe.
The GPT method had lower training overhead than the pre-GPT GAN method. And the various research going on in post-GPT AI improvements, will keep improving this.
And, it may reach an eventual point where, a humanoid robot merely needs to metaphorically to "sit" in the equivalent of a human classroom to absorb further training beyond its baked training (e.g. via whatever means however they decide, like a future robot sitting in class, or a model accepting recorded university lectures), etc, and still "graduate" at the top of respective class. Whatever they are doing, the training-overhead-per-unit-of-intelligence is rapidly falling as the years past by. There is a point where AI is able to learn from less amount of training effort than trying to tech an average human; and the graph of training effort per unit is hurtling to approach that point. They are data-hungry as the parametric counts are going up fast, but the training-compute-per-parameter is also rapidly falling.
It's much like how 1980s video compression researchers originally disbelieved we could do video compression at file size averages of less than 1-bit per pixel (averaged) but it happened, with MPEG then MPEG2, and so on.... Likewise, many disbelieve the training overheads are rapidly falling per unit of intelligence, as time passes. But the best labs are discovering a lot of this;
We have to also consider the implications of future AIs, based on this progress.
It wasn't too long ago that AI couldn't remotely execute caveman-style tool-using abilities. Such as AI picking up a computer/calculator tool automatically [note: this only works under paid subscription] when asked a question that NN is suboptimal for, such as trying to do math without paper/calculator. This is true for both biological and artificial.
Tool-using ability is what makes NN's (biological and artificial) more and more useful. Such improved tool-using ability is an achievement unlocked by only a few modern multimodal AIs. An AI's decision whether to answer directly from it NN (e.g. answering rote, such as 2+2) or "decide" to pick up a calculator (e.g. compound interest rate or an algebraic regression formula).
Although the paid multimodal GPT4 only has the nebulous equivalent of ~6 brain-centres that is currently less integrated together than the approx >100+ brain centres of a human biological brain that have much fuzzier boundaries than the boundaries between the DALLE3/Analysis/GPT4-Vision/GPT4-Turbo/etc silos concurrently enabled via the current paid GPT4 subscription.
Even with the best paid subscriptions merely just a few months ago you had to manually select the silo (select DALLE3 mode versus select Analytic mode). Not anymore. At startup you can ask the same AI to generate an image or a math question (that metaphorically automatically picks up a calculator tool too!). Yet it is not quite fully integrated, e.g. GPT4-Vision and DALLE3 doesn't feedback into each other; GPT4-Vision can't "see" DALLE-3 yet, because it's not currently tweaked that way yet.
Over long term, it becomes more integrated/automatic and the boundaries will blur more and more (e.g. one random example is later, things like GPT4-Vision + DALLE3 feedbacking into each other to improve each other. They don't currently yet). Now extrapolate. In a world where >100 AI-centres feedbacks into each other more and more (at thousands of times per second, in a multipass/iterative way/to get the best answer) -- it becomes more truer and truer AGI. And that's when things get more and more dangerous in this type of increasing amounts of post-GPT AI optimizations. Now add a bunch of centres (motion, vision, 3D vision, geometry, etc), with a bunch of more post-GPT optimizations, that also reduces training-per-unit-of-intelligence.
The ginormous flood of new money into AI research is rapidly unlocking a lot of post-GPT optimizations at a much more rapid pace -- an AGI (DeepMind "Expert" definition) outperforming 90% of humans on 90% of disciplines is no longer "Nuclear fusion is always 20 years away", and no longer needs silicon shrinks much smaller than 2nm, according to estimates I've been reading elsewhere. That gets scary that it's now in the window of less than a human generation away. There are discoveries of architectural optimizations occuring that appears to map out a path to an AI 10x-100x faster than today's NVIDIA H100.
Today's GPU architectures helped AI a lot more than CPU, but GPU architectures are highly optimized for 3D graphics rendering, and today researchers everywhere are only now discovering how to rearchitecture silicon for faster AI; and there are some scary-fast new benchmarks (at same die size), that are not yet hitting AI data centres, that will help path towards 10T parameters. You heard what NVIDIA talks about the giant performance increases of future AI processors, thanks to some of the rearchitecturing discoveries that are ongoing;
People who test only a free AI, or software developers who test a non-multimodal AI engine or highly siloized (poorly-integrated-together) multimodal AI, are currently unaware of some of the big-paid big-money research labs that are now far beyond this.
____________
Anyway:
The point is, AI is dumb now, but AI is less and less dumb than yesterday; at a very rapid pace.
Ignoring this is at our peril. Whether one is excited or scared (or both!!) or indifferent;
And that's the cautionary tale / point I am trying to say, rather than critiquing/goalposting today's AI and not planning/considering/safetymoating for tomorrow;
Consider GPT4 Turbo is estimated to have only 1/50th the number of synapse-equivalents of a house mouse (Wikipedia), yet seems to outperform the intelligence of said mouse mouse. Given such optimizations that are undergoing, there are parties that claims we may discover we end up only need ~10T parameters (instead of 100T-120T) to be able to sit-in human classrooms (or vastly outperform 90% of humans at processing teaching videos, a faster equivalent than a physical robot physically sitting in a classroom, that would also still out-learn a human too). And also achieve roughly 90%-ile "Expert" AGI (DeepMind definition). Given much of a real brain is mostly unused.
We have to plan for what is inevitable, not plan for today. The rubicon has recently already unlocked past "Nuclear fusion is always 20 years away" at a far beyond 50-50 chance of a 90%-ile AGI within less than one human generation from today's date. Dismissing this is at humankind's peril.
Solve the copyright problems, solve the thorny ethics, the societal dangers, the misuse etc.
Last edited:
Upvote
1
(2
/
-1)
Everybody needs to chill, the stuff that i've seen that's been AI generated is mostly compiled clips of already shot stock footage clumsily put together, even those clips above don't look genuine and original, in other words, UNCANNY.... when they get quantum computing up to snuff and they merge it with Sora, then there is reason for bricks to be shat.
Last edited:
Upvote
-4
(0
/
-4)